diff --git "a/related_53K/test_related_long_2404.16325v1.json" "b/related_53K/test_related_long_2404.16325v1.json"
new file mode 100644--- /dev/null
+++ "b/related_53K/test_related_long_2404.16325v1.json"
@@ -0,0 +1,8195 @@
+[
+    {
+        "url": "http://arxiv.org/abs/2404.16325v1",
+        "title": "Semantic Segmentation Refiner for Ultrasound Applications with Zero-Shot Foundation Models",
+        "abstract": "Despite the remarkable success of deep learning in medical imaging analysis,\nmedical image segmentation remains challenging due to the scarcity of\nhigh-quality labeled images for supervision. Further, the significant domain\ngap between natural and medical images in general and ultrasound images in\nparticular hinders fine-tuning models trained on natural images to the task at\nhand. In this work, we address the performance degradation of segmentation\nmodels in low-data regimes and propose a prompt-less segmentation method\nharnessing the ability of segmentation foundation models to segment abstract\nshapes. We do that via our novel prompt point generation algorithm which uses\ncoarse semantic segmentation masks as input and a zero-shot prompt-able\nfoundation model as an optimization target. We demonstrate our method on a\nsegmentation findings task (pathologic anomalies) in ultrasound images. Our\nmethod's advantages are brought to light in varying degrees of low-data regime\nexperiments on a small-scale musculoskeletal ultrasound images dataset,\nyielding a larger performance gain as the training set size decreases.",
+        "authors": "Hedda Cohen Indelman, Elay Dahan, Angeles M. Perez-Agosto, Carmit Shiran, Doron Shaked, Nati Daniel",
+        "published": "2024-04-25",
+        "updated": "2024-04-25",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI"
+        ],
+        "label": "Original Paper",
+        "paper_cat": "Semantic AND Segmentation AND Image",
+        "gt": "2.1 Semantic Segmentation Models Semantic segmentation aims to assign a label or a class to each pixel in an image. Unlike image classification, which assigns a single label to the entire image, semantic segmentation provides a more detailed understanding of the visual scene by segmenting it into distinct regions corresponding to objects or classes. This is an essential technique for applications, such as autonomous vehicles, medical image analysis, and scene understanding in robotics. Like other computer vision tasks, deep learning has demonstrated state-of-the-art results in the semantic segmentation of medical images. The semantic segmentation problem can be formulated as follows: Given an image I \u2208RC\u00d7H\u00d7W , our goal is to train a deep neural network to predict the pixel-wise probability map SN\u00d7H\u00d7W of the classes in the dataset, where N is the number of classes in the dataset. 2 Semantic Segmentation Refiner for U/S Applications with Zero-Shot FMs COHEN H ET AL. DeepLabV3 [9] represents a distinctive approach in semantic image segmentation. Utilizing dilated convolutions, the model strategically enlarges the receptive field and manages the balance between global and local features through padding rates. Notably, the spatial pyramid pooling module proposed by the authors aggregates features from dilated convolutions at various scales, enhancing contextual information. Distinctive from encoder-decoder architectures such as the U-Net [10], it is built upon a robust pre-trained encoder, contributing to its success in generating accurate and detailed segmentation masks across diverse applications. Since DeepLabV3 remains a staple choice for a performant semantic segmentation model, we adopt it as our method\u2019s coarse segmentor. 2.2 Semantic Segmentation Foundation Models Foundation models are trained on broad data at a huge scale and are adaptable to a wide range of downstream tasks [11, 12, 13]. The Segment Anything Model (SAM) [5] emerged as a versatile foundation model for natural image segmentation. Trained on a dataset of over 11 million images and 1B masks, it demonstrates impressive zero-shot generalization in segmenting natural images using an interactive and prompt-driven approach. Prompt types include foreground/background points, bounding boxes, masks, and text prompts. However, SAM achieves subpar generalization on medical images due to substantial domain gaps between natural and medical images [14, 15, 16, 17, 18]. Moreover, SAM obtains the poorest results on ultrasound compared to other medical imaging modalities [15]. These results are attributed to the ultrasound characteristics, e.g., the scan cone, poor image quality, and unique speckled texture. A common methodology to overcome this generalization difficulty is to fine-tune a foundation model on a target dataset [19]. An efficient fine-tuning strategy is Low-Rank Adaptation (LoRA) [20], which has been adopted in fine-tuning SAM to relatively small medical imaging datasets [21, 22, 23]. SonoSAM [6] demonstrates state-of-the-art generalization in segmenting ultrasound images. Fine-tuned on a rich and diverse set of ultrasound image-mask pairs, it has emerged as a prompt-able foundational model for ultrasound image segmentation. Notably, adapting prompt-based models to medical image segmentation is difficult due to the conundrum of crafting high-quality prompts [15]. Manually selecting prompts is time-consuming and requires domain expertise. Methods of extracting prompts from ground-truth masks [23] cannot be applied during inference as they rely on full supervision. Auto-prompting techniques rely on the strong Vision Transformer (ViT-H) image encoder [24] semantic representation capabilities, and suggest generating a segmentation prompt based on SAM\u2019s image encoder embedding [18, 25]. Other strategies suggest replacing the mask decoder with a prediction head requiring no prompts [16]. Nevertheless, SAM\u2019s zero-shot prediction accuracy is typically lower than that of the segmentation models trained with fully supervised methods [26]. Motivated by the generalization abilities of segmentation foundation models, we devise a points selection algorithm from coarse segmentation masks that allows harnessing prompt-based models to ultrasound segmentation in a zero-shot setting.",
+        "pre_questions": [],
+        "main_content": "Introduction Ultrasound is a popular medical imaging modality used to image a large variety of organs and tissues. Ultrasound is often the preferred choice due to its non-radiative and non-invasive nature, relatively easy and fast imaging procedure, and lower costs. Automating the diagnosis or highlighting relevant areas in the image will contribute to faster workflows and potentially more consistent and accurate diagnoses. Artificial Intelligence (AI) has demonstrated remarkable success in automatic medical imaging analysis. Compared to classical methods, previous work based on convolutional neural networks on various medical imaging tasks, such as classification and segmentation, have shown state-of-the-art results [1, 2, 3, 4]. However, effective deep learning segmentation algorithms for medical images is an especially challenging task due to the scarcity of high-quality labeled images for supervision. Moreover, in medical imaging it is often the case that identification of findings regions, namely regions of potentially pathological visual anomalies, having neither a clear boundary nor a typical geometry or position is much more challenging than the identification of an anatomy in its context. Findings are also typically rare, which brings to light the challenge of training such models in limited data regimes. \u2217Corresponding author, e-mail: nati.daniel@gehealthcare.com. \u2020These authors have contributed equally to this work. 1Dept. of AI/ML Research, GE Healthcare, Haifa, Israel. 2Dept. of Clinical Applications, Point of Care Ultrasound & Handheld, Texas, USA. 3Dept. of Clinical Applications, Point of Care Ultrasound & Handheld, Wisconsin, USA. arXiv:2404.16325v1  [cs.CV]  25 Apr 2024 Semantic Segmentation Refiner for U/S Applications with Zero-Shot FMs COHEN H ET AL. Figure 1: A high-level illustration of our semantic segmentation refinement method with zero-shot foundation models. A pre-trained segmentation model predicts a semantic segmentation for each class of an input image. In this example, classes comprise anatomies and pathologies in an ultrasound image, and the coarse segmentor output depicts the predicted semantic segmentation of a pathology. A prompt selection model selects positive and negative points. Consequently, a zero-shot semantic segmentation mask of the pathology is predicted by a foundation segmentation model, prompted by the selected points for the input image. Positive prompt points are depicted in red, and negative prompt points are depicted in blue. The pathology semantic segmentation prediction is highlighted in red. For illustration purposes, the muscle is highlighted in purple, the tendon in yellow, and the bone in green. The freeze symbol indicates preventing gradients from being propagated to the model weights. Recently, new segmentation models have emerged. Trained on data at huge scales, these foundation models aim to be more generic rather than tailored to specific datasets. The Segment Anything Model (SAM) [5] is a foundational model demonstrating zero-shot generalization in segmenting natural images using a prompt-driven approach. The SonoSAM [6] foundational model adapts SAM to ultrasound images by fine-tuning the prompt and mask decoder [6]. Although fine-tuning methods often improve the results on target datasets [7] they essentially downgrade the generalization capabilities of the foundation model. Further, a significant domain gap between natural and medical images, ultrasound images in particular[8], hinders fine-tuning models trained on natural images to the task at hand [7]. In this work, we address the performance degradation of segmentation models in low-data regimes and derive a novel method for harnessing segmentation foundation models\u2019 ability to segment arbitrary regions. Our semantic segmentation refinement method comprises two stages: First, a coarse segmentation is predicted by a model trained on a small subset of the training data. In the second stage, our novel points generation from a coarse pathology segmentation algorithm is used to prompt a segmentation foundation model. Positive prompt points are selected using a partition around medoids method as the most representative pathology points. Negative prompt points are selected by a prompt selection optimization algorithm that identify the context anatomy. Importantly, we do not fine-tune the foundation model to our dataset, i.e., it produces a zero-shot segmentation. The end-to-end pipeline is illustrated in Fig. 1. The method\u2019s advantages are brought to light on varying degrees of low-data regimes experiments on a small-scale images dataset, yielding a larger performance gain compared to a state-of-the-art segmentation model [9] as the training set size decreases. Further, ablation studies validate the effectiveness of our semantic segmentation refinement model. Our approach applies to other ultrasound-based medical diagnostics tasks. The paper is organized as follows: Section 2 presents the semantic segmentation task and leading approaches. Our method is presented in Section 3, and the experimental setup is presented in Section 4. Section 5 presents the results and ablation studies on a discontinuity in tendon fiber (DITF) pathology finding task in a musculoskeletal ultrasound (MSK) dataset, and the conclusions are presented in Section 6. In this section, we present our method for refining a coarse pathology segmentation mask with zero-shot foundation models. This method can be adapted to natural images, as well as to the medical imaging domain. Herein, we validate it based on a specific challenging task of segmenting a discontinuity of the tendon fiber finding (Sec. 4.1), which is the main ultrasound finding of a tendon partial tear pathology. Our key intuition is that although the performance of segmentation models decreases significantly in low-data regimes, even such coarse segmentation masks can be utilized for extracting high-quality prompts that harness segmentation foundation models\u2019 capabilities. Importantly, we use the publicly available pre-trained foundation models without further modification. The flexibility of our method allows for incorporating either SonoSAM or SAM. Though the above-mentioned foundation models allow several types of prompts, we focus on foreground (positive) and background (negative) prompt points. Our method makes use of the ground-truth tendon segmentation, denoted T gt. Since the tendon in the context of the DIFT pathology is usually easy to segment due to its typical geometry and position and relatively simple data acquisition and labeling, we assume that strong segmentation models exist for this task and that their output can be used in lieu of the ground-truth segmentation. With that, we introduce our two-stage method, summarized in Algorithm 1. First, a segmentation model [9] is trained on a random subset of the training data. A coarse semantic segmentation is then predicted for a given test image. Then, k positive and k negative prompt points are selected to prompt a segmentation foundation model. We next describe our prompt points selection algorithm in greater detail. 3 Semantic Segmentation Refiner for U/S Applications with Zero-Shot FMs COHEN H ET AL. Algorithm 1 The Semantic Segmentation Refiner Method Input: \u2022 Input image I \u2022 Ground-truth tendon mask T gt \u2022 Frozen SonoSAM model \u2022 Pre-trained segmentation model S Output: \u2022 Refined pathology segmentation mask O 1: Coarse segmentation mask \u02dc O \u2190S(I) 2: Positive points selection ptspos \u2190k-medoids( \u02dc O) 3: Modified ground-truth tendon mask T \u02dc gt \u2190T gt \\ \u02dc O 4: Initialize complementary problem 5: \u00af ptsneg \u2190ptspos, \u00af ptspos \u2190random from T \u02dc gt 6: for t in range(1, T) do 7: Optimize \u00af ptspos as parameters: 8: \u2113ce( \u00af pts, T \u02dc gt) = \u2212T \u02dc gt log (SonoSAM(I, \u00af pts)) 9: Update \u00af ptspos \u2190\u00af ptspos 10: end for 11: Flip: ptsneg \u2190\u00af ptspos 12: Output O \u2190SonoSAM(I, pts) 3.1 Positive Points Selection We aim to select points that are the most representative of the coarse pathology segmentation mask as the positive prompt points. This selection objective translates to the partitioning around the medoids method\u2019s approach. This approach is preferable compared to a selection based on a minimization of the sum of squared distance (i.e., the k-means) in the case of multiple pathology blobs since the latter might select centroids in between pathology blobs. Thus, k mass centers of the coarse pathology segmentation masks are selected as positive points using the kmedoids clustering algorithm [27]. To reduce the probability of selecting false positive points, a threshold is applied to the coarse pathology segmentation masks before selection. We denote the selected positive points as ptspos = {ptspos i }k i=1. This process is illustrated in Fig. 2. Figure 2: An illustration of our positive (foreground) points selection module, depicted in red. A threshold is applied to the coarse segmentation prediction. A kmedoids clustering algorithm is applied to select k positive pathology points. 3.2 Negative Points Refinement We take inspiration from hard negative selection literature [28, 29, 30], and aim to select the most informative negative points w.r.t. the foreground object. To that end, we formulate a complementary prompt points selection problem w.r.t. the background given the k selected foreground points (3.1), \u00af pts = { \u00af ptspos, \u00af ptsneg}. When the foreground is the pathology, the background is the context anatomy, herein the background is a tendon anatomy. The complementary prompt points selection is optimized to decrease the binary cross-entropy (BCE) loss between the foundation model\u2019s zero-shot tendon segmentation mask prompted on these points and a modified ground-truth tendon 4 Semantic Segmentation Refiner for U/S Applications with Zero-Shot FMs COHEN H ET AL. mask, denoted T \u02dc gt. To avoid predicting tendon points within foreground pathology, the values of the ground-truth tendon mask overlapping with the coarse pathology detection are modified to zero. As points initialization for this complementary problem, we flip the labels of ptspos such that they correspond to negative points, \u00af ptsneg \u2190ptspos. Further, k points are selected at random from T \u02dc gt, denoted \u00af ptspos. While freezing the foundation model, the point prompt optimization is performed for a maximum of 100 steps or until convergence. The optimization is performed such that the selected points are optimal w.r.t. the complementary problem of the tendon segmentation given the foreground pathology predicted by the coarse segmentor. Denote an input image as I, SonoSAM\u2019s zero-shot tendon segmentation given input I and its corresponding optimized prompt points \u00af pts as SonoSAM(I, \u00af pts). Then, the BCE loss of the complementary problem is: \u2113ce( \u00af pts, T \u02dc gt) = \u2212T \u02dc gt log (SonoSAM(I, \u00af pts)) . (1) We used the AdamW [31] optimizer, with learning rate of 4e\u22123, and standard betas to optimize the positive points \u00af ptspos. The optimized positive tendon points selected by this model serve as k negative prompt points, ptsneg \u2190\u00af ptspos, towards the foreground pathology segmentation. This process is illustrated in Fig. 3. Figure 3: An illustration of our negative (background) points selection module. In addition to the positive selected points (Sec. 3.1), negative points are selected randomly from the modified ground-truth tendon mask. The points are flipped to initialize the settings of the complementary tendon segmentation problem. Our points optimization model optimizes prompt points selection w.r.t. the complementary tendon zero-shot segmentation problem (Sec. 3.2). Finally, prompt points are again flipped to account for positive and negative prompt points towards the pathology segmentation. 4 Experiments 4.1 Dataset The data used for this study is ultrasound images of tendons around the shoulder joint. Specifically, we acquired images of the supraspinatus tendon, infraspinatus tendon, and subscapularis. The images are acquired from both the short-axis and the long-axis views. The main parameters of our data are summarized in Table 1. In this work, we aim to segment the partial tear pathology within the tendon, thus our data consists of images paired with the corresponding segmentation mask of anatomies and pathologies. Our data includes semantic labeling of the following classes: DITF, bone, tendon, and muscle. Table 2 summarizes the semantic labeling statistics. In total, our dataset includes 388 images from 124 subjects, 80% of which are used for training, and the remaining 20% are used for validation. The test set comprises 40 images. To prevent data leakage, the test set images are collected from subjects that do not appear in the train data. All images are resized to a constant resolution of 512x512 pixels. All data comply with the Institutional Review Board (IRB) data sharing agreement. 4.2 Evaluation Metric We use the Dice similarity coefficient [32] evaluation metric, commonly used in medical image segmentation research to measure the overlapping pixels between prediction and ground truth masks. The Dice similarity coefficient is defined as 2|A\u2229B| |A|+|B|, where A and B are the pixels of the prediction and the ground truth respectively. 4.3 A Segmentation Model In Low-Data Regimes In this experiment, we investigate the performance and properties of a state-of-the-art semantic segmentation model with a limited training set size of MSK ultrasound images. Our goal is two-fold: (i) to validate our conjecture that high-quality 5 Semantic Segmentation Refiner for U/S Applications with Zero-Shot FMs COHEN H ET AL. (a) 100% of train set. (b) 35% of train set. (c) 15% of train set. (d) 8% of train set. (e) 5% of train set. Figure 4: Positive pathology points retainment in increasingly coarse segmentation mask prediction and our method\u2019s results. Top row: Pathology segmentation mask predicted with a DeepLabV3 model trained on varying percent of the training set. Middle row: Positive points selected on binary pathology mask by our positive points selection module. Bottom row: An illustration of our method\u2019s pathology segmentation output, highlighted in red, compared to the ground-truth segmentation, highlighted in green. The tendon area is shown at the bottom left image for reference. Our method achieves for this test image a Dice similarity coefficient of 0.89, 0.71, 0.73, 0.72, 0.50 when the coarse segmentor is trained on 100%, 35%, 15%, 8%, 5% of the train set, respectively. Table 1: Summary of MSK pathology segmentation dataset main parameters. Parameters/Dataset MSK Ultrasound Images Total frames 388 Original frame size 1536 X 796 or 1044 X 646 pixels Subjects 90 (52.82% males, 47.18% females) Average BMI 24.69 \u00b1 8.92 Vendor GE Healthcare\u2122 Ultrasound system Logiq S8\u2122, Eagle\u2122, LogiqE10\u2122 Data collection Linear Collection Sites USA, Israel prompts can be extracted even from a coarse semantic segmentation prediction, and (ii) to measure the performance degradation in increasingly low-data regimes. These properties are the basis of our two-stage method for exploiting the advantages of a prompt-able foundation segmentation model. Concretely, for an input image I \u2208R512\u00d7512 the segmentation model prediction S \u2208R7\u00d7512\u00d7512 corresponds to a semantic segmentation for each class as detailed in Table 2. 4.4 Segmentation Refinement With Zero-Shot Foundation Models Positive Points Selection A combination of a constant and an adaptive threshold is applied to the coarse segmentation prediction prior to positive point selection. Denote by c0 the coarse segmentation mask prediction at the foreground channel (DITF in our case). 6 Semantic Segmentation Refiner for U/S Applications with Zero-Shot FMs COHEN H ET AL. Table 2: Semantic labeling statistics at the 512X512 patches level. M: Million. Class MSK Type Number of images (% of total) Total Area (pixels) Mean fraction out of total patch area Discontinuity in tendon fiber Pathology 179 (46.13%) 1.11M 1.09% Bone 288 (74.22%) 2.75M 2.7% Tendon Anatomy 388 (100%) 10.64M 10.46% Muscle 388 (100%) 28.13M 27.65% We apply a double thresholding mechanism to disregard the noise in the prediction. \u02dc c = c0 > tmin (2) c = \u02dc c > 0.4 \u2217max(\u02dc c) (3) The initial threshold screens predictions that lack sufficient global (cross-classes) certainty, when the minimum threshold is set to tmin = 0.15. The second thresholding term adaptively screens all predictions that lack sufficient local (classwise) certainty. Further, we set the k-medoids++ medoid initialization method [33] which selects more separated initial medoids than those selected by the other methods. The hyper-parameter k is adaptively set such that the sum of distances of samples to their closest cluster center (inertia) is minimized, k \u2208[4, 6]. Negative Points Refinement We deploy in our experiments the SonoSAM semantic segmentation foundation model since it is expected to better generalize to zero-shot segmentation of ultrasound images than SAM. Due to the randomness in the initialization of the complementary positive points \u00af ptspos selection problem, evaluation is performed over 10 random initialization. 4.5 Training Procedure Our coarse segmentor is DeepLabV3 [9], a state-of-the-art convolutional approach to handle objects in images of varying scales, with a ResNet-50 backbone [34]. As our complete dataset consists of only 275 training images, the model is pre-trained on the ImageNet dataset [35]. To evaluate our method across different data regimes we trained our coarse segmentor on varying n percentage of the training data, n \u2208[5, 8, 12, 20, 35, 60, 100], sub-sampled at random. The model is trained with equally weighted BCE loss and a Dice similarity coefficient loss between the predicted and ground-truth segmentation for each class. Each such experiment is trained for 100 epochs, where the weights of the maximal validation loss have been selected for testing. We used the robust AdamW [31] optimizer, with no learning rate scheduler and parameters of \u03b21 = 0.9, \u03b22 = 0.999 and learning rate of 4e\u22123. The test set remains constant across the different training experiments. The model training and evaluation code is implemented with the PyTorch [36] framework. 5 Results 5.1 Semantic Segmentation Model In Low-Data Regimes The results of this experiment validate our conjecture that positive pathology points are consistently selected in increasingly coarse segmentation mask predictions. As the segmentation model is trained on an increasingly smaller training set, the segmentation mask prediction becomes coarse: the pathology segmentation boundaries become less defined and its prediction probability decreases (Fig. 4, top row). Nevertheless, the positive pathology points selected by our method remain generally consistent (Fig. 4, middle row). Consistent with these results, we find that the average Dice similarity coefficient of the segmentation model decreases rapidly when the model is trained on increasingly smaller training set sizes (Fig. 5, \u2018Segmentation Model\u2019). These results validate our method\u2019s motivation and approach. 5.2 Semantic Segmentation Refinement With Zero-Shot Foundation Model Fig. 5 summarizes the results of our method in comparison with those of the baseline segmentation model in various training set sizes. Our method\u2019s average Dice is higher than the baseline\u2019s in every training set size. Moreover, 7 Semantic Segmentation Refiner for U/S Applications with Zero-Shot FMs COHEN H ET AL. our method\u2019s performance gain is larger as the training set size decreases (\u223c10% average Dice increase in 5% and 8% training set sizes), substantiating the advantage of our method in low-data regimes. Our method\u2019s pathology segmentation output in varying training set sizes compared to the ground-truth segmentation is illustrated in Fig. 4, bottom row. 100.0 60.0 35.0 20.0 15.0 12.0 8.0 5.0 Percent of the training set size (%) 0.20 0.25 0.30 0.35 0.40 0.45 Average Dice score Segmentation Model (DeepLabV3) Ours Figure 5: A summary of the average DITF Dice similarity coefficient of methods in various training set sizes. Depicted are the results of the baseline segmentation model[9] and our segmentation refinement with zero-shot SonoSAM foundation model. Error bars depict the standard deviation of our method\u2019s statistics. To analyze the stochasticity effect of our method\u2019s negative points random initialization (Sec. 3.2), we compare our method\u2019s DITF Dice score statistics over ten random initialization and the baseline segmentation model\u2019s average DITF Dice similarity coefficient. Results show that our method\u2019s performance is robust, exhibiting relatively low standard deviation in all train set sizes (Fig. 5). Additionally, our method\u2019s mean DITF Dice surpasses that of the baseline\u2019s in all but one train set size, and is higher by 4% on average than the baseline. 5.3 Ablation Studies In this section, we present ablation studies substantiating the effectiveness of our negative prompt points refinement (NPPR) model, as well as examining our method\u2019s performance when replacing the SonoSAM foundation model with SAM. 5.3.1 SAM vs. SonoSAM as a segmentation foundation model In this study, we investigate the impact of replacing SonoSAM with SAM as the zero-shot semantic segmentation foundation model in our method. Table 3 shows that harnessing SonoSAM\u2019s generalizability for MSK ultrasound images is preferable to SAM in low-data regimes and on par with SAM otherwise. 5.3.2 Random negative prompt points section In this experiment, we investigate the effectiveness of our negative prompt points refinement model by comparing it to a random negative prompt points selection algorithm. Concretely, k negative prompt points are randomly selected from the modified ground-truth tendon mask, T \u02dc gt. Our positive points selection approach remains unchanged. Results in Table 3 demonstrate that this naive selection algorithm achieves subpar average Dice scores across almost all train set sizes, especially in low-data regimes. These results establish the advantage of our negative points optimization algorithm. 6 Conclusions In this paper, we address the performance degradation of a state-of-the-art semantic segmentation model in low-data regimes. A novel prompt points selection algorithm optimized on a zero-shot segmentation foundation model was presented, as a means of refining a coarse pathology segmentation. Our method\u2019s advantages are brought to light in varying degrees of low-data regimes experiments, demonstrating a larger performance gain compared to the baseline segmentation model as the training set size decreases (Fig. 5). 8 Semantic Segmentation Refiner for U/S Applications with Zero-Shot FMs COHEN H ET AL. Table 3: Ablation studies: quantitative segmentation test results of the mean DITF Dice similarity coefficient (DSC) for different approaches over 10 run cycles. Our method is using zero-shot SonoSAM [6] foundation model. A higher DSC is better, with the best scores marked in bold. NPPR: Negative Prompt Points Refinement. Methods Percent of the training set 100% 60% 35% 20% 15% 12% 8% 5% Ours without NPPR 44.6% 40.0% 34.2% 27.8% 30.3% 27.5% 20.7% 16.6% Ours with SAM 45.5% 41.6% 39.7% 29.3% 32.9% 28.3% 27.6% 23.0% Ours 46.3% 39.3% 39.6% 31.9% 32.8% 31.8% 32.0% 24.6% Further, we validate our method\u2019s robustness to negative point initialization stochasticity and study the effectiveness of our prompt points refinement model (Section 5.3.2). Results demonstrate that the generalization of SonoSAM in extremely low data regimes is better than SAM\u2019s (Section 5.3.1). Our approach can be used for other ultrasound-based medical diagnostics tasks. An inherent limitation of our two-stage method is that its latency is higher than that of a core segmentation model.",
+        "additional_info": [
+            [
+                {
+                    "url": "http://arxiv.org/abs/2404.07410v1",
+                    "title": "Improving Shift Invariance in Convolutional Neural Networks with Translation Invariant Polyphase Sampling",
+                    "abstract": "Downsampling operators break the shift invariance of convolutional neural\nnetworks (CNNs) and this affects the robustness of features learned by CNNs\nwhen dealing with even small pixel-level shift. Through a large-scale\ncorrelation analysis framework, we study shift invariance of CNNs by inspecting\nexisting downsampling operators in terms of their maximum-sampling bias (MSB),\nand find that MSB is negatively correlated with shift invariance. Based on this\ncrucial insight, we propose a learnable pooling operator called Translation\nInvariant Polyphase Sampling (TIPS) and two regularizations on the intermediate\nfeature maps of TIPS to reduce MSB and learn translation-invariant\nrepresentations. TIPS can be integrated into any CNN and can be trained\nend-to-end with marginal computational overhead. Our experiments demonstrate\nthat TIPS results in consistent performance gains in terms of accuracy, shift\nconsistency, and shift fidelity on multiple benchmarks for image classification\nand semantic segmentation compared to previous methods and also leads to\nimprovements in adversarial and distributional robustness. TIPS results in the\nlowest MSB compared to all previous methods, thus explaining our strong\nempirical results.",
+                    "authors": "Sourajit Saha, Tejas Gokhale",
+                    "published": "2024-04-11",
+                    "updated": "2024-04-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Original Paper",
+                    "paper_cat": "Semantic AND Segmentation AND Image",
+                    "gt": "Robustness of CNNs has been examined under different types of input perturbations and transformations such as rotation, reflection and scaling (Cohen & Welling, 2016; Poulenard et al., 2019), geometric transformations (Liu et al., 2019), affine transformations (Engstrom et al., 2019), domain shift (Venkateswara et al., 2017), attribute shift (Gokhale et al., 2021), adversarial attacks and perturbations (Agarwal et al., 2020; Zhang et al., 2021), and natural corruptions (Hendrycks & Dietterich, 2018). Distributional robustness of CNNs has been explored through various approaches including static, random, or learned data augmentation (Hendrycks et al., 2019; Xu et al., 2020; Gokhale et al., 2023), contrastive learning (Khosla et al., 2020), and Bayesian approaches (Cheng et al., 2023). Dense Sampling and Anti-aliasing. Conventional sliding window downsampling in computer vision algorithms (Fukushima, 1980; Lowe, 1999) is typically applied with stride that is bigger than 1 which breaks shift equivariance (Simoncelli et al., 1992). Shift invariance can be improved through dense sampling (Leung & Malik, 2001) with dilated convolutions (Yu et al., 2017) with susceptibility to griding artifacts. Zhang (2019) suggest BlurPool to enhance SI through anti-aliasing before downsampling, whereas Zou et al. (2020) propose DDAC, to learn low pass anti-aliasing filter. Polyphase Sampling. Recent works such as APS (Chaman & Dokmanic, 2021) and LPS (Rojas-Gomez et al., 2022) use polyphase sampling to meet the Nyquist sampling theorem (Nyquist, 1928) and permutation invariance which provides robustness against circular shifts. APS enhances shift invariance by sampling the highest energy polyphase index (\u2113p norm) while LPS learns the sampling. LPS being sensitive to gumble softmax temperature can sample polyphases to maximize downstream objective which does not consider shift invariance unless training data is shift-augmented. Evidence suggests that although polyphase sampling methods can improve shift invariance for circular shift, they still struggle to deal with standard shift. The focus of this study is to improve robustness of CNNs against both standard and circular shifts, which constitute significant aspects of model evaluation.",
+                    "pre_questions": [],
+                    "main_content": "Introduction Shift invariance is an ideal property for visual recognition models and necessitates that predictions remain invariant to small pixel-level shifts in input images. Shifting an image by a few pixels horizontally and/or vertically should not affect the category predicted by an image classifier such as a convolutional neural 1Code is available at https://github.com/sourajitcs/tips/ 1 arXiv:2404.07410v1  [cs.CV]  11 Apr 2024 network (CNN). Figure 2 depicts three scenarios where an input x undergoes a transformation g before being fed into a model f to generate a prediction \u02c6 y = f(g(x)) = g\u2032(f(x)): shift equivariance, shift noninvariance, and shift invariance. If g\u2032 = g, then f is g-equivariant and if g\u2032 = I then f is g-invariant. Shift-invariance is desirable for image classification to ensure that categorical outputs are invariant to pixel shift, and shift-equivariance is desirable for semantic segmentation to ensure that pixel-shift in the image results in equivalent shift in segmentation. Recent studies have also found shift invariant visual recognition models to be more robust on out-of-distribution testing and adversarial attacks, therefore improving shift invariance in CNNs is consequential. Figure 2: An illustration of shift equivariance, noninvariance, and invariance. Invariant models map shifted, non-shifted inputs to identical outputs, while equivariant models mirror the input shift in outputs. Although individual convolution operations in CNNs are shift-equivariant (Fukushima, 1980; LeCun et al., 1989; 2015), recent studies (Zhang, 2019; Azulay & Weiss, 2019; Zou et al., 2023) reveal that conventional pooling operators in CNNs such as max pooling, average pooling, and strided convolution break shift invariance by violating the Nyquist sampling theorem (Nyquist, 1928) and aliasing highfrequency signals, and impact model prediction at pooling boundaries under small shift transformations. Pooling techniques such as Max pooling, Avg Pooling and, Adaptive Polyphase Sampling (APS) (Chaman & Dokmanic, 2021) are subject to activation strength within pooling windows, i.e. the maximum feature activation in a pooling window influences the pooled value. While pooling methods such as Adaptive Polyphase Sampling (APS) (Chaman & Dokmanic, 2021) and Learnable Polyphase Sampling (LPS) (Rojas-Gomez et al., 2022) work well for circular shift, empirical evidence (Rojas-Gomez et al., 2023; Ding et al., 2023; Zhong et al., 2023) suggests lack of robustness for standard shifts. This observation raises an important question: while sampling the strongest signal works well for downstream tasks, does it affect the network\u2019s performance under pixel shift? In this work, we study the correlation between shift invariance and the tendency to downsample strongest features by introducing the concept of Maximum-Sampling Bias (MSB). We observe a strong negative correlation between MSB and shift invariance, i.e. models with higher MSB are the least shift invariant. Based on insights from our large-scale correlation study, we design a novel pooling operator called Translation Invariant Polyphase Sampling (TIPS) that discourages MSB and improves invariance under shift transformations. To further improve visual recognition performance and shift invariance, we introduce two loss functions: LF M \u2013 to discourage known failure modes of shift invariance and Lundo \u2013 to learn to undo standard shift. In real world scenarios, standard shifts are more like to occur than circular shift; however current literature largely focuses on circular shift invariance. We show that the TIPS pooling operator and regularization improves shift invariance on both circular shift and standard shift. Our contributions and findings are summarized below and results are highlighted in Figure 1. \u2022 We identify maximum-sampling bias (MSB) as a factor that hurts the shift invariance of existing pooling methods in CNNs. \u2022 We propose a learnable pooling method called Translation Invariant Polyphase Sampling (TIPS) and two regularizations called LF M and Lundo to improve shift invariance and discourage MSB when training visual recognition models. \u2022 We demonstrate that this approach consistently improves robustness under shift transformation on multiple image classification and semantic segmentation benchmarks, outperforming data augmentation and contrastive learning strategies, and resulting in state-of-the-art performance in terms of accuracy, shift consistency, and shift fidelity under standard and circular shift transformations, while operating at a small computational overhead. \u2022 When tested on adversarial attacks, patch attacks, and natural corruption of images, models trained with TIPS exhibit greater robustness than previous shift-invariant pooling operators. 2 In this section, we discuss the design of the TIPS layer and the workflow for training CNNs with TIPS. Let X \u2208Rc\u00d7h\u00d7w be a ReLU-activated input feature map, where c, h, w denote the number of channels, height, and width of feature maps. A pooling layer with stride s, downsamples X into \u02c6 X where \u02c6 X \u2208R+c\u00d7h/s\u00d7w/s. 3.1 TIPS: A Learnable Pooling Layer TIPS learns to sample polyphase decompositions of input feature maps X using two branches. In the first branch, polyphase components of X with stride s are computed similar to Chaman & Dokmanic (2021): polyis+j(X) = X[k, sn1 + i, sn2 + j], \u2200i, j \u2208Z s\u22121 0 ; k \u2208Z c\u22121 0 ; n1 \u2208Z\u2308h s  0 \u2308h s \u2309 0 ; n2 \u2208Z\u2308w s  0 \u2308w s \u2309 0 . (1) Note that polyphase sampling can also be achieved by a strided convolution with a s \u00d7 s kernel equal to 1 at index (i, j) and 0 elsewhere. A visualization of polyphase sampling with s = 2 is shown in Figure 3. The second branch of TIPS is a small fully convolutional function f\u03b8 : X \u2192\u03c4 that learns the mixing coefficients \u03c4 \u2208[0, 1]c\u00d7s2. \u03c8() is the first 3\u00d73 convolutional layer followed by ReLU activation as shown in Figure 3; thus \u03c8(X) represents intermediate feature maps of X in f\u03b8(). All operations in f\u03b8 are shift invariant, including Global Average Pooling (GAP) (He et al., 2016b). The output of the TIPS layer is computed as a weighted linear combination of the polyphase components: \u02c6 X = \ufffd i,j \ufffd i,j \u03c4 is+j polyis+j(X) (2) Regularizing TIPS to Discourage Known Failure Modes of Shift Invariance. Chaman & Dokmanic (2021); Rojas-Gomez et al. (2022) have shown that having extremely skewed mixing coefficients (e.g. 3 Figure 3: TIPS downsamples ReLU activated intermediate feature map X into \u02c6 X with stride of s. From input feature map X, TIPS learns polyphase mixing coefficients \u03c4 using a small fully convolutional function f\u03b8. The polyphase decomposition on input feature map X results in polyi which are then mixed as a weighted linear combination with \u03c4 (Equation 2) to compute \u02c6 X. \u03c4={0, 1, 0, 0} for s=2) is not robust against standard shift. TIPS with uniform mixing coefficients and LPS with higher softmax temperature (e.g. \u03c4={0.25, 0.25, 0.25, 0.25} for s=2) is identical to average pooling, which has been shown to hurt shift invariance (Zhang, 2019; Zou et al., 2020). Based on these observations, we introduce a regularization on the mixing coefficients in TIPS to discourage known failure modes. In LF M, the first term discourages skewed \u03c4 and the second term discourages uniform \u03c4. LF M = (\u2225\u03c4\u22252 \u22121) + (1 \u2212s2 \u2225\u03c4\u22252) = (1 \u2212s2) \u2225\u03c4\u22252 . (3) Figure 4: The end-to-end training pipeline with TIPS, Translation invariant regularization Lundo, regularization to discourage known failure modes of shift invariance LF M, and downstream task loss Ltask. Learning to Undo Standard Shift. Although prior work (Chaman & Dokmanic, 2021; RojasGomez et al., 2022; 2023; Ding et al., 2023; Zhong et al., 2023) has shown the benefits of using polyphase sampling to counter circular shifts, there is still a performance degradation with standard shifts due to information loss beyond the pooling boundary. To improve robustness against standard shift, we shift the input feature map with a random amount of vertical and horizontal standard shift sampled from uniform distribution U(0, h 10) and U(0, w 10) respectively to obtain a shifted Xt. We then regularize training by setting up the objective of undoing this shift between \u03c8(X) and the shifted Xt, via the additional loss term Lundo: Lundo = Eh\u2032\u2208hEw\u2032\u2208w[Xt h\u2032,w\u2032 \u2212\u03c8(Xh\u2032,w\u2032 )]2 (4) 3.2 Training CNNs with TIPS Let N be the number of training epochs. For the first \u03f5N epochs, we train only with the task loss Ltask and the regularization to discourage failure models LF M. For subsequent epochs, the undo regularization Lundo is introduced. The final training loss is the Lagrangian (with \u03b1 \u2208[0, 1]): L = ( (1 \u2212\u03b1)Ltask + LF M for epoch < \u03f5N (1 \u2212\u03b1)Ltask + \u03b1Lundo + LF M otherwise. (5) \u03c8(X) contains a 3 \u00d7 3 convolution layer, followed by ReLU. This is followed by global average pooling, 1 \u00d7 1 convolution, flattening, and a softmax operation to obtain \u03c4 as shown in Figure 3. Weights of f\u03b8 are initialized using the using Kaiming normal approach (He et al., 2016a). 4 Figure 5: Our large-scale study of shift invariance of CNN-based models for image classification and semantic segmentation, with multiple CNN architectures, datasets, and pooling methods, shows a strong negative correlation between each evaluation metric and MSB (%), as indicated by the Pearson correlation coefficient (r). Linear clusters with negative correlation are also observed for points belonging to each pooling method. 4 Maximum-Sampling Bias and its Correlation with Shift Invariance In this section, we setup a framework to study shift invariance in CNNs, by defining maximum-sampling bias (MSB). We show that MSB is a common preference exhibited by both conventional pooling operators and those designed to improve shift invariance and through a large-scale analysis, we show that MSB is negatively correlated with shift invariance. Definition of MSB. Existing pooling operators exhibit a common tendency to propagate signals based on activation strength. We denote this phenomenon as maximum-sampling bias (MSB), defined as the fraction of window locations for which the the maximum signal value is sampled. MSB quantifies the bias of a pooling operator to select and propagate the maximum value of the signal; a higher MSB indicates a higher probability of maximum signal values being selected during pooling. Let p() be a pooling operator in a convolutional neural network and s be the downsampling factor; for example, s=2 for a max-pooling operator with window size 2 \u00d7 2. Let X \u2208Rh\u00d7w be the 2-dimensional input to a pooling layer. Applying a pooling operator p() on X with downsampling factor s results in an output \u02c6 X = p(X) \u2208R h s \u00d7 w s . It is trivial to see that MSB = 1 for max-pool, as max-pool by definition always selects the maximum signal, \u02c6 X[i, j] = max m,n X[is+m, js+n] \u2200(i, j). Average pooling produces X[i, j] = E m,nx[is+m, js+n] and is equivalent to max-pooling if all values within the window are identical. For all other cases, the average value is sampled, which is necessarily less than the maximum and thus MSB \u22641. APS pooling (Chaman & Dokmanic, 2021) samples the polyphase component of X with maximum \u2113p norm; \u02c6 X[i, j] = max is+j {\u2225polyj(x)\u2225p}s\u22121 i,j=0. As the polyphase function poly() in APS and LPS (Rojas-Gomez et al., 2022) is a monotonic function, it also exhibits a preference for sampling larger signals in the pooling window. From the above definition, we can see that existing pooling operators implicitly prefer selecting larger elements in the pooling window. We investigate whether this preference (or bias) towards maximum-sampling is linked to shift invariance, by conducting a large-scale analysis of the correlation between MSB and shift invariance on a number of visual recognition benchmarks with multiple CNN architectures and pooling methods. Our work is the first to identify maximum-sampling bias and its connection to shift invariance. 5 Standard Shift Circular Shift Method Acc. Consistency Fidelity Consistency Fidelity CNN (ResNet-18) MaxPool 91.43\u00b10.04 87.43\u00b10.05 79.94\u00b10.05 90.18\u00b10.03 82.45\u00b10.08 APS 94.02\u00b10.07 92.89\u00b10.08 87.33\u00b10.05 100.00\u00b10.00 94.02\u00b10.07 LPS 94.45\u00b10.05 93.11\u00b10.07 87.94\u00b10.03 100.00\u00b10.00 94.45\u00b10.05 TIPS 95.75\u00b10.11 98.38\u00b10.37 94.20\u00b10.08 100.00\u00b10.00 95.75\u00b10.11 BlurPool (LPF-5) 94.29\u00b10.11 91.04\u00b10.09 85.84\u00b10.12 98.27\u00b10.11 92.66\u00b10.07 APS (LPF-5) 94.44\u00b10.09 93.25\u00b10.13 88.06\u00b10.17 100.00\u00b10.00 94.44\u00b10.09 LPS (LPF-5) 95.17\u00b10.12 94.87\u00b10.08 90.09\u00b10.15 100.00\u00b10.00 95.17\u00b10.12 TIPS (LPF-5) 96.05\u00b10.13 98.65\u00b10.11 94.75\u00b10.10 100.00\u00b10.00 96.05\u00b10.13 ViT ViT-B/16 (I21k) 98.89\u00b10.04 82.34\u00b10.07 81.43\u00b10.05 83.79\u00b10.15 82.86\u00b10.12 ViT-L/16 (I21k) 99.15\u00b10.02 82.72\u00b10.09 82.01\u00b10.08 84.41\u00b10.11 83.69\u00b10.06 Swin-B (I21k) 99.22\u00b10.03 83.19\u00b10.07 82.54\u00b10.05 84.05\u00b10.04 83.40\u00b10.04 (a) CIFAR-10 Standard Shift Circular Shift Method Acc. Consistency Fidelity Consistency Fidelity CNN (ResNet-34) MaxPool 88.38\u00b10.04 90.25\u00b10.07 79.76\u00b10.13 88.21\u00b10.08 77.96\u00b11.05 APS 88.49\u00b10.12 93.54\u00b10.11 82.77\u00b10.09 100.00\u00b10.00 88.49\u00b10.12 LPS 87.62\u00b10.07 92.73\u00b10.18 81.25\u00b10.15 100.00\u00b10.00 87.62\u00b10.07 TIPS 91.86\u00b10.03 95.77\u00b10.04 87.97\u00b10.12 100.00\u00b10.00 91.86\u00b10.03 BlurPool (LPF-5) 87.79\u00b10.11 92.65\u00b10.14 81.34\u00b10.14 95.39\u00b10.10 73.81\u00b10.14 APS (LPF-5) 88.57\u00b10.07 93.97\u00b10.04 83.20\u00b10.02 100.00\u00b10.00 88.57\u00b10.07 LPS (LPF-5) 88.79\u00b10.12 93.41\u00b10.08 83.00\u00b10.11 100.00\u00b10.00 88.79\u00b10.12 TIPS (LPF-5) 92.34\u00b10.09 95.96\u00b10.11 88.61\u00b10.07 100.00\u00b10.00 92.34\u00b10.09 ViT ViT-B/16 (I21k) 91.54\u00b10.07 87.25\u00b10.08 79.87\u00b10.10 82.39\u00b10.04 75.42\u00b10.06 ViT-L/16 (I21k) 93.39\u00b10.05 87.11\u00b10.18 81.35\u00b10.15 81.49\u00b10.07 76.11\u00b10.11 Swin-B (I21k) 93.78\u00b10.03 87.34\u00b10.06 81.91\u00b10.17 83.57\u00b10.11 78.37\u00b10.13 (b) CIFAR-100 Table 1: Image classification performance on CIFAR-10 and CIFAR-100 averaged over five trials. Negative Correlation between MSB and Shift Invariance. To understand how MSB affects shift invariance in CNNs we evaluated 576 models across different architectures, datasets, and pooling methods 2 and conducted a correlation study as shown in Figure 5 with MSB on x-axis and performance metrics on the y-axis for both image classification and semantic segmentation. A strong negative correlation is observed between MSB and shift consistency and fidelity (discussed in the next subsection), and surprisingly also with downstream task performance (accuracy, mIoU). In all scenarios, when MSB decreases, shift invariance and downstream performance improves. Figure 5 further depicts that circular consistency is more negatively correlated with MSB than standard consistency for both tasks. Linear relationships are also observed for points corresponding to specific pooling methods across architectures and datasets. Using Global Average Pooling (GAP) (He et al., 2016a) before classification layer with no spatial downsampling of the intermediate feature maps leads to additional computational expense since there are more feature grids to convolve. While this design choice helps improving shift invariance, the computational expense for additional convolution operations renders such designs impractical (see Appendix Table 11 for computational costs). TIPS achieves high shift invariance with marginal computational overhead. 5 Experiments We perform experiments on multiple benchmarks for both image classification and semantic segmentation. For image classification, we evaluate shift invariance while for semantic segmentation we evaluate shift equivariance; following conventions used in prior work, this is also referred to as \u201cshift invariance\u201d in the results. For both classification and segmentation, we avoid using pre-trained CNNs since the pre-training step uses strided convolution and maxpool. 5.1 Image Classification Experiments Datasets and Baselines. We benchmark the performance of TIPS and prior work on six image classification datasets: CIFAR-10, 100 (Krizhevsky, 2009), Food-101 (Bossard et al., 2014), Oxford-102 (Nilsback & Zisserman, 2008), Tiny ImageNet (Le & Yang, 2015), and ImageNet (Krizhevsky et al., 2012). Our baselines include MaxPool, APS (p=2), and LPS (\u03c4=0.23), as well as BlurPool, APS, and LPS with anti-aliasing using n\u00d7n Gaussian low-pass filter (LPF-5). We also compare with three Vision Transformer (ViT) architectures: ViT-B/16, ViT-L-16 (Dosovitskiy et al., 2020), and Swin-B (Liu et al., 2021) which are pre-trained on the larger ImageNet-21k dataset (Deng et al., 2009). Hyperparameters. For TIPS, we choose \u03f5 = 0.4 and \u03b1 = 0.35 in Equation 5. All models are trained using an SGD optimizer with initial learning rate 0.05, momentum 0.9, and weight decay 1e-4 with early stopping. No models in our experiments were trained on shifted images. For each dataset-backbone pair, for fair comparison, TIPS and all baselines are trained with identical hyperparameters. 2The appendix has details about architectures, datasets, pooling methods, and hyperparameters for the all experiments. 6 Standard Shift Circular Shift Method Acc. Consistency Fidelity Consistency Fidelity CNN (ResNet-50) MaxPool 92.96\u00b10.08 82.13\u00b10.57 76.18\u00b10.07 83.61\u00b10.12 77.72\u00b10.05 APS 94.68\u00b10.11 91.34\u00b10.04 86.48\u00b10.13 100.00\u00b10.00 94.68\u00b10.11 LPS 94.71\u00b10.02 92.41\u00b10.03 87.52\u00b10.11 99.48\u00b10.11 94.22\u00b10.05 TIPS 95.63\u00b10.15 95.02\u00b10.09 90.87\u00b11.08 100.00\u00b10.00 95.63\u00b10.15 BlurPool (LPF-5) 93.77\u00b10.03 88.18\u00b10.17 82.69\u00b11.08 93.49\u00b10.13 87.67\u00b10.03 APS (LPF-5) 94.07\u00b10.13 92.51\u00b10.06 87.03\u00b10.20 100.00\u00b10.00 94.07\u00b10.13 LPS (LPF-5) 95.62\u00b10.07 94.10\u00b10.07 89.99\u00b10.19 100.00\u00b10.00 95.62\u00b10.07 TIPS (LPF-5) 96.42\u00b10.16 95.50\u00b10.13 92.08\u00b10.19 100.00\u00b10.00 96.42\u00b10.16 ViT ViT-B/16 (I21k) 96.88\u00b10.13 81.45\u00b10.04 78.91\u00b10.15 78.39\u00b10.12 75.94\u00b10.12 ViT-L/16 (I21k) 97.00\u00b10.03 81.84\u00b10.11 79.38\u00b10.08 78.06\u00b10.18 75.72\u00b10.17 Swin-B (I21k) 97.49\u00b10.05 82.85\u00b10.14 80.77\u00b10.09 78.05\u00b10.02 76.10\u00b10.08 (a) Food-101 Standard Shift Circular Shift Method Acc. Consistency Fidelity Consistency Fidelity CNN (ResNet-50) MaxPool 93.48\u00b10.15 85.63\u00b10.11 80.05\u00b10.17 89.38\u00b10.17 83.55\u00b10.12 APS 94.68\u00b10.03 92.47\u00b10.05 87.55\u00b11.09 100.00\u00b10.00 94.68\u00b10.03 LPS 95.31\u00b10.08 93.63\u00b10.17 89.24\u00b10.11 100.00\u00b10.00 95.31\u00b10.08 TIPS 97.18\u00b10.06 95.78\u00b10.03 93.08\u00b10.16 100.00\u00b10.00 97.18\u00b10.06 BlurPool (LPF-5) 92.71\u00b10.08 90.32\u00b10.13 83.74\u00b10.05 94.07\u00b10.13 87.21\u00b10.08 APS (LPF-5) 94.71\u00b10.11 93.00\u00b10.08 88.09\u00b10.14 100.00\u00b10.00 94.71\u00b10.11 LPS (LPF-5) 96.28\u00b10.05 94.33\u00b10.06 90.82\u00b10.09 100.00\u00b10.00 96.28\u00b10.05 TIPS (LPF-5) 97.62\u00b10.11 96.51\u00b10.14 94.21\u00b10.14 100.00\u00b10.00 97.62\u00b10.11 ViT ViT-B/16 (I21k) 99.33\u00b10.05 88.47\u00b10.04 87.88\u00b10.08 82.24\u00b10.03 81.69\u00b10.06 ViT-L/16 (I21k) 99.59\u00b10.03 87.25\u00b10.09 86.89\u00b10.18 82.39\u00b10.13 82.05\u00b10.03 Swin-B (I21k) 99.68\u00b10.02 87.06\u00b10.16 80.16\u00b10.07 83.57\u00b10.11 83.30\u00b10.05 (b) Oxford-102 Table 2: Image classification performance on Food-101 and Oxford-102 datasets averaged over five trials. Standard Shift Circular Shift Method Acc. Consistency Fidelity Consistency Fidelity CNN (ResNet-101) MaxPool 78.54\u00b10.22 88.45\u00b10.15 69.47\u00b10.14 92.82\u00b1.14 79.20\u00b10.15 APS 83.01\u00b10.08 91.37\u00b10.06 75.85\u00b10.04 100.00\u00b10.00 83.01\u00b10.08 LPS 85.67\u00b10.18 92.95\u00b10.04 79.63\u00b10.05 100.00\u00b10.00 85.67\u00b10.18 TIPS 86.78\u00b10.19 94.27\u00b10.14 81.80\u00b10.15 100.00\u00b10.00 86.78\u00b10.19 BlurPool (LPF-5) 82.83\u00b10.13 90.81\u00b10.17 75.22\u00b10.12 95.87\u00b10.19 79.41\u00b11.12 APS (LPF-5) 83.52\u00b10.03 92.00\u00b10.20 76.84\u00b10.11 100.00\u00b10.00 83.52\u00b10.03 LPS (LPF-5) 86.74\u00b10.09 93.38\u00b10.17 80.99\u00b10.04 100.00\u00b10.00 86.74\u00b10.09 TIPS (LPF-5) 86.91\u00b10.13 94.55\u00b10.06 88.20\u00b10.07 100.00\u00b10.00 86.91\u00b10.13 ViT ViT-B/16 (I21k) 89.34\u00b10.06 73.47\u00b10.03 65.64\u00b10.11 72.94\u00b10.19 65.16\u00b10.08 ViT-L/16 (I21k) 90.75\u00b10.15 74.39\u00b10.16 67.51\u00b10.14 73.85\u00b10.06 67.02\u00b10.19 Swin-B (I21k) 91.19\u00b10.04 72.14\u00b10.15 65.78\u00b10.17 75.49\u00b10.05 68.84\u00b10.10 (a) TinyImageNet Standard Shift Circular Shift Method Acc. Consistency Fidelity Consistency Fidelity CNN (ResNet-101) MaxPool 76.31\u00b10.18 89.05\u00b10.19 67.05\u00b10.06 87.56\u00b10.13 66.82\u00b10.17 APS 76.07\u00b10.15 90.95\u00b10.13 69.19\u00b10.13 100.00\u00b10.00 76.07\u00b10.15 LPS 78.29\u00b10.14 91.74\u00b10.03 71.82\u00b10.13 100.00\u00b10.00 78.29\u00b10.14 TIPS 80.24\u00b10.09 92.87\u00b10.08 74.52\u00b10.18 100.00\u00b10.00 80.24\u00b10.09 BlurPool (LPF-5) 76.33\u00b10.08 90.70\u00b10.14 69.23\u00b10.15 90.55\u00b10.17 69.12\u00b10.19 APS (LPF-5) 76.49\u00b10.08 91.23\u00b10.17 69.78\u00b10.05 99.98\u00b10.00 76.41\u00b10.06 LPS (LPF-5) 78.31\u00b10.05 92.49\u00b10.15 72.43\u00b10.08 100.00\u00b10.00 78.31\u00b10.05 TIPS (LPF-5) 81.36\u00b10.10 93.11\u00b10.03 75.75\u00b10.14 100.00\u00b10.00 81.36\u00b10.10 ViT ViT-B/16 (I21k) 83.89\u00b10.07 84.38\u00b10.05 70.79\u00b10.27 81.03\u00b10.11 67.98\u00b10.19 ViT-L/16 (I21k) 85.06\u00b10.02 83.19\u00b10.12 70.76\u00b10.17 81.64\u00b10.15 69.44\u00b10.14 Swin-B (I21k) 85.16\u00b10.05 85.24\u00b10.19 72.59\u00b10.05 82.79\u00b10.08 70.50\u00b10.18 (b) ImageNet Table 3: Image classification performance on TinyImageNet and ImageNet averaged over five trials. Evaluation Metrics. In addition to reporting classification accuracy on the unshifted test set, we use the consistency definition from Zou et al. (2020) which compares the predictions for two shifted images. However, as consistency does not consider the ground truth label (y) for evaluation, we introduce fidelity as a new metric. Note: xh1,w1 denotes image x shifted by h \u223cU(0, h 8 ) vertically and w \u223cU(0, w 8 ) horizontally. Consistency = E x E (h1,w1),(h2,w2)1[f(xh1,w1) = f(xh2,w2)]. (6) Fidelity = E x E (h1,w1),(h2,w2)1[y = f(xh1,w1) = f(xh2,w2)]. (7) Results. Tables 1, 2, and 3 show strong datasetand backbone-agnostic evidence for the efficacy of TIPS in terms of accuracy and shift invariance for both standard shift and circular shift. TIPS results in large gains in consistency and fidelity on standard shift, which was a challenge for prior work. It is important to note that TIPS with LPF-5 also improves upon prior work that uses LPF-5 anti-aliasing. For ViTs, shift invariance performance is inferior to CNNS, even though they consistently achieve higher accuracy. ViT architectures despite being pre-trained on a very large scale dataset ImageNet21k (I21k) cannot improve shift invariance which depicts that large-scale pre-training has no implications on shift invariace. While CNNs in general perform better on circular shift than standard shift, there is no such clear trend for ViT \u2013 for example, ViTs are more robust on standard shift for Oxford-102 and Tiny ImageNet and more robust on circular shift for the other four datasets. 5.2 Semantic Segmentation Experiments Datasets and Baselines. We use the following datasets: PASCAL VOC 2012 (Everingham et al., 2010), Cityscapes (Cordts et al., 2016), Kvasir (Jha et al., 2020), and CVC-ClinicDB (Bernal et al., 2015). Our baselines include MaxPool, APS, LPS, BlurPool (LPF-3), and DDAC (groups g=8, LPF-3). BlurPool and DDAC (Zou et al., 2020) perform antialiasing by either using a fixed low-pass filter (BlurPool) or a learnable low pass group-wise convolution filter (DDAC). 7 PASCAL VOC 2012 DeepLabV3+ (ResNet-18) Cityscapes DeepLabV3+ (ResNet-101) Unshifted Standard Shift Circular Shift Unshifted Standard Shift Circular Shift Method Anti-Alias mIOU Consistency Fidelity Consistency Fidelity mIOU Consistency Fidelity Consistency Fidelity MaxPool 70.03 95.17 66.65 95.42 66.82 78.50 96.03 75.38 97.07 76.20 Blurpool LPF-3 71.02 95.52 67.84 96.03 68.20 78.90 96.09 75.82 97.94 77.27 DDAC LPF-3 72.28 96.77 69.95 95.98 69.37 79.52 96.28 76.54 98.21 78.09 APS LPF-3 72.37 97.05 70.24 96.70 69.98 79.84 97.53 77.87 98.32 78.50 LPS LPF-3 72.37 97.98 70.92 100.00 72.37 80.15 98.60 79.03 100.00 80.15 TIPS LPF-3 73.84 98.65 72.84 100.00 73.84 81.37 99.02 80.57 100.00 81.37 Table 4: Semantic segmentation performance on Pascal VOC and Cityscapes datasets. Kvasir U-Net CVC-ClinicDB U-Net Unshifted Standard Shift Circular Shift Unshifted Standard Shift Circular Shift Method Anti-Alias mIOU Consistency Fidelity Consistency Fidelity mIOU Consistency Fidelity Consistency Fidelity MaxPool 75.60 92.84 70.19 97.91 74.02 73.81 90.24 66.61 95.50 70.50 Blurpool LPF-3 78.39 94.63 74.18 98.30 77.06 76.32 93.87 71.64 96.36 73.54 DDAC LPF-3 79.24 95.17 75.41 98.49 78.04 77.89 92.17 71.80 97.73 76.12 APS LPF-3 81.97 96.32 78.95 100.00 81.97 79.31 95.63 75.84 100.00 79.31 LPS LPF-3 82.38 97.86 80.62 100.00 82.38 78.59 96.21 75.61 100.00 78.59 TIPS LPF-3 86.10 98.09 84.46 100.00 86.10 80.05 97.89 78.36 100.00 80.05 Table 5: Semantic segmentation performance on Kvasir and CVC-ClinicDB datasets. Figure 6: Qualitative comparison of segmentation masks predicted on original and shifted images. Images from Cityscapes, Pascal VOC are standard-shifted by (43,-17), (-38,0) respectively. Regions where TIPS achieve improvements (i.e. consistent segmentation quality) under linear shifts are highlighted with circles. Hyperparameters. We use SGD optimizer with a initial learning rate 0.01, momentum 0.9, weight decay 5e-4 with early stopping. We use DeepLabV3+ (Chen et al., 2018) with ResNet-18 as the backbone for the Pascal-VOC dataset and with ResNet-101 as the backbone for the Cityscapes dataset. For Kvasir and CVC-ClinicDB, we use a UNet (Ronneberger et al., 2015) model with \u201cKaiming Normal\u201d initialization. Evaluation Metrics. To report shift invariance for semantic segmentation, we use consistency and fidelity similar to image classification experiments, by comparing a common cropped area among images with different shift amounts. Within the common crop, we compute the percentage of pixels that have identical predictions in terms of segmentation categories. Results. Comparison of mIOU, consistency and fidelity in Tables 4, 5 shows that TIPS improves mIOU in comparison to all baselines on all four benchmarks. Consistent with our finding in image classification, we observe a sharper increase in shift consistency under standard shift than with circular. Models trained with TIPS pooling have higher fidelity on both standard and circular shifts, depicting the efficacy of TIPS in learning both shift invariant and high quality segmentation. In Figure 6, we compare the quality of the masks predicted on shifted image when using prior work or TIPS. The areas highlighted with red circles in the first two rows (Cityscapes) demonstrate that TIPS segments objects with higher consistency than other pooling operators under image shifts. The yellow boxes in the last two rows (Pascal-VOC) further illustrate improved segmentation consistency with TIPS under small shifts. 8 Unshifted Standard Shift Circular Shift Stategy Method Acc. Consistency Fidelity Consistency Fidelity Pooling MaxPool 64.88 82.41 53.14 80.39 50.71 (without anti-aliasing) DDAC 67.59 85.43 57.74 80.90 54.68 APS 67.05 86.39 57.92 100.00 67.05 LPS 67.39 86.17 58.07 100.00 67.39 TIPS 69.02 87.42 60.34 100.00 69.02 Pooling BlurPool (LPF-5) 66.85 87.43 58.54 87.88 58.75 (with LPF-5) DDAC (LPF-5) 66.98 86.92 58.22 80.35 53.82 (anti-aliasing) APS (LPF-5) 67.52 87.02 58.76 99.98 67.51 LPS (LPF-5) 69.11 86.58 59.84 100.00 69.11 TIPS (LPF-5) 70.01 87.51 61.27 100.00 70.01 Data Augmentation circular 64.25 83.58 53.71 84.27 54.14 standard 63.91 84.45 53.97 81.27 51.94 both 64.87 84.99 55.13 85.64 55.55 Contrastive Learning SimCLR 71.15 85.63 60.93 78.26 55.68 SupCon 72.49 86.17 62.46 81.75 59.26 Table 6: A comparison of accuracy and shift consistency and fidelity for additional methods including data augmentation, contrastive learning, and pooling with or without anti-aliasing. The models are trained on the ImageNet dataset with a ResNet18 backbone. Best performance in each section of the table is in bold, performance lower than MaxPool is highlighted in red and overall best performance is in cyan . 6 Analysis We further investigate the effectiveness of TIPS by comparing with non-pooling strategies, conducting ablation studies to examine the impact of our novel loss functions, understanding the effect of hyperparameters, and evaluating the effect of TIPS on various measures of robustness. 6.1 Investigating Other Strategies for Improving Shift Invariance of CNNs In Section 5 we compared TIPS with different pooling methods. There are other approaches besides pooling that could be useful for mitigating failures with pixel-level shift such as data augmentation and contrastive learning. To understand the efficacy of these approaches and compare them with TIPS and other pooling operators, we experiment with three types of data augmentation while training: standard shift, circular shift, and their combination, and two contrastive learning approaches: self-supervised SimCLR (Chen et al., 2020) and supervised SupCon (Khosla et al., 2020). For contrastive learning, representations are learned via the contrastive objective of aligning shifted samples closer and are used for downstream image classification. Table 6 shows a comparison of these techniques with TIPS and previous pooling-based approaches, for image classification with a ResNet-18 backbone, evaluated on the ImageNet dataset. While data augmentation does not significantly improve performance, both contrastive learning methods outperform MaxPool on all metrics. However, TIPS, without any contrastive learning or data augmentation, results in greater shift invariance. 6.2 Effect of Lundo and LF M Regularization Figure 7: The effect of Lundo in terms of |\u03c8(X) \u2212 Xt| and example feature maps (ResNet-101 with TIPS trained for 90 epochs on ImageNet; \u03f5=0.4). In Figure 7 we analyze the impact of Lundo on learning shift invariant intermediate features, by visualizing |\u03c8(X)\u2212Xt| at different stages of training. We observe that as training progresses, Lundo is able to guide |\u03c8(X) \u2212Xt| closer to 0 and thus TIPS learns to offset standard shift transformation on intermediate feature maps. In Figure 8 we quantify this observation by plotting accuracy and four shift invariance measurements for different values of \u03f5 in 9 Figure 8: Inspecting the effect of training varying % of epochs on Lundo for Tiny ImageNet classification. Lower \u03f5 indicates more epochs with Lundo and vice versa. \u03f5 = 0.4 (i.e. training without Lundo for the first 40% of epochs and with Lundo for the rest of the epochs is optimal. Values of \u03f5 higher than 0.4 yields sub-optimal shift invariance, but is better than low values of \u03f5, demonstrating the impact of Lundo. Figure 9: The impact of each component of our training objective is quantified through an ablation study. The top row shows results for six image classification datasets and the bottom row shows results for four semantic segmentation datasets. Regularization using both Lundo and LF M results in the best performance. Equation 5. For accuracy and both shift fidelity metrics, Lundo helps, but there is an optimal value of \u03f5 (=0.4 for Tiny ImageNet). A very low value of \u03f5 hurts performance. The training objective in Equation 5 includes task loss Ltask and two regularizations LF M and Lundo. In Figure 9, we perform an ablation study to examine the efficacy of each term in the loss function. Our results reveal a clear trend: \u25e6> \u22c6> \u25b3> \u25a1; across all datasets for image classification and segmentation, regularizing with both Lundo and LF M (denoted by \u25e6in the plots) leads to the highest accuracy, highest shift invariance in terms of all four evaluation metrics, and lowest MSB. Only using one of Lundo or LF M also improves performance compared to training only with Ltask. These results demonstrate the impact of each component of our loss function on shift invariance and further demonstrate the inverse relationship between MSB and shift invariance. 6.3 Effect of the Number of TIPS Layers Figure 10: The diameter of the bubbles denotes MSB. As the number of downsampling layers increases, MSB decreases and shift invariance increases. Figure 10 (a) portrays mean MSB, consistency, fidelity on Tiny ImageNet classification and Figure 10 (b) shows shift consistency and fidelity on Pascal VOC for semantic segmentation. As we train with more TIPS layer, shift invariance does not always strictly improve, in fact sometimes it decreases. However, for both Tiny ImageNet classification and Pascal VOC semantic segmentation, MSB always decreases as we train with more TIPS layers, indicating the efficacy of using TIPS in reducing MSB. 10 Figure 11: Shift invariance (consistency and fidelity) for standard and circular shifts on image classification (ImageNet) and semantic segmentation (Cityscapes) for varying degrees of shift. TIPS outperform existing pooling operators in all the evaluation metrics for shift invariance. Figure 12: Adversarial robustness under different levels \u03b5 of input perturbations. 6.4 Finegrained Results for Different Levels of Shift In our experiments the level of pixel shift, d is sampled from the range {0, 1, ..., D} where D = h 8 or w 8 for vertical and horizontal shift. In Figure 11, we demonstrate shift invariance under all possible levels of shifts d \u2208{0, 1, ..., D}, and observe that shift consistency drops faster with higher degrees of shift when using existing pooling methods whereas with TIPS this degradation is much slower. TIPS not only outperforms other pooling methods on average but at all degrees of shifts \u2208{0, 1, ..., D}. We observe that gain with TIPS in comparison to existing pooling operators is higher for shift fidelity than shift consistency. This suggests that TIPS improves both downstream task performance and shift invariance simultaneously. 6.5 Robustness Evaluation Adversarial Attacks. Recent studies reveal that deep models with ReLU are vulnerable against adversarial attacks if they are optimized for domain generalization (Frei et al., 2023) or shift invariance (Singla et al., 2021). Studies have also revealed a trade-off between adversarial robustness and other forms of generalization (Gokhale et al., 2022; Moayeri et al., 2022; Teney et al., 2024). We investigate the \u21132 and \u2113\u221e adversarial robustness of TIPS (with ResNet-34 backbone trained on CIFAR-10 and Tiny-ImageNet) using PGD (Madry et al., 2018) and FGSM(Goodfellow et al., 2014) attacks from Foolbox (Rauber et al., 2017). Figure 12 shows that TIPS exhibits superior adversarial robustness compared to previous methods. We also observe that better shift invariance is generally correlated with better adversarial robustness Patch Attacks. We adopt the experiment setup from Chaman & Dokmanic (2021) where square patches are randomly erased from the input image and test models trained on the clean CIFAR-10 and ImageNet datasets using a ResNet-18 backbone. Figure 13 demonstrates that TIPS outperforms other methods (pooling and 11 Figure 13: Evaluation of shift invariance under patch attacks (randomly erasing image patches) shows that TIPS exhibits higher robustness than existing pooling and data augmentation methods. Noise Blur Weather Digital Method Clean mCE Gauss. Shot Impulse Defocus Glass Motion Zoom Snow Frost Fog Bright Contrast Elastic Pixel JPEG VGG-19 25.8 81.6 82.0 83.0 88.0 82.0 94.0 84.0 86.0 80.0 78.0 69.0 61.0 74.0 94.0 85.0 83.0 VGG-19+TIPS 25.1 81.1 82.1 83.5 86.9 82.1 93.5 82.2 86.7 80.0 77.2 68.3 60.1 74.4 93.8 83.7 82.2 ResNet-18 30.2 84.7 87.0 88.0 91.0 84.0 91.0 87.0 89.0 86.0 84.0 78.0 69.0 78.0 90.0 80.0 85.0 ResNet-18+TIPS 28.7 83.9 85.3 87.9 91.6 83.6 91.2 85.7 88.3 85.4 82.6 77.1 68.5 77.3 89.9 80.1 84.6 Table 7: Errors on clean (ImageNet) and corrupted (ImageNet-C) test sets. mCE is the mean corruption error. Models are trained only on clean ImageNet training dataset. data augmentation) in robustness to such patch attacks. On ImageNet, shift consistency is more pronounced than other methods, especially for larger erased patches. Natural Corruptions. We evaluated the robustness of TIPS under an out-of-distribution setting, where models are trained on clean images, but tested on images with natural corruptions due to noise, blur, weather artifacts, or digital corruptions. We test robustness to natural corruptions using the ImageNet-C test dataset (Hendrycks & Dietterich, 2018) and report error on clean ImageNet (complement of classification accuracy). Table 7 shows that with TIPS, the mCE (mean corruption error) for VGG-19 and ResNet-18 architectures decreased by 0.61 % and 0.94 % respectively. 6.6 Applicability of TIPS to Vision Transformers Our work is focused on improving shift invariance of CNNs \u2013 models that are already in use in may real-world applications. We note that in vision tranformers, three modules break shift invariance: \u2022 Patch embeddings convert image patches into vectors using strided convolution (not shift invariant). \u2022 Positional encodings for both shifted and non-shifted inputs are identical (amount of shift is not encoded). \u2022 Window-based self-attention is computationally cheap, but applying local attention on windows of sizes larger than amount of input shift causes token values to change invariantly w.r.t. input shift. Since these mechanisms are not analogous to downsampling, polyphase sampling cannot be directly applied to ViTs as conveniently as CNNs. Although TIPS is currently limited to CNNs, in our experiments we show that ViTs are also not shift invariant and our simple plug-in solution for CNNs (TIPS) outperforms ViTs. 7 Conclusion Through a large scale correlation analysis we identify a strong inverse relationship of shift invariance of convolutional neural networks with the maximum-sampling bias (MSB) of pooling operators. We find that optimizing neural network weights to reduce MSB is a good strategy for improving shift invariance. With our proposed learnable Translation Invariant Polyphase Sampling (TIPS) pooling layer and regularization that promotes low MSB, we achieve state-of-the-art results for shift invariance on a variety of image classification and semantic segmentation benchmarks, outperforming data augmentation and contrastive learning strategies. Our analysis reveals additional benefits of TIPS, including improved robustness to adversarial attacks and corruptions. Our work serves as a starting point for further empirical or theoretical investigations into factors that cause sensitivity to shift. 12"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2008.09604v1",
+                    "title": "Delving Deeper into Anti-aliasing in ConvNets",
+                    "abstract": "Aliasing refers to the phenomenon that high frequency signals degenerate into\ncompletely different ones after sampling. It arises as a problem in the context\nof deep learning as downsampling layers are widely adopted in deep\narchitectures to reduce parameters and computation. The standard solution is to\napply a low-pass filter (e.g., Gaussian blur) before downsampling. However, it\ncan be suboptimal to apply the same filter across the entire content, as the\nfrequency of feature maps can vary across both spatial locations and feature\nchannels. To tackle this, we propose an adaptive content-aware low-pass\nfiltering layer, which predicts separate filter weights for each spatial\nlocation and channel group of the input feature maps. We investigate the\neffectiveness and generalization of the proposed method across multiple tasks\nincluding ImageNet classification, COCO instance segmentation, and Cityscapes\nsemantic segmentation. Qualitative and quantitative results demonstrate that\nour approach effectively adapts to the different feature frequencies to avoid\naliasing while preserving useful information for recognition. Code is available\nat https://maureenzou.github.io/ddac/.",
+                    "authors": "Xueyan Zou, Fanyi Xiao, Zhiding Yu, Yong Jae Lee",
+                    "published": "2020-08-21",
+                    "updated": "2020-08-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2307.09520v2",
+                    "title": "Adversarial Bayesian Augmentation for Single-Source Domain Generalization",
+                    "abstract": "Generalizing to unseen image domains is a challenging problem primarily due\nto the lack of diverse training data, inaccessible target data, and the large\ndomain shift that may exist in many real-world settings. As such data\naugmentation is a critical component of domain generalization methods that seek\nto address this problem. We present Adversarial Bayesian Augmentation (ABA), a\nnovel algorithm that learns to generate image augmentations in the challenging\nsingle-source domain generalization setting. ABA draws on the strengths of\nadversarial learning and Bayesian neural networks to guide the generation of\ndiverse data augmentations -- these synthesized image domains aid the\nclassifier in generalizing to unseen domains. We demonstrate the strength of\nABA on several types of domain shift including style shift, subpopulation\nshift, and shift in the medical imaging setting. ABA outperforms all previous\nstate-of-the-art methods, including pre-specified augmentations, pixel-based\nand convolutional-based augmentations.",
+                    "authors": "Sheng Cheng, Tejas Gokhale, Yezhou Yang",
+                    "published": "2023-07-18",
+                    "updated": "2023-10-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1705.09914v1",
+                    "title": "Dilated Residual Networks",
+                    "abstract": "Convolutional networks for image classification progressively reduce\nresolution until the image is represented by tiny feature maps in which the\nspatial structure of the scene is no longer discernible. Such loss of spatial\nacuity can limit image classification accuracy and complicate the transfer of\nthe model to downstream applications that require detailed scene understanding.\nThese problems can be alleviated by dilation, which increases the resolution of\noutput feature maps without reducing the receptive field of individual neurons.\nWe show that dilated residual networks (DRNs) outperform their non-dilated\ncounterparts in image classification without increasing the model's depth or\ncomplexity. We then study gridding artifacts introduced by dilation, develop an\napproach to removing these artifacts (`degridding'), and show that this further\nincreases the performance of DRNs. In addition, we show that the accuracy\nadvantage of DRNs is further magnified in downstream applications such as\nobject localization and semantic segmentation.",
+                    "authors": "Fisher Yu, Vladlen Koltun, Thomas Funkhouser",
+                    "published": "2017-05-28",
+                    "updated": "2017-05-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1712.02779v4",
+                    "title": "Exploring the Landscape of Spatial Robustness",
+                    "abstract": "The study of adversarial robustness has so far largely focused on\nperturbations bound in p-norms. However, state-of-the-art models turn out to be\nalso vulnerable to other, more natural classes of perturbations such as\ntranslations and rotations. In this work, we thoroughly investigate the\nvulnerability of neural network--based classifiers to rotations and\ntranslations. While data augmentation offers relatively small robustness, we\nuse ideas from robust optimization and test-time input aggregation to\nsignificantly improve robustness. Finally we find that, in contrast to the\np-norm case, first-order methods cannot reliably find worst-case perturbations.\nThis highlights spatial robustness as a fundamentally different setting\nrequiring additional study. Code available at\nhttps://github.com/MadryLab/adversarial_spatial and\nhttps://github.com/MadryLab/spatial-pytorch.",
+                    "authors": "Logan Engstrom, Brandon Tran, Dimitris Tsipras, Ludwig Schmidt, Aleksander Madry",
+                    "published": "2017-12-07",
+                    "updated": "2019-09-16",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.CV",
+                        "cs.NE",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1906.11555v2",
+                    "title": "Effective Rotation-invariant Point CNN with Spherical Harmonics kernels",
+                    "abstract": "We present a novel rotation invariant architecture operating directly on\npoint cloud data. We demonstrate how rotation invariance can be injected into a\nrecently proposed point-based PCNN architecture, at all layers of the network,\nachieving invariance to both global shape transformations, and to local\nrotations on the level of patches or parts, useful when dealing with non-rigid\nobjects. We achieve this by employing a spherical harmonics based kernel at\ndifferent layers of the network, which is guaranteed to be invariant to rigid\nmotions. We also introduce a more efficient pooling operation for PCNN using\nspace-partitioning data-structures. This results in a flexible, simple and\nefficient architecture that achieves accurate results on challenging shape\nanalysis tasks including classification and segmentation, without requiring\ndata-augmentation, typically employed by non-invariant approaches.",
+                    "authors": "Adrien Poulenard, Marie-Julie Rakotosaona, Yann Ponty, Maks Ovsjanikov",
+                    "published": "2019-06-27",
+                    "updated": "2019-09-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2004.11362v5",
+                    "title": "Supervised Contrastive Learning",
+                    "abstract": "Contrastive learning applied to self-supervised representation learning has\nseen a resurgence in recent years, leading to state of the art performance in\nthe unsupervised training of deep image models. Modern batch contrastive\napproaches subsume or significantly outperform traditional contrastive losses\nsuch as triplet, max-margin and the N-pairs loss. In this work, we extend the\nself-supervised batch contrastive approach to the fully-supervised setting,\nallowing us to effectively leverage label information. Clusters of points\nbelonging to the same class are pulled together in embedding space, while\nsimultaneously pushing apart clusters of samples from different classes. We\nanalyze two possible versions of the supervised contrastive (SupCon) loss,\nidentifying the best-performing formulation of the loss. On ResNet-200, we\nachieve top-1 accuracy of 81.4% on the ImageNet dataset, which is 0.8% above\nthe best number reported for this architecture. We show consistent\noutperformance over cross-entropy on other datasets and two ResNet variants.\nThe loss shows benefits for robustness to natural corruptions and is more\nstable to hyperparameter settings such as optimizers and data augmentations.\nOur loss function is simple to implement, and reference TensorFlow code is\nreleased at https://t.ly/supcon.",
+                    "authors": "Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, Dilip Krishnan",
+                    "published": "2020-04-23",
+                    "updated": "2021-03-10",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.CV",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1602.07576v3",
+                    "title": "Group Equivariant Convolutional Networks",
+                    "abstract": "We introduce Group equivariant Convolutional Neural Networks (G-CNNs), a\nnatural generalization of convolutional neural networks that reduces sample\ncomplexity by exploiting symmetries. G-CNNs use G-convolutions, a new type of\nlayer that enjoys a substantially higher degree of weight sharing than regular\nconvolution layers. G-convolutions increase the expressive capacity of the\nnetwork without increasing the number of parameters. Group convolution layers\nare easy to use and can be implemented with negligible computational overhead\nfor discrete groups generated by translations, reflections and rotations.\nG-CNNs achieve state of the art results on CIFAR10 and rotated MNIST.",
+                    "authors": "Taco S. Cohen, Max Welling",
+                    "published": "2016-02-24",
+                    "updated": "2016-06-03",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.01806v3",
+                    "title": "Attribute-Guided Adversarial Training for Robustness to Natural Perturbations",
+                    "abstract": "While existing work in robust deep learning has focused on small pixel-level\nnorm-based perturbations, this may not account for perturbations encountered in\nseveral real-world settings. In many such cases although test data might not be\navailable, broad specifications about the types of perturbations (such as an\nunknown degree of rotation) may be known. We consider a setup where robustness\nis expected over an unseen test domain that is not i.i.d. but deviates from the\ntraining domain. While this deviation may not be exactly known, its broad\ncharacterization is specified a priori, in terms of attributes. We propose an\nadversarial training approach which learns to generate new samples so as to\nmaximize exposure of the classifier to the attributes-space, without having\naccess to the data from the test domain. Our adversarial training solves a\nmin-max optimization problem, with the inner maximization generating\nadversarial perturbations, and the outer minimization finding model parameters\nby optimizing the loss on adversarial perturbations generated from the inner\nmaximization. We demonstrate the applicability of our approach on three types\nof naturally occurring perturbations -- object-related shifts, geometric\ntransformations, and common image corruptions. Our approach enables deep neural\nnetworks to be robust against a wide range of naturally occurring\nperturbations. We demonstrate the usefulness of the proposed approach by\nshowing the robustness gains of deep neural networks trained using our\nadversarial training on MNIST, CIFAR-10, and a new variant of the CLEVR\ndataset.",
+                    "authors": "Tejas Gokhale, Rushil Anirudh, Bhavya Kailkhura, Jayaraman J. Thiagarajan, Chitta Baral, Yezhou Yang",
+                    "published": "2020-12-03",
+                    "updated": "2021-04-08",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1903.12261v1",
+                    "title": "Benchmarking Neural Network Robustness to Common Corruptions and Perturbations",
+                    "abstract": "In this paper we establish rigorous benchmarks for image classifier\nrobustness. Our first benchmark, ImageNet-C, standardizes and expands the\ncorruption robustness topic, while showing which classifiers are preferable in\nsafety-critical applications. Then we propose a new dataset called ImageNet-P\nwhich enables researchers to benchmark a classifier's robustness to common\nperturbations. Unlike recent robustness research, this benchmark evaluates\nperformance on common corruptions and perturbations not worst-case adversarial\nperturbations. We find that there are negligible changes in relative corruption\nrobustness from AlexNet classifiers to ResNet classifiers. Afterward we\ndiscover ways to enhance corruption and perturbation robustness. We even find\nthat a bypassed adversarial defense provides substantial common perturbation\nrobustness. Together our benchmarks may aid future work toward networks that\nrobustly generalize.",
+                    "authors": "Dan Hendrycks, Thomas Dietterich",
+                    "published": "2019-03-28",
+                    "updated": "2019-03-28",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.CV",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2008.09604v1",
+                    "title": "Delving Deeper into Anti-aliasing in ConvNets",
+                    "abstract": "Aliasing refers to the phenomenon that high frequency signals degenerate into\ncompletely different ones after sampling. It arises as a problem in the context\nof deep learning as downsampling layers are widely adopted in deep\narchitectures to reduce parameters and computation. The standard solution is to\napply a low-pass filter (e.g., Gaussian blur) before downsampling. However, it\ncan be suboptimal to apply the same filter across the entire content, as the\nfrequency of feature maps can vary across both spatial locations and feature\nchannels. To tackle this, we propose an adaptive content-aware low-pass\nfiltering layer, which predicts separate filter weights for each spatial\nlocation and channel group of the input feature maps. We investigate the\neffectiveness and generalization of the proposed method across multiple tasks\nincluding ImageNet classification, COCO instance segmentation, and Cityscapes\nsemantic segmentation. Qualitative and quantitative results demonstrate that\nour approach effectively adapts to the different feature frequencies to avoid\naliasing while preserving useful information for recognition. Code is available\nat https://maureenzou.github.io/ddac/.",
+                    "authors": "Xueyan Zou, Fanyi Xiao, Zhiding Yu, Yong Jae Lee",
+                    "published": "2020-08-21",
+                    "updated": "2020-08-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.05932v1",
+                    "title": "GIFT: Learning Transformation-Invariant Dense Visual Descriptors via Group CNNs",
+                    "abstract": "Finding local correspondences between images with different viewpoints\nrequires local descriptors that are robust against geometric transformations.\nAn approach for transformation invariance is to integrate out the\ntransformations by pooling the features extracted from transformed versions of\nan image. However, the feature pooling may sacrifice the distinctiveness of the\nresulting descriptors. In this paper, we introduce a novel visual descriptor\nnamed Group Invariant Feature Transform (GIFT), which is both discriminative\nand robust to geometric transformations. The key idea is that the features\nextracted from the transformed versions of an image can be viewed as a function\ndefined on the group of the transformations. Instead of feature pooling, we use\ngroup convolutions to exploit underlying structures of the extracted features\non the group, resulting in descriptors that are both discriminative and\nprovably invariant to the group of transformations. Extensive experiments show\nthat GIFT outperforms state-of-the-art methods on several benchmark datasets\nand practically improves the performance of relative pose estimation.",
+                    "authors": "Yuan Liu, Zehong Shen, Zhixuan Lin, Sida Peng, Hujun Bao, Xiaowei Zhou",
+                    "published": "2019-11-14",
+                    "updated": "2019-11-14",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.02525v1",
+                    "title": "Semantic Segmentation Alternative Technique: Segmentation Domain Generation",
+                    "abstract": "Detecting objects of interest in images was always a compelling task to\nautomate. In recent years this task was more and more explored using deep\nlearning techniques, mostly using region-based convolutional networks. In this\nproject we propose an alternative semantic segmentation technique making use of\nGenerative Adversarial Networks. We consider semantic segmentation to be a\ndomain transfer problem. Thus, we train a feed forward network (FFNN) to\nreceive as input a seed real image and generate as output its segmentation\nmask.",
+                    "authors": "Ana-Cristina Rogoz, Radu Muntean, Stefan Cobeli",
+                    "published": "2021-07-06",
+                    "updated": "2021-07-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2304.02713v1",
+                    "title": "NUMSnet: Nested-U Multi-class Segmentation network for 3D Medical Image Stacks",
+                    "abstract": "Semantic segmentation for medical 3D image stacks enables accurate volumetric\nreconstructions, computer-aided diagnostics and follow up treatment planning.\nIn this work, we present a novel variant of the Unet model called the NUMSnet\nthat transmits pixel neighborhood features across scans through nested layers\nto achieve accurate multi-class semantic segmentations with minimal training\ndata. We analyze the semantic segmentation performance of the NUMSnet model in\ncomparison with several Unet model variants to segment 3-7 regions of interest\nusing only 10% of images for training per Lung-CT and Heart-CT volumetric image\nstacks. The proposed NUMSnet model achieves up to 20% improvement in\nsegmentation recall with 4-9% improvement in Dice scores for Lung-CT stacks and\n2.5-10% improvement in Dice scores for Heart-CT stacks when compared to the\nUnet++ model. The NUMSnet model needs to be trained by ordered images around\nthe central scan of each volumetric stack. Propagation of image feature\ninformation from the 6 nested layers of the Unet++ model are found to have\nbetter computation and segmentation performances than propagation of all\nup-sampling layers in a Unet++ model. The NUMSnet model achieves comparable\nsegmentation performances to existing works, while being trained on as low as\n5\\% of the training images. Also, transfer learning allows faster convergence\nof the NUMSnet model for multi-class semantic segmentation from pathology in\nLung-CT images to cardiac segmentations in Heart-CT stacks. Thus, the proposed\nmodel can standardize multi-class semantic segmentation on a variety of\nvolumetric image stacks with minimal training dataset. This can significantly\nreduce the cost, time and inter-observer variabilities associated with\ncomputer-aided detections and treatment.",
+                    "authors": "Sohini Roychowdhury",
+                    "published": "2023-04-05",
+                    "updated": "2023-04-05",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.07898v5",
+                    "title": "ISLE: A Framework for Image Level Semantic Segmentation Ensemble",
+                    "abstract": "One key bottleneck of employing state-of-the-art semantic segmentation\nnetworks in the real world is the availability of training labels. Conventional\nsemantic segmentation networks require massive pixel-wise annotated labels to\nreach state-of-the-art prediction quality. Hence, several works focus on\nsemantic segmentation networks trained with only image-level annotations.\nHowever, when scrutinizing the results of state-of-the-art in more detail, we\nnotice that they are remarkably close to each other on average prediction\nquality, different approaches perform better in different classes while\nproviding low quality in others. To address this problem, we propose a novel\nframework, ISLE, which employs an ensemble of the \"pseudo-labels\" for a given\nset of different semantic segmentation techniques on a class-wise level.\nPseudo-labels are the pixel-wise predictions of the image-level semantic\nsegmentation frameworks used to train the final segmentation model. Our\npseudo-labels seamlessly combine the strong points of multiple segmentation\ntechniques approaches to reach superior prediction quality. We reach up to 2.4%\nimprovement over ISLE's individual components. An exhaustive analysis was\nperformed to demonstrate ISLE's effectiveness over state-of-the-art frameworks\nfor image-level semantic segmentation.",
+                    "authors": "Erik Ostrowski, Muhammad Shafique",
+                    "published": "2023-03-14",
+                    "updated": "2023-09-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1709.01956v1",
+                    "title": "Learning Dilation Factors for Semantic Segmentation of Street Scenes",
+                    "abstract": "Contextual information is crucial for semantic segmentation. However, finding\nthe optimal trade-off between keeping desired fine details and at the same time\nproviding sufficiently large receptive fields is non trivial. This is even more\nso, when objects or classes present in an image significantly vary in size.\nDilated convolutions have proven valuable for semantic segmentation, because\nthey allow to increase the size of the receptive field without sacrificing\nimage resolution. However, in current state-of-the-art methods, dilation\nparameters are hand-tuned and fixed. In this paper, we present an approach for\nlearning dilation parameters adaptively per channel, consistently improving\nsemantic segmentation results on street-scene datasets like Cityscapes and\nCamvid.",
+                    "authors": "Yang He, Margret Keuper, Bernt Schiele, Mario Fritz",
+                    "published": "2017-09-06",
+                    "updated": "2017-09-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.04297v1",
+                    "title": "SHDM-NET: Heat Map Detail Guidance with Image Matting for Industrial Weld Semantic Segmentation Network",
+                    "abstract": "In actual industrial production, the assessment of the steel plate welding\neffect is an important task, and the segmentation of the weld section is the\nbasis of the assessment. This paper proposes an industrial weld segmentation\nnetwork based on a deep learning semantic segmentation algorithm fused with\nheatmap detail guidance and Image Matting to solve the automatic segmentation\nproblem of weld regions. In the existing semantic segmentation networks, the\nboundary information can be preserved by fusing the features of both high-level\nand low-level layers. However, this method can lead to insufficient expression\nof the spatial information in the low-level layer, resulting in inaccurate\nsegmentation boundary positioning. We propose a detailed guidance module based\non heatmaps to fully express the segmented region boundary information in the\nlow-level network to address this problem. Specifically, the expression of\nboundary information can be enhanced by adding a detailed branch to predict\nsegmented boundary and then matching it with the boundary heat map generated by\nmask labels to calculate the mean square error loss. In addition, although deep\nlearning has achieved great success in the field of semantic segmentation, the\nprecision of the segmentation boundary region is not high due to the loss of\ndetailed information caused by the classical segmentation network in the\nprocess of encoding and decoding process. This paper introduces a matting\nalgorithm to calibrate the boundary of the segmentation region of the semantic\nsegmentation network to solve this problem. Through many experiments on\nindustrial weld data sets, the effectiveness of our method is demonstrated, and\nthe MIOU reaches 97.93%. It is worth noting that this performance is comparable\nto human manual segmentation ( MIOU 97.96%).",
+                    "authors": "Qi Wang, Jingwu Mei",
+                    "published": "2022-07-09",
+                    "updated": "2022-07-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2404.04807v1",
+                    "title": "D2SL: Decouple Defogging and Semantic Learning for Foggy Domain-Adaptive Segmentation",
+                    "abstract": "We investigated domain adaptive semantic segmentation in foggy weather\nscenarios, which aims to enhance the utilization of unlabeled foggy data and\nimprove the model's adaptability to foggy conditions. Current methods rely on\nclear images as references, jointly learning defogging and segmentation for\nfoggy images. Despite making some progress, there are still two main drawbacks:\n(1) the coupling of segmentation and defogging feature representations,\nresulting in a decrease in semantic representation capability, and (2) the\nfailure to leverage real fog priors in unlabeled foggy data, leading to\ninsufficient model generalization ability. To address these issues, we propose\na novel training framework, Decouple Defogging and Semantic learning, called\nD2SL, aiming to alleviate the adverse impact of defogging tasks on the final\nsegmentation task. In this framework, we introduce a domain-consistent transfer\nstrategy to establish a connection between defogging and segmentation tasks.\nFurthermore, we design a real fog transfer strategy to improve defogging\neffects by fully leveraging the fog priors from real foggy images. Our approach\nenhances the semantic representations required for segmentation during the\ndefogging learning process and maximizes the representation capability of fog\ninvariance by effectively utilizing real fog data. Comprehensive experiments\nvalidate the effectiveness of the proposed method.",
+                    "authors": "Xuan Sun, Zhanfu An, Yuyu Liu",
+                    "published": "2024-04-07",
+                    "updated": "2024-04-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.MM"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.05564v1",
+                    "title": "Hypergraph Convolutional Networks for Weakly-Supervised Semantic Segmentation",
+                    "abstract": "Semantic segmentation is a fundamental topic in computer vision. Several deep\nlearning methods have been proposed for semantic segmentation with outstanding\nresults. However, these models require a lot of densely annotated images. To\naddress this problem, we propose a new algorithm that uses HyperGraph\nConvolutional Networks for Weakly-supervised Semantic Segmentation\n(HyperGCN-WSS). Our algorithm constructs spatial and k-Nearest Neighbor (k-NN)\ngraphs from the images in the dataset to generate the hypergraphs. Then, we\ntrain a specialized HyperGraph Convolutional Network (HyperGCN) architecture\nusing some weak signals. The outputs of the HyperGCN are denominated\npseudo-labels, which are later used to train a DeepLab model for semantic\nsegmentation. HyperGCN-WSS is evaluated on the PASCAL VOC 2012 dataset for\nsemantic segmentation, using scribbles or clicks as weak signals. Our algorithm\nshows competitive performance against previous methods.",
+                    "authors": "Jhony H. Giraldo, Vincenzo Scarrica, Antonino Staiano, Francesco Camastra, Thierry Bouwmans",
+                    "published": "2022-10-11",
+                    "updated": "2022-10-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.06974v1",
+                    "title": "A One Stop 3D Target Reconstruction and multilevel Segmentation Method",
+                    "abstract": "3D object reconstruction and multilevel segmentation are fundamental to\ncomputer vision research. Existing algorithms usually perform 3D scene\nreconstruction and target objects segmentation independently, and the\nperformance is not fully guaranteed due to the challenge of the 3D\nsegmentation. Here we propose an open-source one stop 3D target reconstruction\nand multilevel segmentation framework (OSTRA), which performs segmentation on\n2D images, tracks multiple instances with segmentation labels in the image\nsequence, and then reconstructs labelled 3D objects or multiple parts with\nMulti-View Stereo (MVS) or RGBD-based 3D reconstruction methods. We extend\nobject tracking and 3D reconstruction algorithms to support continuous\nsegmentation labels to leverage the advances in the 2D image segmentation,\nespecially the Segment-Anything Model (SAM) which uses the pretrained neural\nnetwork without additional training for new scenes, for 3D object segmentation.\nOSTRA supports most popular 3D object models including point cloud, mesh and\nvoxel, and achieves high performance for semantic segmentation, instance\nsegmentation and part segmentation on several 3D datasets. It even surpasses\nthe manual segmentation in scenes with complex structures and occlusions. Our\nmethod opens up a new avenue for reconstructing 3D targets embedded with rich\nmulti-scale segmentation information in complex scenes. OSTRA is available from\nhttps://github.com/ganlab/OSTRA.",
+                    "authors": "Jiexiong Xu, Weikun Zhao, Zhiyan Tang, Xiangchao Gan",
+                    "published": "2023-08-14",
+                    "updated": "2023-08-14",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2211.08352v1",
+                    "title": "Visual Semantic Segmentation Based on Few/Zero-Shot Learning: An Overview",
+                    "abstract": "Visual semantic segmentation aims at separating a visual sample into diverse\nblocks with specific semantic attributes and identifying the category for each\nblock, and it plays a crucial role in environmental perception. Conventional\nlearning-based visual semantic segmentation approaches count heavily on\nlarge-scale training data with dense annotations and consistently fail to\nestimate accurate semantic labels for unseen categories. This obstruction spurs\na craze for studying visual semantic segmentation with the assistance of\nfew/zero-shot learning. The emergence and rapid progress of few/zero-shot\nvisual semantic segmentation make it possible to learn unseen-category from a\nfew labeled or zero-labeled samples, which advances the extension to practical\napplications. Therefore, this paper focuses on the recently published\nfew/zero-shot visual semantic segmentation methods varying from 2D to 3D space\nand explores the commonalities and discrepancies of technical settlements under\ndifferent segmentation circumstances. Specifically, the preliminaries on\nfew/zero-shot visual semantic segmentation, including the problem definitions,\ntypical datasets, and technical remedies, are briefly reviewed and discussed.\nMoreover, three typical instantiations are involved to uncover the interactions\nof few/zero-shot learning with visual semantic segmentation, including image\nsemantic segmentation, video object segmentation, and 3D segmentation. Finally,\nthe future challenges of few/zero-shot visual semantic segmentation are\ndiscussed.",
+                    "authors": "Wenqi Ren, Yang Tang, Qiyu Sun, Chaoqiang Zhao, Qing-Long Han",
+                    "published": "2022-11-13",
+                    "updated": "2022-11-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.10122v2",
+                    "title": "Weakly-supervised Semantic Segmentation in Cityscape via Hyperspectral Image",
+                    "abstract": "High-resolution hyperspectral images (HSIs) contain the response of each\npixel in different spectral bands, which can be used to effectively distinguish\nvarious objects in complex scenes. While HSI cameras have become low cost,\nalgorithms based on it have not been well exploited. In this paper, we focus on\na novel topic, weakly-supervised semantic segmentation in cityscape via HSIs.\nIt is based on the idea that high-resolution HSIs in city scenes contain rich\nspectral information, which can be easily associated to semantics without\nmanual labeling. Therefore, it enables low cost, highly reliable semantic\nsegmentation in complex scenes. Specifically, in this paper, we theoretically\nanalyze the HSIs and introduce a weakly-supervised HSI semantic segmentation\nframework, which utilizes spectral information to improve the coarse labels to\na finer degree. The experimental results show that our method can obtain highly\ncompetitive labels and even have higher edge fineness than artificial fine\nlabels in some classes. At the same time, the results also show that the\nrefined labels can effectively improve the effect of semantic segmentation. The\ncombination of HSIs and semantic segmentation proves that HSIs have great\npotential in high-level visual tasks.",
+                    "authors": "Yuxing Huang, Shaodi You, Ying Fu, Qiu Shen",
+                    "published": "2020-12-18",
+                    "updated": "2021-07-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2006.07601v1",
+                    "title": "NoPeopleAllowed: The Three-Step Approach to Weakly Supervised Semantic Segmentation",
+                    "abstract": "We propose a novel approach to weakly supervised semantic segmentation, which\nconsists of three consecutive steps. The first two steps extract high-quality\npseudo masks from image-level annotated data, which are then used to train a\nsegmentation model on the third step. The presented approach also addresses two\nproblems in the data: class imbalance and missing labels. Using only\nimage-level annotations as supervision, our method is capable of segmenting\nvarious classes and complex objects. It achieves 37.34 mean IoU on the test\nset, placing 3rd at the LID Challenge in the task of weakly supervised semantic\nsegmentation.",
+                    "authors": "Mariia Dobko, Ostap Viniavskyi, Oles Dobosevych",
+                    "published": "2020-06-13",
+                    "updated": "2020-06-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.00487v1",
+                    "title": "Linear Semantics in Generative Adversarial Networks",
+                    "abstract": "Generative Adversarial Networks (GANs) are able to generate high-quality\nimages, but it remains difficult to explicitly specify the semantics of\nsynthesized images. In this work, we aim to better understand the semantic\nrepresentation of GANs, and thereby enable semantic control in GAN's generation\nprocess. Interestingly, we find that a well-trained GAN encodes image semantics\nin its internal feature maps in a surprisingly simple way: a linear\ntransformation of feature maps suffices to extract the generated image\nsemantics. To verify this simplicity, we conduct extensive experiments on\nvarious GANs and datasets; and thanks to this simplicity, we are able to learn\na semantic segmentation model for a trained GAN from a small number (e.g., 8)\nof labeled images. Last but not least, leveraging our findings, we propose two\nfew-shot image editing approaches, namely Semantic-Conditional Sampling and\nSemantic Image Editing. Given a trained GAN and as few as eight semantic\nannotations, the user is able to generate diverse images subject to a\nuser-provided semantic layout, and control the synthesized image semantics. We\nhave made the code publicly available.",
+                    "authors": "Jianjin Xu, Changxi Zheng",
+                    "published": "2021-04-01",
+                    "updated": "2021-04-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1511.06988v1",
+                    "title": "Learning High-level Prior with Convolutional Neural Networks for Semantic Segmentation",
+                    "abstract": "This paper proposes a convolutional neural network that can fuse high-level\nprior for semantic image segmentation. Motivated by humans' vision recognition\nsystem, our key design is a three-layer generative structure consisting of\nhigh-level coding, middle-level segmentation and low-level image to introduce\nglobal prior for semantic segmentation. Based on this structure, we proposed a\ngenerative model called conditional variational auto-encoder (CVAE) that can\nbuild up the links behind these three layers. These important links include an\nimage encoder that extracts high level info from image, a segmentation encoder\nthat extracts high level info from segmentation, and a hybrid decoder that\noutputs semantic segmentation from the high level prior and input image. We\ntheoretically derive the semantic segmentation as an optimization problem\nparameterized by these links. Finally, the optimization problem enables us to\ntake advantage of state-of-the-art fully convolutional network structure for\nthe implementation of the above encoders and decoder. Experimental results on\nseveral representative datasets demonstrate our supreme performance for\nsemantic segmentation.",
+                    "authors": "Haitian Zheng, Yebin Liu, Mengqi Ji, Feng Wu, Lu Fang",
+                    "published": "2015-11-22",
+                    "updated": "2015-11-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.00949v3",
+                    "title": "Synthetic Instance Segmentation from Semantic Image Segmentation Masks",
+                    "abstract": "In recent years, the development of instance segmentation has garnered\nsignificant attention in a wide range of applications. However, the training of\na fully-supervised instance segmentation model requires costly both\ninstance-level and pixel-level annotations. In contrast, weakly-supervised\ninstance segmentation methods (i.e., with image-level class labels or point\nlabels) struggle to satisfy the accuracy and recall requirements of practical\nscenarios. In this paper, we propose a novel paradigm called synthetic instance\nsegmentation (SISeg), which achieves Instance Segmentation results from image\nmasks predicted using off-the-shelf semantic segmentation models. SISeg does\nnot require training a semantic or/and instance segmentation model and avoids\nthe need for instance-level image annotations. Therefore, it is highly\nefficient. Specifically, we first obtain a semantic segmentation mask of the\ninput image via a trained semantic segmentation model. Then, we calculate a\ndisplacement field vector for each pixel based on the segmentation mask, which\ncan indicate representations belonging to the same class but different\ninstances, i.e., obtaining the instance-level object information. Finally,\ninstance segmentation results are obtained after being refined by a learnable\ncategory-agnostic object boundary branch. Extensive experimental results on two\nchallenging datasets and representative semantic segmentation baselines\n(including CNNs and Transformers) demonstrate that SISeg can achieve\ncompetitive results compared to the state-of-the-art fully-supervised instance\nsegmentation methods without the need for additional human resources or\nincreased computational costs. The code is available at: SISeg",
+                    "authors": "Yuchen Shen, Dong Zhang, Yuhui Zheng, Zechao Li, Liyong Fu, Qiaolin Ye",
+                    "published": "2023-08-02",
+                    "updated": "2023-10-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.02094v1",
+                    "title": "Segment Anything Meets Semantic Communication",
+                    "abstract": "In light of the diminishing returns of traditional methods for enhancing\ntransmission rates, the domain of semantic communication presents promising new\nfrontiers. Focusing on image transmission, this paper explores the application\nof foundation models, particularly the Segment Anything Model (SAM) developed\nby Meta AI Research, to improve semantic communication. SAM is a promptable\nimage segmentation model that has gained attention for its ability to perform\nzero-shot segmentation tasks without explicit training or domain-specific\nknowledge. By employing SAM's segmentation capability and lightweight neural\nnetwork architecture for semantic coding, we propose a practical approach to\nsemantic communication. We demonstrate that this approach retains critical\nsemantic features, achieving higher image reconstruction quality and reducing\ncommunication overhead. This practical solution eliminates the\nresource-intensive stage of training a segmentation model and can be applied to\nany semantic coding architecture, paving the way for real-world applications.",
+                    "authors": "Shehbaz Tariq, Brian Estadimas Arfeto, Chaoning Zhang, Hyundong Shin",
+                    "published": "2023-06-03",
+                    "updated": "2023-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2205.03186v1",
+                    "title": "Semantics-Guided Moving Object Segmentation with 3D LiDAR",
+                    "abstract": "Moving object segmentation (MOS) is a task to distinguish moving objects,\ne.g., moving vehicles and pedestrians, from the surrounding static environment.\nThe segmentation accuracy of MOS can have an influence on odometry, map\nconstruction, and planning tasks. In this paper, we propose a semantics-guided\nconvolutional neural network for moving object segmentation. The network takes\nsequential LiDAR range images as inputs. Instead of segmenting the moving\nobjects directly, the network conducts single-scan-based semantic segmentation\nand multiple-scan-based moving object segmentation in turn. The semantic\nsegmentation module provides semantic priors for the MOS module, where we\npropose an adjacent scan association (ASA) module to convert the semantic\nfeatures of adjacent scans into the same coordinate system to fully exploit the\ncross-scan semantic features. Finally, by analyzing the difference between the\ntransformed features, reliable MOS result can be obtained quickly. Experimental\nresults on the SemanticKITTI MOS dataset proves the effectiveness of our work.",
+                    "authors": "Shuo Gu, Suling Yao, Jian Yang, Hui Kong",
+                    "published": "2022-05-06",
+                    "updated": "2022-05-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.RO"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1803.10464v2",
+                    "title": "Learning Pixel-level Semantic Affinity with Image-level Supervision for Weakly Supervised Semantic Segmentation",
+                    "abstract": "The deficiency of segmentation labels is one of the main obstacles to\nsemantic segmentation in the wild. To alleviate this issue, we present a novel\nframework that generates segmentation labels of images given their image-level\nclass labels. In this weakly supervised setting, trained models have been known\nto segment local discriminative parts rather than the entire object area. Our\nsolution is to propagate such local responses to nearby areas which belong to\nthe same semantic entity. To this end, we propose a Deep Neural Network (DNN)\ncalled AffinityNet that predicts semantic affinity between a pair of adjacent\nimage coordinates. The semantic propagation is then realized by random walk\nwith the affinities predicted by AffinityNet. More importantly, the supervision\nemployed to train AffinityNet is given by the initial discriminative part\nsegmentation, which is incomplete as a segmentation annotation but sufficient\nfor learning semantic affinities within small image areas. Thus the entire\nframework relies only on image-level class labels and does not require any\nextra data or annotations. On the PASCAL VOC 2012 dataset, a DNN learned with\nsegmentation labels generated by our method outperforms previous models trained\nwith the same level of supervision, and is even as competitive as those relying\non stronger supervision.",
+                    "authors": "Jiwoon Ahn, Suha Kwak",
+                    "published": "2018-03-28",
+                    "updated": "2018-04-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.01369v2",
+                    "title": "Exploring Limits of Diffusion-Synthetic Training with Weakly Supervised Semantic Segmentation",
+                    "abstract": "The advance of generative models for images has inspired various training\ntechniques for image recognition utilizing synthetic images. In semantic\nsegmentation, one promising approach is extracting pseudo-masks from attention\nmaps in text-to-image diffusion models, which enables\nreal-image-and-annotation-free training. However, the pioneering training\nmethod using the diffusion-synthetic images and pseudo-masks, i.e., DiffuMask\nhas limitations in terms of mask quality, scalability, and ranges of applicable\ndomains. To overcome these limitations, this work introduces three techniques\nfor diffusion-synthetic semantic segmentation training. First,\nreliability-aware robust training, originally used in weakly supervised\nlearning, helps segmentation with insufficient synthetic mask quality. %Second,\nlarge-scale pretraining of whole segmentation models, not only backbones, on\nsynthetic ImageNet-1k-class images with pixel-labels benefits downstream\nsegmentation tasks. Second, we introduce prompt augmentation, data augmentation\nto the prompt text set to scale up and diversify training images with a limited\ntext resources. Finally, LoRA-based adaptation of Stable Diffusion enables the\ntransfer to a distant domain, e.g., auto-driving images. Experiments in PASCAL\nVOC, ImageNet-S, and Cityscapes show that our method effectively closes gap\nbetween real and synthetic training in semantic segmentation.",
+                    "authors": "Ryota Yoshihashi, Yuya Otsuka, Kenji Doi, Tomohiro Tanaka, Hirokatsu Kataoka",
+                    "published": "2023-09-04",
+                    "updated": "2024-04-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1812.10885v1",
+                    "title": "Coarse-to-fine Semantic Segmentation from Image-level Labels",
+                    "abstract": "Deep neural network-based semantic segmentation generally requires\nlarge-scale cost extensive annotations for training to obtain better\nperformance. To avoid pixel-wise segmentation annotations which are needed for\nmost methods, recently some researchers attempted to use object-level labels\n(e.g. bounding boxes) or image-level labels (e.g. image categories). In this\npaper, we propose a novel recursive coarse-to-fine semantic segmentation\nframework based on only image-level category labels. For each image, an initial\ncoarse mask is first generated by a convolutional neural network-based\nunsupervised foreground segmentation model and then is enhanced by a graph\nmodel. The enhanced coarse mask is fed to a fully convolutional neural network\nto be recursively refined. Unlike existing image-level label-based semantic\nsegmentation methods which require to label all categories for images contain\nmultiple types of objects, our framework only needs one label for each image\nand can handle images contains multi-category objects. With only trained on\nImageNet, our framework achieves comparable performance on PASCAL VOC dataset\nas other image-level label-based state-of-the-arts of semantic segmentation.\nFurthermore, our framework can be easily extended to foreground object\nsegmentation task and achieves comparable performance with the state-of-the-art\nsupervised methods on the Internet Object dataset.",
+                    "authors": "Longlong Jing, Yucheng Chen, Yingli Tian",
+                    "published": "2018-12-28",
+                    "updated": "2018-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1809.10245v1",
+                    "title": "Cylindrical Transform: 3D Semantic Segmentation of Kidneys With Limited Annotated Images",
+                    "abstract": "In this paper, we propose a novel technique for sampling sequential images\nusing a cylindrical transform in a cylindrical coordinate system for kidney\nsemantic segmentation in abdominal computed tomography (CT). The images\ngenerated from a cylindrical transform augment a limited annotated set of\nimages in three dimensions. This approach enables us to train contemporary\nclassification deep convolutional neural networks (DCNNs) instead of fully\nconvolutional networks (FCNs) for semantic segmentation. Typical semantic\nsegmentation models segment a sequential set of images (e.g. CT or video) by\nsegmenting each image independently. However, the proposed method not only\nconsiders the spatial dependency in the x-y plane, but also the spatial\nsequential dependency along the z-axis. The results show that classification\nDCNNs, trained on cylindrical transformed images, can achieve a higher\nsegmentation performance value than FCNs using a limited number of annotated\nimages.",
+                    "authors": "Hojjat Salehinejad, Sumeya Naqvi, Errol Colak, Joseph Barfett, Shahrokh Valaee",
+                    "published": "2018-09-24",
+                    "updated": "2018-09-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.NE"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2201.05869v2",
+                    "title": "Prototype Guided Network for Anomaly Segmentation",
+                    "abstract": "Semantic segmentation methods can not directly identify abnormal objects in\nimages. Anomaly Segmentation algorithm from this realistic setting can\ndistinguish between in-distribution objects and Out-Of-Distribution (OOD)\nobjects and output the anomaly probability for pixels. In this paper, a\nPrototype Guided Anomaly segmentation Network (PGAN) is proposed to extract\nsemantic prototypes for in-distribution training data from limited annotated\nimages. In the model, prototypes are used to model the hierarchical category\nsemantic information and distinguish OOD pixels. The proposed PGAN model\nincludes a semantic segmentation network and a prototype extraction network.\nSimilarity measures are adopted to optimize the prototypes. The learned\nsemantic prototypes are used as category semantics to compare the similarity\nwith features extracted from test images and then to generate semantic\nsegmentation prediction. The proposed prototype extraction network can also be\nintegrated into most semantic segmentation networks and recognize OOD pixels.\nOn the StreetHazards dataset, the proposed PGAN model produced mIoU of 53.4%\nfor anomaly segmentation. The experimental results demonstrate PGAN may achieve\nthe SOTA performance in the anomaly segmentation tasks.",
+                    "authors": "Yiqing Hao, Yi Jin, Gaoyun An",
+                    "published": "2022-01-15",
+                    "updated": "2022-03-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.00242v1",
+                    "title": "3D Guided Weakly Supervised Semantic Segmentation",
+                    "abstract": "Pixel-wise clean annotation is necessary for fully-supervised semantic\nsegmentation, which is laborious and expensive to obtain. In this paper, we\npropose a weakly supervised 2D semantic segmentation model by incorporating\nsparse bounding box labels with available 3D information, which is much easier\nto obtain with advanced sensors. We manually labeled a subset of the 2D-3D\nSemantics(2D-3D-S) dataset with bounding boxes, and introduce our 2D-3D\ninference module to generate accurate pixel-wise segment proposal masks. Guided\nby 3D information, we first generate a point cloud of objects and calculate\nobjectness probability score for each point. Then we project the point cloud\nwith objectness probabilities back to 2D images followed by a refinement step\nto obtain segment proposals, which are treated as pseudo labels to train a\nsemantic segmentation network. Our method works in a recursive manner to\ngradually refine the above-mentioned segment proposals. Extensive experimental\nresults on the 2D-3D-S dataset show that the proposed method can generate\naccurate segment proposals when bounding box labels are available on only a\nsmall subset of training images. Performance comparison with recent\nstate-of-the-art methods further illustrates the effectiveness of our method.",
+                    "authors": "Weixuan Sun, Jing Zhang, Nick Barnes",
+                    "published": "2020-12-01",
+                    "updated": "2020-12-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.11857v1",
+                    "title": "Joint Learning of Intrinsic Images and Semantic Segmentation",
+                    "abstract": "Semantic segmentation of outdoor scenes is problematic when there are\nvariations in imaging conditions. It is known that albedo (reflectance) is\ninvariant to all kinds of illumination effects. Thus, using reflectance images\nfor semantic segmentation task can be favorable. Additionally, not only\nsegmentation may benefit from reflectance, but also segmentation may be useful\nfor reflectance computation. Therefore, in this paper, the tasks of semantic\nsegmentation and intrinsic image decomposition are considered as a combined\nprocess by exploring their mutual relationship in a joint fashion. To that end,\nwe propose a supervised end-to-end CNN architecture to jointly learn intrinsic\nimage decomposition and semantic segmentation. We analyze the gains of\naddressing those two problems jointly. Moreover, new cascade CNN architectures\nfor intrinsic-for-segmentation and segmentation-for-intrinsic are proposed as\nsingle tasks. Furthermore, a dataset of 35K synthetic images of natural\nenvironments is created with corresponding albedo and shading (intrinsics), as\nwell as semantic labels (segmentation) assigned to each object/scene. The\nexperiments show that joint learning of intrinsic image decomposition and\nsemantic segmentation is beneficial for both tasks for natural scenes. Dataset\nand models are available at: https://ivi.fnwi.uva.nl/cv/intrinseg",
+                    "authors": "Anil S. Baslamisli, Thomas T. Groenestege, Partha Das, Hoang-An Le, Sezer Karaoglu, Theo Gevers",
+                    "published": "2018-07-31",
+                    "updated": "2018-07-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2307.13215v1",
+                    "title": "Image Segmentation Keras : Implementation of Segnet, FCN, UNet, PSPNet and other models in Keras",
+                    "abstract": "Semantic segmentation plays a vital role in computer vision tasks, enabling\nprecise pixel-level understanding of images. In this paper, we present a\ncomprehensive library for semantic segmentation, which contains implementations\nof popular segmentation models like SegNet, FCN, UNet, and PSPNet. We also\nevaluate and compare these models on several datasets, offering researchers and\npractitioners a powerful toolset for tackling diverse segmentation challenges.",
+                    "authors": "Divam Gupta",
+                    "published": "2023-07-25",
+                    "updated": "2023-07-25",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.15608v1",
+                    "title": "Semantic Segmentation by Semantic Proportions",
+                    "abstract": "Semantic segmentation is a critical task in computer vision that aims to\nidentify and classify individual pixels in an image, with numerous applications\nfor example autonomous driving and medical image analysis. However, semantic\nsegmentation can be super challenging particularly due to the need for large\namounts of annotated data. Annotating images is a time-consuming and costly\nprocess, often requiring expert knowledge and significant effort. In this\npaper, we propose a novel approach for semantic segmentation by eliminating the\nneed of ground-truth segmentation maps. Instead, our approach requires only the\nrough information of individual semantic class proportions, shortened as\nsemantic proportions. It greatly simplifies the data annotation process and\nthus will significantly reduce the annotation time and cost, making it more\nfeasible for large-scale applications. Moreover, it opens up new possibilities\nfor semantic segmentation tasks where obtaining the full ground-truth\nsegmentation maps may not be feasible or practical. Extensive experimental\nresults demonstrate that our approach can achieve comparable and sometimes even\nbetter performance against the benchmark method that relies on the ground-truth\nsegmentation maps. Utilising semantic proportions suggested in this work offers\na promising direction for future research in the field of semantic\nsegmentation.",
+                    "authors": "Halil Ibrahim Aysel, Xiaohao Cai, Adam Pr\u00fcgel-Bennett",
+                    "published": "2023-05-24",
+                    "updated": "2023-05-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.05457v1",
+                    "title": "Instance Segmentation based Semantic Matting for Compositing Applications",
+                    "abstract": "Image compositing is a key step in film making and image editing that aims to\nsegment a foreground object and combine it with a new background. Automatic\nimage compositing can be done easily in a studio using chroma-keying when the\nbackground is pure blue or green. However, image compositing in natural scenes\nwith complex backgrounds remains a tedious task, requiring experienced artists\nto hand-segment. In order to achieve automatic compositing in natural scenes,\nwe propose a fully automated method that integrates instance segmentation and\nimage matting processes to generate high-quality semantic mattes that can be\nused for image editing task. Our approach can be seen both as a refinement of\nexisting instance segmentation algorithms and as a fully automated semantic\nimage matting method. It extends automatic image compositing techniques such as\nchroma-keying to scenes with complex natural backgrounds without the need for\nany kind of user interaction. The output of our approach can be considered as\nboth refined instance segmentations and alpha mattes with semantic meanings. We\nprovide experimental results which show improved performance results as\ncompared to existing approaches.",
+                    "authors": "Guanqing Hu, James J. Clark",
+                    "published": "2019-04-10",
+                    "updated": "2019-04-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2101.09642v2",
+                    "title": "Image Compression with Encoder-Decoder Matched Semantic Segmentation",
+                    "abstract": "In recent years, layered image compression is demonstrated to be a promising\ndirection, which encodes a compact representation of the input image and apply\nan up-sampling network to reconstruct the image. To further improve the quality\nof the reconstructed image, some works transmit the semantic segment together\nwith the compressed image data. Consequently, the compression ratio is also\ndecreased because extra bits are required for transmitting the semantic\nsegment. To solve this problem, we propose a new layered image compression\nframework with encoder-decoder matched semantic segmentation (EDMS). And then,\nfollowed by the semantic segmentation, a special convolution neural network is\nused to enhance the inaccurate semantic segment. As a result, the accurate\nsemantic segment can be obtained in the decoder without requiring extra bits.\nThe experimental results show that the proposed EDMS framework can get up to\n35.31% BD-rate reduction over the HEVC-based (BPG) codec, 5% bitrate, and 24%\nencoding time saving compare to the state-of-the-art semantic-based image\ncodec.",
+                    "authors": "Trinh Man Hoang, Jinjia Zhou, Yibo Fan",
+                    "published": "2021-01-24",
+                    "updated": "2021-01-30",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV",
+                        "cs.MM"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1609.09220v1",
+                    "title": "CNN-aware Binary Map for General Semantic Segmentation",
+                    "abstract": "In this paper we introduce a novel method for general semantic segmentation\nthat can benefit from general semantics of Convolutional Neural Network (CNN).\nOur segmentation proposes visually and semantically coherent image segments. We\nuse binary encoding of CNN features to overcome the difficulty of the\nclustering on the high-dimensional CNN feature space. These binary codes are\nvery robust against noise and non-semantic changes in the image. These binary\nencoding can be embedded into the CNN as an extra layer at the end of the\nnetwork. This results in real-time segmentation. To the best of our knowledge\nour method is the first attempt on general semantic image segmentation using\nCNN. All the previous papers were limited to few number of category of the\nimages (e.g. PASCAL VOC). Experiments show that our segmentation algorithm\noutperform the state-of-the-art non-semantic segmentation methods by large\nmargin.",
+                    "authors": "Mahdyar Ravanbakhsh, Hossein Mousavi, Moin Nabi, Mohammad Rastegari, Carlo Regazzoni",
+                    "published": "2016-09-29",
+                    "updated": "2016-09-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2212.07623v1",
+                    "title": "SBSS: Stacking-Based Semantic Segmentation Framework for Very High Resolution Remote Sensing Image",
+                    "abstract": "Semantic segmentation of Very High Resolution (VHR) remote sensing images is\na fundamental task for many applications. However, large variations in the\nscales of objects in those VHR images pose a challenge for performing accurate\nsemantic segmentation. Existing semantic segmentation networks are able to\nanalyse an input image at up to four resizing scales, but this may be\ninsufficient given the diversity of object scales. Therefore, Multi Scale (MS)\ntest-time data augmentation is often used in practice to obtain more accurate\nsegmentation results, which makes equal use of the segmentation results\nobtained at the different resizing scales. However, it was found in this study\nthat different classes of objects had their preferred resizing scale for more\naccurate semantic segmentation. Based on this behaviour, a Stacking-Based\nSemantic Segmentation (SBSS) framework is proposed to improve the segmentation\nresults by learning this behaviour, which contains a learnable Error Correction\nModule (ECM) for segmentation result fusion and an Error Correction Scheme\n(ECS) for computational complexity control. Two ECS, i.e., ECS-MS and ECS-SS,\nare proposed and investigated in this study. The Floating-point operations\n(Flops) required for ECS-MS and ECS-SS are similar to the commonly used MS test\nand the Single-Scale (SS) test, respectively. Extensive experiments on four\ndatasets (i.e., Cityscapes, UAVid, LoveDA and Potsdam) show that SBSS is an\neffective and flexible framework. It achieved higher accuracy than MS when\nusing ECS-MS, and similar accuracy as SS with a quarter of the memory footprint\nwhen using ECS-SS.",
+                    "authors": "Yuanzhi Cai, Lei Fan, Yuan Fang",
+                    "published": "2022-12-15",
+                    "updated": "2022-12-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.08988v1",
+                    "title": "Heterogeneous Feature Distillation Network for SAR Image Semantic Segmentation",
+                    "abstract": "Semantic segmentation for SAR (Synthetic Aperture Radar) images has attracted\nincreasing attention in the remote sensing community recently, due to SAR's\nall-time and all-weather imaging capability. However, SAR images are generally\nmore difficult to be segmented than their EO (Electro-Optical) counterparts,\nsince speckle noises and layovers are inevitably involved in SAR images. To\naddress this problem, we investigate how to introduce EO features to assist the\ntraining of a SAR-segmentation model, and propose a heterogeneous feature\ndistillation network for segmenting SAR images, called HFD-Net, where a\nSAR-segmentation student model gains knowledge from a pre-trained\nEO-segmentation teacher model. In the proposed HFD-Net, both the student and\nteacher models employ an identical architecture but different parameter\nconfigurations, and a heterogeneous feature distillation model is explored for\ntransferring latent EO features from the teacher model to the student model and\nthen enhancing the ability of the student model for SAR image segmentation. In\naddition, a heterogeneous feature alignment module is explored to aggregate\nmulti-scale features for segmentation in each of the student model and teacher\nmodel. Extensive experimental results on two public datasets demonstrate that\nthe proposed HFD-Net outperforms seven state-of-the-art SAR image semantic\nsegmentation methods.",
+                    "authors": "Gao Mengyu, Dong Qiulei",
+                    "published": "2022-10-17",
+                    "updated": "2022-10-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2009.12232v4",
+                    "title": "From Pixel to Patch: Synthesize Context-aware Features for Zero-shot Semantic Segmentation",
+                    "abstract": "Zero-shot learning has been actively studied for image classification task to\nrelieve the burden of annotating image labels. Interestingly, semantic\nsegmentation task requires more labor-intensive pixel-wise annotation, but\nzero-shot semantic segmentation has only attracted limited research interest.\nThus, we focus on zero-shot semantic segmentation, which aims to segment unseen\nobjects with only category-level semantic representations provided for unseen\ncategories. In this paper, we propose a novel Context-aware feature Generation\nNetwork (CaGNet), which can synthesize context-aware pixel-wise visual features\nfor unseen categories based on category-level semantic representations and\npixel-wise contextual information. The synthesized features are used to\nfinetune the classifier to enable segmenting unseen objects. Furthermore, we\nextend pixel-wise feature generation and finetuning to patch-wise feature\ngeneration and finetuning, which additionally considers inter-pixel\nrelationship. Experimental results on Pascal-VOC, Pascal-Context, and\nCOCO-stuff show that our method significantly outperforms the existing\nzero-shot semantic segmentation methods. Code is available at\nhttps://github.com/bcmi/CaGNetv2-Zero-Shot-Semantic-Segmentation.",
+                    "authors": "Zhangxuan Gu, Siyuan Zhou, Li Niu, Zihan Zhao, Liqing Zhang",
+                    "published": "2020-09-25",
+                    "updated": "2022-01-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1811.00174v4",
+                    "title": "Pixel Level Data Augmentation for Semantic Image Segmentation using Generative Adversarial Networks",
+                    "abstract": "Semantic segmentation is one of the basic topics in computer vision, it aims\nto assign semantic labels to every pixel of an image. Unbalanced semantic label\ndistribution could have a negative influence on segmentation accuracy. In this\npaper, we investigate using data augmentation approach to balance the semantic\nlabel distribution in order to improve segmentation performance. We propose\nusing generative adversarial networks (GANs) to generate realistic images for\nimproving the performance of semantic segmentation networks. Experimental\nresults show that the proposed method can not only improve segmentation\nperformance on those classes with low accuracy, but also obtain 1.3% to 2.1%\nincrease in average segmentation accuracy. It shows that this augmentation\nmethod can boost accuracy and be easily applicable to any other segmentation\nmodels.",
+                    "authors": "Shuangting Liu, Jiaqi Zhang, Yuxin Chen, Yifan Liu, Zengchang Qin, Tao Wan",
+                    "published": "2018-11-01",
+                    "updated": "2019-11-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1905.08748v3",
+                    "title": "RIU-Net: Embarrassingly simple semantic segmentation of 3D LiDAR point cloud",
+                    "abstract": "This paper proposes RIU-Net (for Range-Image U-Net), the adaptation of a\npopular semantic segmentation network for the semantic segmentation of a 3D\nLiDAR point cloud. The point cloud is turned into a 2D range-image by\nexploiting the topology of the sensor. This image is then used as input to a\nU-net. This architecture has already proved its efficiency for the task of\nsemantic segmentation of medical images. We demonstrate how it can also be used\nfor the accurate semantic segmentation of a 3D LiDAR point cloud and how it\nrepresents a valid bridge between image processing and 3D point cloud\nprocessing. Our model is trained on range-images built from KITTI 3D object\ndetection dataset. Experiments show that RIU-Net, despite being very simple,\noffers results that are comparable to the state-of-the-art of range-image based\nmethods. Finally, we demonstrate that this architecture is able to operate at\n90fps on a single GPU, which enables deployment for real-time segmentation.",
+                    "authors": "Pierre Biasutti, Aur\u00e9lie Bugeau, Jean-Fran\u00e7ois Aujol, Mathieu Br\u00e9dif",
+                    "published": "2019-05-21",
+                    "updated": "2019-06-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.10834v1",
+                    "title": "DANNet: A One-Stage Domain Adaptation Network for Unsupervised Nighttime Semantic Segmentation",
+                    "abstract": "Semantic segmentation of nighttime images plays an equally important role as\nthat of daytime images in autonomous driving, but the former is much more\nchallenging due to poor illuminations and arduous human annotations. In this\npaper, we propose a novel domain adaptation network (DANNet) for nighttime\nsemantic segmentation without using labeled nighttime image data. It employs an\nadversarial training with a labeled daytime dataset and an unlabeled dataset\nthat contains coarsely aligned day-night image pairs. Specifically, for the\nunlabeled day-night image pairs, we use the pixel-level predictions of static\nobject categories on a daytime image as a pseudo supervision to segment its\ncounterpart nighttime image. We further design a re-weighting strategy to\nhandle the inaccuracy caused by misalignment between day-night image pairs and\nwrong predictions of daytime images, as well as boost the prediction accuracy\nof small objects. The proposed DANNet is the first one stage adaptation\nframework for nighttime semantic segmentation, which does not train additional\nday-night image transfer models as a separate pre-processing stage. Extensive\nexperiments on Dark Zurich and Nighttime Driving datasets show that our method\nachieves state-of-the-art performance for nighttime semantic segmentation.",
+                    "authors": "Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, Song Wang",
+                    "published": "2021-04-22",
+                    "updated": "2021-04-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.12903v1",
+                    "title": "Land Cover Change Detection via Semantic Segmentation",
+                    "abstract": "This paper presents a change detection method that identifies land cover\nchanges from aerial imagery, using semantic segmentation, a machine learning\napproach. We present a land cover classification training pipeline with Deeplab\nv3+, state-of-the-art semantic segmentation technology, including data\npreparation, model training for seven land cover types, and model exporting\nmodules. In the land cover change detection system, the inputs are images\nretrieved from Google Earth at the same location but from different times. The\nsystem then predicts semantic segmentation results on these images using the\ntrained model and calculates the land cover class percentage for each input\nimage. We see an improvement in the accuracy of the land cover semantic\nsegmentation model, with a mean IoU of 0.756 compared to 0.433, as reported in\nthe DeepGlobe land cover classification challenge. The land cover change\ndetection system that leverages the state-of-the-art semantic segmentation\ntechnology is proposed and can be used for deforestation analysis, land\nmanagement, and urban planning.",
+                    "authors": "Renee Su, Rong Chen",
+                    "published": "2019-11-28",
+                    "updated": "2019-11-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2111.11103v1",
+                    "title": "Improving Semantic Image Segmentation via Label Fusion in Semantically Textured Meshes",
+                    "abstract": "Models for semantic segmentation require a large amount of hand-labeled\ntraining data which is costly and time-consuming to produce. For this purpose,\nwe present a label fusion framework that is capable of improving semantic pixel\nlabels of video sequences in an unsupervised manner. We make use of a 3D mesh\nrepresentation of the environment and fuse the predictions of different frames\ninto a consistent representation using semantic mesh textures. Rendering the\nsemantic mesh using the original intrinsic and extrinsic camera parameters\nyields a set of improved semantic segmentation images. Due to our optimized\nCUDA implementation, we are able to exploit the entire $c$-dimensional\nprobability distribution of annotations over $c$ classes in an\nuncertainty-aware manner. We evaluate our method on the Scannet dataset where\nwe improve annotations produced by the state-of-the-art segmentation network\nESANet from $52.05 \\%$ to $58.25 \\%$ pixel accuracy. We publish the source code\nof our framework online to foster future research in this area\n(\\url{https://github.com/fferflo/semantic-meshes}). To the best of our\nknowledge, this is the first publicly available label fusion framework for\nsemantic image segmentation based on meshes with semantic textures.",
+                    "authors": "Florian Fervers, Timo Breuer, Gregor Stachowiak, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens",
+                    "published": "2021-11-22",
+                    "updated": "2021-11-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.02840v1",
+                    "title": "Attention-based fusion of semantic boundary and non-boundary information to improve semantic segmentation",
+                    "abstract": "This paper introduces a method for image semantic segmentation grounded on a\nnovel fusion scheme, which takes place inside a deep convolutional neural\nnetwork. The main goal of our proposal is to explore object boundary\ninformation to improve the overall segmentation performance. Unlike previous\nworks that combine boundary and segmentation features, or those that use\nboundary information to regularize semantic segmentation, we instead propose a\nnovel approach that embodies boundary information onto segmentation. For that,\nour semantic segmentation method uses two streams, which are combined through\nan attention gate, forming an end-to-end Y-model. To the best of our knowledge,\nours is the first work to show that boundary detection can improve semantic\nsegmentation when fused through a semantic fusion gate (attention model). We\nperformed an extensive evaluation of our method over public data sets. We found\ncompetitive results on all data sets after comparing our proposed model with\nother twelve state-of-the-art segmenters, considering the same training\nconditions. Our proposed model achieved the best mIoU on the CityScapes,\nCamVid, and Pascal Context data sets, and the second best on Mapillary Vistas.",
+                    "authors": "Jefferson Fontinele, Gabriel Lefundes, Luciano Oliveira",
+                    "published": "2021-08-05",
+                    "updated": "2021-08-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2110.04487v1",
+                    "title": "Colour augmentation for improved semi-supervised semantic segmentation",
+                    "abstract": "Consistency regularization describes a class of approaches that have yielded\nstate-of-the-art results for semi-supervised classification. While\nsemi-supervised semantic segmentation proved to be more challenging, a number\nof successful approaches have been recently proposed. Recent work explored the\nchallenges involved in using consistency regularization for segmentation\nproblems. In their self-supervised work Chen et al. found that colour\naugmentation prevents a classification network from using image colour\nstatistics as a short-cut for self-supervised learning via instance\ndiscrimination. Drawing inspiration from this we find that a similar problem\nimpedes semi-supervised semantic segmentation and offer colour augmentation as\na solution, improving semi-supervised semantic segmentation performance on\nchallenging photographic imagery.",
+                    "authors": "Geoff French, Michal Mackiewicz",
+                    "published": "2021-10-09",
+                    "updated": "2021-10-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1705.09052v3",
+                    "title": "Weakly Supervised Semantic Segmentation Based on Web Image Co-segmentation",
+                    "abstract": "Training a Fully Convolutional Network (FCN) for semantic segmentation\nrequires a large number of masks with pixel level labelling, which involves a\nlarge amount of human labour and time for annotation. In contrast, web images\nand their image-level labels are much easier and cheaper to obtain. In this\nwork, we propose a novel method for weakly supervised semantic segmentation\nwith only image-level labels. The method utilizes the internet to retrieve a\nlarge number of images and uses a large scale co-segmentation framework to\ngenerate masks for the retrieved images. We first retrieve images from search\nengines, e.g. Flickr and Google, using semantic class names as queries, e.g.\nclass names in the dataset PASCAL VOC 2012. We then use high quality masks\nproduced by co-segmentation on the retrieved images as well as the target\ndataset images with image level labels to train segmentation networks. We\nobtain an IoU score of 56.9 on test set of PASCAL VOC 2012, which reaches the\nstate-of-the-art performance.",
+                    "authors": "Tong Shen, Guosheng Lin, Lingqiao Liu, Chunhua Shen, Ian Reid",
+                    "published": "2017-05-25",
+                    "updated": "2017-08-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.17874v1",
+                    "title": "SmooSeg: Smoothness Prior for Unsupervised Semantic Segmentation",
+                    "abstract": "Unsupervised semantic segmentation is a challenging task that segments images\ninto semantic groups without manual annotation. Prior works have primarily\nfocused on leveraging prior knowledge of semantic consistency or priori\nconcepts from self-supervised learning methods, which often overlook the\ncoherence property of image segments. In this paper, we demonstrate that the\nsmoothness prior, asserting that close features in a metric space share the\nsame semantics, can significantly simplify segmentation by casting unsupervised\nsemantic segmentation as an energy minimization problem. Under this paradigm,\nwe propose a novel approach called SmooSeg that harnesses self-supervised\nlearning methods to model the closeness relationships among observations as\nsmoothness signals. To effectively discover coherent semantic segments, we\nintroduce a novel smoothness loss that promotes piecewise smoothness within\nsegments while preserving discontinuities across different segments.\nAdditionally, to further enhance segmentation quality, we design an asymmetric\nteacher-student style predictor that generates smoothly updated pseudo labels,\nfacilitating an optimal fit between observations and labeling outputs. Thanks\nto the rich supervision cues of the smoothness prior, our SmooSeg significantly\noutperforms STEGO in terms of pixel accuracy on three datasets: COCOStuff\n(+14.9%), Cityscapes (+13.0%), and Potsdam-3 (+5.7%).",
+                    "authors": "Mengcheng Lan, Xinjiang Wang, Yiping Ke, Jiaxing Xu, Litong Feng, Wayne Zhang",
+                    "published": "2023-10-27",
+                    "updated": "2023-10-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.03138v1",
+                    "title": "Quantity beats quality for semantic segmentation of corrosion in images",
+                    "abstract": "Dataset creation is typically one of the first steps when applying Artificial\nIntelligence methods to a new task; and the real world performance of models\nhinges on the quality and quantity of data available. Producing an image\ndataset for semantic segmentation is resource intensive, particularly for\nspecialist subjects where class segmentation is not able to be effectively\nfarmed out. The benefit of producing a large, but poorly labelled, dataset\nversus a small, expertly segmented dataset for semantic segmentation is an open\nquestion. Here we show that a large, noisy dataset outperforms a small,\nexpertly segmented dataset for training a Fully Convolutional Network model for\nsemantic segmentation of corrosion in images. A large dataset of 250 images\nwith segmentations labelled by undergraduates and a second dataset of just 10\nimages, with segmentations labelled by subject matter experts were produced.\nThe mean Intersection over Union and micro F-score metrics were compared after\ntraining for 50,000 epochs. This work is illustrative for researchers setting\nout to develop deep learning models for detection and location of specialist\nfeatures.",
+                    "authors": "Will Nash, Tom Drummond, Nick Birbilis",
+                    "published": "2018-06-30",
+                    "updated": "2018-06-30",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1606.01481v1",
+                    "title": "Better Image Segmentation by Exploiting Dense Semantic Predictions",
+                    "abstract": "It is well accepted that image segmentation can benefit from utilizing\nmultilevel cues. The paper focuses on utilizing the FCNN-based dense semantic\npredictions in the bottom-up image segmentation, arguing to take semantic cues\ninto account from the very beginning. By this we can avoid merging regions of\nsimilar appearance but distinct semantic categories as possible. The semantic\ninefficiency problem is handled. We also propose a straightforward way to use\nthe contour cues to suppress the noise in multilevel cues, thus to improve the\nsegmentation robustness. The evaluation on the BSDS500 shows that we obtain the\ncompetitive region and boundary performance. Furthermore, since all individual\nregions can be assigned with appropriate semantic labels during the\ncomputation, we are capable of extracting the adjusted semantic segmentations.\nThe experiment on Pascal VOC 2012 shows our improvement to the original\nsemantic segmentations which derives directly from the dense predictions.",
+                    "authors": "Qiyang Zhao, Lewis D Griffin",
+                    "published": "2016-06-05",
+                    "updated": "2016-06-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.05490v1",
+                    "title": "Learning Semantic Segmentation with Query Points Supervision on Aerial Images",
+                    "abstract": "Semantic segmentation is crucial in remote sensing, where high-resolution\nsatellite images are segmented into meaningful regions. Recent advancements in\ndeep learning have significantly improved satellite image segmentation.\nHowever, most of these methods are typically trained in fully supervised\nsettings that require high-quality pixel-level annotations, which are expensive\nand time-consuming to obtain. In this work, we present a weakly supervised\nlearning algorithm to train semantic segmentation algorithms that only rely on\nquery point annotations instead of full mask labels. Our proposed approach\nperforms accurate semantic segmentation and improves efficiency by\nsignificantly reducing the cost and time required for manual annotation.\nSpecifically, we generate superpixels and extend the query point labels into\nthose superpixels that group similar meaningful semantics. Then, we train\nsemantic segmentation models, supervised with images partially labeled with the\nsuperpixels pseudo-labels. We benchmark our weakly supervised training approach\non an aerial image dataset and different semantic segmentation architectures,\nshowing that we can reach competitive performance compared to fully supervised\ntraining while reducing the annotation effort.",
+                    "authors": "Santiago Rivier, Carlos Hinojosa, Silvio Giancola, Bernard Ghanem",
+                    "published": "2023-09-11",
+                    "updated": "2023-09-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1804.04882v2",
+                    "title": "Learning to Exploit the Prior Network Knowledge for Weakly-Supervised Semantic Segmentation",
+                    "abstract": "Training a Convolutional Neural Network (CNN) for semantic segmentation\ntypically requires to collect a large amount of accurate pixel-level\nannotations, a hard and expensive task. In contrast, simple image tags are\neasier to gather. With this paper we introduce a novel weakly-supervised\nsemantic segmentation model able to learn from image labels, and just image\nlabels. Our model uses the prior knowledge of a network trained for image\nrecognition, employing these image annotations as an attention mechanism to\nidentify semantic regions in the images. We then present a methodology that\nbuilds accurate class-specific segmentation masks from these regions, where\nneither external objectness nor saliency algorithms are required. We describe\nhow to incorporate this mask generation strategy into a fully end-to-end\ntrainable process where the network jointly learns to classify and segment\nimages. Our experiments on PASCAL VOC 2012 dataset show that exploiting these\ngenerated class-specific masks in conjunction with our novel end-to-end\nlearning process outperforms several recent weakly-supervised semantic\nsegmentation methods that use image tags only, and even some models that\nleverage additional supervision or training data.",
+                    "authors": "Carolina Redondo-Cabrera, Marcos Baptista-R\u00edos, Roberto J. L\u00f3pez-Sastre",
+                    "published": "2018-04-13",
+                    "updated": "2019-02-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.13118v1",
+                    "title": "Automatic Preprocessing and Ensemble Learning for Low Quality Cell Image Segmentation",
+                    "abstract": "We propose an automatic preprocessing and ensemble learning for segmentation\nof cell images with low quality. It is difficult to capture cells with strong\nlight. Therefore, the microscopic images of cells tend to have low image\nquality but these images are not good for semantic segmentation. Here we\npropose a method to translate an input image to the images that are easy to\nrecognize by deep learning. The proposed method consists of two deep neural\nnetworks. The first network is the usual training for semantic segmentation,\nand penultimate feature maps of the first network are used as filters to\ntranslate an input image to the images that emphasize each class. This is the\nautomatic preprocessing and translated cell images are easily classified. The\ninput cell image with low quality is translated by the feature maps in the\nfirst network, and the translated images are fed into the second network for\nsemantic segmentation. Since the outputs of the second network are multiple\nsegmentation results, we conduct the weighted ensemble of those segmentation\nimages. Two networks are trained by end-to-end manner, and we do not need to\nprepare images with high quality for the translation. We confirmed that our\nproposed method can translate cell images with low quality to the images that\nare easy to segment, and segmentation accuracy has improved using the weighted\nensemble learning.",
+                    "authors": "Sota Kato, Kazuhiro Hotta",
+                    "published": "2021-08-30",
+                    "updated": "2021-08-30",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.13697v1",
+                    "title": "Generalizable Semantic Vision Query Generation for Zero-shot Panoptic and Semantic Segmentation",
+                    "abstract": "Zero-shot Panoptic Segmentation (ZPS) aims to recognize foreground instances\nand background stuff without images containing unseen categories in training.\nDue to the visual data sparsity and the difficulty of generalizing from seen to\nunseen categories, this task remains challenging. To better generalize to\nunseen classes, we propose Conditional tOken aligNment and Cycle trAnsiTion\n(CONCAT), to produce generalizable semantic vision queries. First, a feature\nextractor is trained by CON to link the vision and semantics for providing\ntarget queries. Formally, CON is proposed to align the semantic queries with\nthe CLIP visual CLS token extracted from complete and masked images. To address\nthe lack of unseen categories, a generator is required. However, one of the\ngaps in synthesizing pseudo vision queries, ie, vision queries for unseen\ncategories, is describing fine-grained visual details through semantic\nembeddings. Therefore, we approach CAT to train the generator in\nsemantic-vision and vision-semantic manners. In semantic-vision, visual query\ncontrast is proposed to model the high granularity of vision by pulling the\npseudo vision queries with the corresponding targets containing segments while\npushing those without segments away. To ensure the generated queries retain\nsemantic information, in vision-semantic, the pseudo vision queries are mapped\nback to semantic and supervised by real semantic embeddings. Experiments on ZPS\nachieve a 5.2% hPQ increase surpassing SOTA. We also examine inductive ZPS and\nopen-vocabulary semantic segmentation and obtain comparative results while\nbeing 2 times faster in testing.",
+                    "authors": "Jialei Chen, Daisuke Deguchi, Chenkai Zhang, Hiroshi Murase",
+                    "published": "2024-02-21",
+                    "updated": "2024-02-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1509.02441v1",
+                    "title": "Semantic Video Segmentation : Exploring Inference Efficiency",
+                    "abstract": "We explore the efficiency of the CRF inference beyond image level semantic\nsegmentation and perform joint inference in video frames. The key idea is to\ncombine best of two worlds: semantic co-labeling and more expressive models.\nOur formulation enables us to perform inference over ten thousand images within\nseconds and makes the system amenable to perform video semantic segmentation\nmost effectively. On CamVid dataset, with TextonBoost unaries, our proposed\nmethod achieves up to 8% improvement in accuracy over individual semantic image\nsegmentation without additional time overhead. The source code is available at\nhttps://github.com/subtri/video_inference",
+                    "authors": "Subarna Tripathi, Serge Belongie, Youngbae Hwang, Truong Nguyen",
+                    "published": "2015-09-04",
+                    "updated": "2015-09-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.03273v1",
+                    "title": "Semantic Segmentation using Vision Transformers: A survey",
+                    "abstract": "Semantic segmentation has a broad range of applications in a variety of\ndomains including land coverage analysis, autonomous driving, and medical image\nanalysis. Convolutional neural networks (CNN) and Vision Transformers (ViTs)\nprovide the architecture models for semantic segmentation. Even though ViTs\nhave proven success in image classification, they cannot be directly applied to\ndense prediction tasks such as image segmentation and object detection since\nViT is not a general purpose backbone due to its patch partitioning scheme. In\nthis survey, we discuss some of the different ViT architectures that can be\nused for semantic segmentation and how their evolution managed the above-stated\nchallenge. The rise of ViT and its performance with a high success rate\nmotivated the community to slowly replace the traditional convolutional neural\nnetworks in various computer vision tasks. This survey aims to review and\ncompare the performances of ViT architectures designed for semantic\nsegmentation using benchmarking datasets. This will be worthwhile for the\ncommunity to yield knowledge regarding the implementations carried out in\nsemantic segmentation and to discover more efficient methodologies using ViTs.",
+                    "authors": "Hans Thisanke, Chamli Deshan, Kavindu Chamith, Sachith Seneviratne, Rajith Vidanaarachchi, Damayanthi Herath",
+                    "published": "2023-05-05",
+                    "updated": "2023-05-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.02270v1",
+                    "title": "Weak-shot Semantic Segmentation via Dual Similarity Transfer",
+                    "abstract": "Semantic segmentation is an important and prevalent task, but severely\nsuffers from the high cost of pixel-level annotations when extending to more\nclasses in wider applications. To this end, we focus on the problem named\nweak-shot semantic segmentation, where the novel classes are learnt from\ncheaper image-level labels with the support of base classes having\noff-the-shelf pixel-level labels. To tackle this problem, we propose SimFormer,\nwhich performs dual similarity transfer upon MaskFormer. Specifically,\nMaskFormer disentangles the semantic segmentation task into two sub-tasks:\nproposal classification and proposal segmentation for each proposal. Proposal\nsegmentation allows proposal-pixel similarity transfer from base classes to\nnovel classes, which enables the mask learning of novel classes. We also learn\npixel-pixel similarity from base classes and distill such class-agnostic\nsemantic similarity to the semantic masks of novel classes, which regularizes\nthe segmentation model with pixel-level semantic relationship across images. In\naddition, we propose a complementary loss to facilitate the learning of novel\nclasses. Comprehensive experiments on the challenging COCO-Stuff-10K and ADE20K\ndatasets demonstrate the effectiveness of our method. Codes are available at\nhttps://github.com/bcmi/SimFormer-Weak-Shot-Semantic-Segmentation.",
+                    "authors": "Junjie Chen, Li Niu, Siyuan Zhou, Jianlou Si, Chen Qian, Liqing Zhang",
+                    "published": "2022-10-05",
+                    "updated": "2022-10-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1809.10198v1",
+                    "title": "Recent progress in semantic image segmentation",
+                    "abstract": "Semantic image segmentation, which becomes one of the key applications in\nimage processing and computer vision domain, has been used in multiple domains\nsuch as medical area and intelligent transportation. Lots of benchmark datasets\nare released for researchers to verify their algorithms. Semantic segmentation\nhas been studied for many years. Since the emergence of Deep Neural Network\n(DNN), segmentation has made a tremendous progress. In this paper, we divide\nsemantic image segmentation methods into two categories: traditional and recent\nDNN method. Firstly, we briefly summarize the traditional method as well as\ndatasets released for segmentation, then we comprehensively investigate recent\nmethods based on DNN which are described in the eight aspects: fully\nconvolutional network, upsample ways, FCN joint with CRF methods, dilated\nconvolution approaches, progresses in backbone network, pyramid methods,\nMulti-level feature and multi-stage method, supervised, weakly-supervised and\nunsupervised methods. Finally, a conclusion in this area is drawn.",
+                    "authors": "Xiaolong Liu, Zhidong Deng, Yuhan Yang",
+                    "published": "2018-09-20",
+                    "updated": "2018-09-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1611.08408v1",
+                    "title": "Semantic Segmentation using Adversarial Networks",
+                    "abstract": "Adversarial training has been shown to produce state of the art results for\ngenerative image modeling. In this paper we propose an adversarial training\napproach to train semantic segmentation models. We train a convolutional\nsemantic segmentation network along with an adversarial network that\ndiscriminates segmentation maps coming either from the ground truth or from the\nsegmentation network. The motivation for our approach is that it can detect and\ncorrect higher-order inconsistencies between ground truth segmentation maps and\nthe ones produced by the segmentation net. Our experiments show that our\nadversarial training approach leads to improved accuracy on the Stanford\nBackground and PASCAL VOC 2012 datasets.",
+                    "authors": "Pauline Luc, Camille Couprie, Soumith Chintala, Jakob Verbeek",
+                    "published": "2016-11-25",
+                    "updated": "2016-11-25",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2405.04913v1",
+                    "title": "Weakly-supervised Semantic Segmentation via Dual-stream Contrastive Learning of Cross-image Contextual Information",
+                    "abstract": "Weakly supervised semantic segmentation (WSSS) aims at learning a semantic\nsegmentation model with only image-level tags. Despite intensive research on\ndeep learning approaches over a decade, there is still a significant\nperformance gap between WSSS and full semantic segmentation. Most current WSSS\nmethods always focus on a limited single image (pixel-wise) information while\nignoring the valuable inter-image (semantic-wise) information. From this\nperspective, a novel end-to-end WSSS framework called DSCNet is developed along\nwith two innovations: i) pixel-wise group contrast and semantic-wise graph\ncontrast are proposed and introduced into the WSSS framework; ii) a novel\ndual-stream contrastive learning (DSCL) mechanism is designed to jointly handle\npixel-wise and semantic-wise context information for better WSSS performance.\nSpecifically, the pixel-wise group contrast learning (PGCL) and semantic-wise\ngraph contrast learning (SGCL) tasks form a more comprehensive solution.\nExtensive experiments on PASCAL VOC and MS COCO benchmarks verify the\nsuperiority of DSCNet over SOTA approaches and baseline models.",
+                    "authors": "Qi Lai, Chi-Man Vong",
+                    "published": "2024-05-08",
+                    "updated": "2024-05-08",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2102.12095v1",
+                    "title": "Synergy Between Semantic Segmentation and Image Denoising via Alternate Boosting",
+                    "abstract": "The capability of image semantic segmentation may be deteriorated due to\nnoisy input image, where image denoising prior to segmentation helps. Both\nimage denoising and semantic segmentation have been developed significantly\nwith the advance of deep learning. Thus, we are interested in the synergy\nbetween them by using a holistic deep model. We observe that not only denoising\nhelps combat the drop of segmentation accuracy due to noise, but also\npixel-wise semantic information boosts the capability of denoising. We then\npropose a boosting network to perform denoising and segmentation alternately.\nThe proposed network is composed of multiple segmentation and denoising blocks\n(SDBs), each of which estimates semantic map then uses the map to regularize\ndenoising. Experimental results show that the denoised image quality is\nimproved substantially and the segmentation accuracy is improved to close to\nthat of clean images. Our code and models will be made publicly available.",
+                    "authors": "Shunxin Xu, Ke Sun, Dong Liu, Zhiwei Xiong, Zheng-Jun Zha",
+                    "published": "2021-02-24",
+                    "updated": "2021-02-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2106.04108v3",
+                    "title": "Fully Transformer Networks for Semantic Image Segmentation",
+                    "abstract": "Transformers have shown impressive performance in various natural language\nprocessing and computer vision tasks, due to the capability of modeling\nlong-range dependencies. Recent progress has demonstrated that combining such\nTransformers with CNN-based semantic image segmentation models is very\npromising. However, it is not well studied yet on how well a pure Transformer\nbased approach can achieve for image segmentation. In this work, we explore a\nnovel framework for semantic image segmentation, which is encoder-decoder based\nFully Transformer Networks (FTN). Specifically, we first propose a Pyramid\nGroup Transformer (PGT) as the encoder for progressively learning hierarchical\nfeatures, meanwhile reducing the computation complexity of the standard Visual\nTransformer (ViT). Then, we propose a Feature Pyramid Transformer (FPT) to fuse\nsemantic-level and spatial-level information from multiple levels of the PGT\nencoder for semantic image segmentation. Surprisingly, this simple baseline can\nachieve better results on multiple challenging semantic segmentation and face\nparsing benchmarks, including PASCAL Context, ADE20K, COCOStuff, and\nCelebAMask-HQ. The source code will be released on\nhttps://github.com/BR-IDL/PaddleViT.",
+                    "authors": "Sitong Wu, Tianyi Wu, Fangjian Lin, Shengwei Tian, Guodong Guo",
+                    "published": "2021-06-08",
+                    "updated": "2021-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.17091v1",
+                    "title": "SSSegmenation: An Open Source Supervised Semantic Segmentation Toolbox Based on PyTorch",
+                    "abstract": "This paper presents SSSegmenation, which is an open source supervised\nsemantic image segmentation toolbox based on PyTorch. The design of this\ntoolbox is motivated by MMSegmentation while it is easier to use because of\nfewer dependencies and achieves superior segmentation performance under a\ncomparable training and testing setup. Moreover, the toolbox also provides\nplenty of trained weights for popular and contemporary semantic segmentation\nmethods, including Deeplab, PSPNet, OCRNet, MaskFormer, \\emph{etc}. We expect\nthat this toolbox can contribute to the future development of semantic\nsegmentation. Codes and model zoos are available at\n\\href{https://github.com/SegmentationBLWX/sssegmentation/}{SSSegmenation}.",
+                    "authors": "Zhenchao Jin",
+                    "published": "2023-05-26",
+                    "updated": "2023-05-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2403.01482v4",
+                    "title": "EAGLE: Eigen Aggregation Learning for Object-Centric Unsupervised Semantic Segmentation",
+                    "abstract": "Semantic segmentation has innately relied on extensive pixel-level annotated\ndata, leading to the emergence of unsupervised methodologies. Among them,\nleveraging self-supervised Vision Transformers for unsupervised semantic\nsegmentation (USS) has been making steady progress with expressive deep\nfeatures. Yet, for semantically segmenting images with complex objects, a\npredominant challenge remains: the lack of explicit object-level semantic\nencoding in patch-level features. This technical limitation often leads to\ninadequate segmentation of complex objects with diverse structures. To address\nthis gap, we present a novel approach, EAGLE, which emphasizes object-centric\nrepresentation learning for unsupervised semantic segmentation. Specifically,\nwe introduce EiCue, a spectral technique providing semantic and structural cues\nthrough an eigenbasis derived from the semantic similarity matrix of deep image\nfeatures and color affinity from an image. Further, by incorporating our\nobject-centric contrastive loss with EiCue, we guide our model to learn\nobject-level representations with intra- and inter-image object-feature\nconsistency, thereby enhancing semantic accuracy. Extensive experiments on\nCOCO-Stuff, Cityscapes, and Potsdam-3 datasets demonstrate the state-of-the-art\nUSS results of EAGLE with accurate and consistent semantic segmentation across\ncomplex scenes.",
+                    "authors": "Chanyoung Kim, Woojung Han, Dayun Ju, Seong Jae Hwang",
+                    "published": "2024-03-03",
+                    "updated": "2024-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1502.04983v1",
+                    "title": "Context Tricks for Cheap Semantic Segmentation",
+                    "abstract": "Accurate semantic labeling of image pixels is difficult because intra-class\nvariability is often greater than inter-class variability. In turn, fast\nsemantic segmentation is hard because accurate models are usually too\ncomplicated to also run quickly at test-time. Our experience with building and\nrunning semantic segmentation systems has also shown a reasonably obvious\nbottleneck on model complexity, imposed by small training datasets. We\ntherefore propose two simple complementary strategies that leverage context to\ngive better semantic segmentation, while scaling up or down to train on\ndifferent-sized datasets.\n  As easy modifications for existing semantic segmentation algorithms, we\nintroduce Decorrelated Semantic Texton Forests, and the Context Sensitive Image\nLevel Prior. The proposed modifications are tested using a Semantic Texton\nForest (STF) system, and the modifications are validated on two standard\nbenchmark datasets, MSRC-21 and PascalVOC-2010. In Python based comparisons,\nour system is insignificantly slower than STF at test-time, yet produces\nsuperior semantic segmentations overall, with just push-button training.",
+                    "authors": "Thanapong Intharah, Gabriel J. Brostow",
+                    "published": "2015-02-17",
+                    "updated": "2015-02-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2011.00674v1",
+                    "title": "Highway Driving Dataset for Semantic Video Segmentation",
+                    "abstract": "Scene understanding is an essential technique in semantic segmentation.\nAlthough there exist several datasets that can be used for semantic\nsegmentation, they are mainly focused on semantic image segmentation with large\ndeep neural networks. Therefore, these networks are not useful for real time\napplications, especially in autonomous driving systems. In order to solve this\nproblem, we make two contributions to semantic segmentation task. The first\ncontribution is that we introduce the semantic video dataset, the Highway\nDriving dataset, which is a densely annotated benchmark for a semantic video\nsegmentation task. The Highway Driving dataset consists of 20 video sequences\nhaving a 30Hz frame rate, and every frame is densely annotated. Secondly, we\npropose a baseline algorithm that utilizes a temporal correlation. Together\nwith our attempt to analyze the temporal correlation, we expect the Highway\nDriving dataset to encourage research on semantic video segmentation.",
+                    "authors": "Byungju Kim, Junho Yim, Junmo Kim",
+                    "published": "2020-11-02",
+                    "updated": "2020-11-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2304.11332v2",
+                    "title": "Input Augmentation with SAM: Boosting Medical Image Segmentation with Segmentation Foundation Model",
+                    "abstract": "The Segment Anything Model (SAM) is a recently developed large model for\ngeneral-purpose segmentation for computer vision tasks. SAM was trained using\n11 million images with over 1 billion masks and can produce segmentation\nresults for a wide range of objects in natural scene images. SAM can be viewed\nas a general perception model for segmentation (partitioning images into\nsemantically meaningful regions). Thus, how to utilize such a large foundation\nmodel for medical image segmentation is an emerging research target. This paper\nshows that although SAM does not immediately give high-quality segmentation for\nmedical image data, its generated masks, features, and stability scores are\nuseful for building and training better medical image segmentation models. In\nparticular, we demonstrate how to use SAM to augment image input for\ncommonly-used medical image segmentation models (e.g., U-Net). Experiments on\nthree segmentation tasks show the effectiveness of our proposed SAMAug method.\nThe code is available at \\url{https://github.com/yizhezhang2000/SAMAug}.",
+                    "authors": "Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Danny Z. Chen",
+                    "published": "2023-04-22",
+                    "updated": "2023-06-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2103.00286v1",
+                    "title": "A Novel Adaptive Deep Network for Building Footprint Segmentation",
+                    "abstract": "Building footprint segmentations for high resolution images are increasingly\ndemanded for many remote sensing applications. By the emerging deep learning\napproaches, segmentation networks have made significant advances in the\nsemantic segmentation of objects. However, these advances and the increased\naccess to satellite images require the generation of accurate object boundaries\nin satellite images. In the current paper, we propose a novel network-based on\nPix2Pix methodology to solve the problem of inaccurate boundaries obtained by\nconverting satellite images into maps using segmentation networks in order to\nsegment building footprints. To define the new network named G2G, our framework\nincludes two generators where the first generator extracts localization\nfeatures in order to merge them with the boundary features extracted from the\nsecond generator to segment all detailed building edges. Moreover, different\nstrategies are implemented to enhance the quality of the proposed networks'\nresults, implying that the proposed network outperforms state-of-the-art\nnetworks in segmentation accuracy with a large margin for all evaluation\nmetrics. The implementation is available at\nhttps://github.com/A2Amir/A-Novel-Adaptive-Deep-Network-for-Building-Footprint-Segmentation.",
+                    "authors": "A. Ziaee, R. Dehbozorgi, M. D\u00f6ller",
+                    "published": "2021-02-27",
+                    "updated": "2021-02-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG",
+                        "cs.NA",
+                        "eess.IV",
+                        "math.NA"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.06753v1",
+                    "title": "3rd Place Solution for PVUW Challenge 2023: Video Panoptic Segmentation",
+                    "abstract": "In order to deal with the task of video panoptic segmentation in the wild, we\npropose a robust integrated video panoptic segmentation solution. In our\nsolution, we regard the video panoptic segmentation task as a segmentation\ntarget querying task, represent both semantic and instance targets as a set of\nqueries, and then combine these queries with video features extracted by neural\nnetworks to predict segmentation masks. In order to improve the learning\naccuracy and convergence speed of the solution, we add additional tasks of\nvideo semantic segmentation and video instance segmentation for joint training.\nIn addition, we also add an additional image semantic segmentation model to\nfurther improve the performance of semantic classes. In addition, we also add\nsome additional operations to improve the robustness of the model. Extensive\nexperiments on the VIPSeg dataset show that the proposed solution achieves\nstate-of-the-art performance with 50.04\\% VPQ on the VIPSeg test set, which is\n3rd place on the video panoptic segmentation track of the PVUW Challenge 2023.",
+                    "authors": "Jinming Su, Wangwang Yang, Junfeng Luo, Xiaolin Wei",
+                    "published": "2023-06-11",
+                    "updated": "2023-06-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2001.00335v1",
+                    "title": "Graph-FCN for image semantic segmentation",
+                    "abstract": "Semantic segmentation with deep learning has achieved great progress in\nclassifying the pixels in the image. However, the local location information is\nusually ignored in the high-level feature extraction by the deep learning,\nwhich is important for image semantic segmentation. To avoid this problem, we\npropose a graph model initialized by a fully convolutional network (FCN) named\nGraph-FCN for image semantic segmentation. Firstly, the image grid data is\nextended to graph structure data by a convolutional network, which transforms\nthe semantic segmentation problem into a graph node classification problem.\nThen we apply graph convolutional network to solve this graph node\nclassification problem. As far as we know, it is the first time that we apply\nthe graph convolutional network in image semantic segmentation. Our method\nachieves competitive performance in mean intersection over union (mIOU) on the\nVOC dataset(about 1.34% improvement), compared to the original FCN model.",
+                    "authors": "Yi Lu, Yaran Chen, Dongbin Zhao, Jianxin Chen",
+                    "published": "2020-01-02",
+                    "updated": "2020-01-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2203.15943v1",
+                    "title": "Self-Supervised Leaf Segmentation under Complex Lighting Conditions",
+                    "abstract": "As an essential prerequisite task in image-based plant phenotyping, leaf\nsegmentation has garnered increasing attention in recent years. While\nself-supervised learning is emerging as an effective alternative to various\ncomputer vision tasks, its adaptation for image-based plant phenotyping remains\nrather unexplored. In this work, we present a self-supervised leaf segmentation\nframework consisting of a self-supervised semantic segmentation model, a\ncolor-based leaf segmentation algorithm, and a self-supervised color correction\nmodel. The self-supervised semantic segmentation model groups the semantically\nsimilar pixels by iteratively referring to the self-contained information,\nallowing the pixels of the same semantic object to be jointly considered by the\ncolor-based leaf segmentation algorithm for identifying the leaf regions.\nAdditionally, we propose to use a self-supervised color correction model for\nimages taken under complex illumination conditions. Experimental results on\ndatasets of different plant species demonstrate the potential of the proposed\nself-supervised framework in achieving effective and generalizable leaf\nsegmentation.",
+                    "authors": "Xufeng Lin, Chang-Tsun Li, Scott Adams, Abbas Kouzani, Richard Jiang, Ligang He, Yongjian Hu, Michael Vernon, Egan Doeven, Lawrence Webb, Todd Mcclellan, Adam Guskic",
+                    "published": "2022-03-29",
+                    "updated": "2022-03-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.05321v1",
+                    "title": "Image Segmentation Semantic Communication over Internet of Vehicles",
+                    "abstract": "In this paper, the problem of semantic-based efficient image transmission is\nstudied over the Internet of Vehicles (IoV). In the considered model, a vehicle\nshares massive amount of visual data perceived by its visual sensors to assist\nother vehicles in making driving decisions. However, it is hard to maintain a\nhigh reliable visual data transmission due to the limited spectrum resources.\nTo tackle this problem, a semantic communication approach is introduced to\nreduce the transmission data amount while ensuring the semantic-level accuracy.\nParticularly, an image segmentation semantic communication (ISSC) system is\nproposed, which can extract the semantic features from the perceived images and\ntransmit the features to the receiving vehicle that reconstructs the image\nsegmentations. The ISSC system consists of an encoder and a decoder at the\ntransmitter and the receiver, respectively. To accurately extract the image\nsemantic features, the ISSC system encoder employs a Swin Transformer based\nmulti-scale semantic feature extractor. Then, to resist the wireless noise and\nreconstruct the image segmentation, a semantic feature decoder and a\nreconstructor are designed at the receiver. Simulation results show that the\nproposed ISSC system can reconstruct the image segmentation accurately with a\nhigh compression ratio, and can achieve robust transmission performance against\nchannel noise, especially at the low signal-to-noise ratio (SNR). In terms of\nmean Intersection over Union (mIoU), the ISSC system can achieve an increase by\n75%, compared to the baselines using traditional coding method",
+                    "authors": "Qiang Pan, Haonan Tong, Jie Lv, Tao Luo, Zhilong Zhang, Changchuan Yin, Jianfeng Li",
+                    "published": "2022-10-11",
+                    "updated": "2022-10-11",
+                    "primary_cat": "cs.NI",
+                    "cats": [
+                        "cs.NI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1605.06885v1",
+                    "title": "Bridging Category-level and Instance-level Semantic Image Segmentation",
+                    "abstract": "We propose an approach to instance-level image segmentation that is built on\ntop of category-level segmentation. Specifically, for each pixel in a semantic\ncategory mask, its corresponding instance bounding box is predicted using a\ndeep fully convolutional regression network. Thus it follows a different\npipeline to the popular detect-then-segment approaches that first predict\ninstances' bounding boxes, which are the current state-of-the-art in instance\nsegmentation. We show that, by leveraging the strength of our state-of-the-art\nsemantic segmentation models, the proposed method can achieve comparable or\neven better results to detect-then-segment approaches. We make the following\ncontributions. (i) First, we propose a simple yet effective approach to\nsemantic instance segmentation. (ii) Second, we propose an online bootstrapping\nmethod during training, which is critically important for achieving good\nperformance for both semantic category segmentation and instance-level\nsegmentation. (iii) As the performance of semantic category segmentation has a\nsignificant impact on the instance-level segmentation, which is the second step\nof our approach, we train fully convolutional residual networks to achieve the\nbest semantic category segmentation accuracy. On the PASCAL VOC 2012 dataset,\nwe obtain the currently best mean intersection-over-union score of 79.1%. (iv)\nWe also achieve state-of-the-art results for instance-level segmentation.",
+                    "authors": "Zifeng Wu, Chunhua Shen, Anton van den Hengel",
+                    "published": "2016-05-23",
+                    "updated": "2016-05-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.03212v2",
+                    "title": "Hierarchical Semantic Segmentation using Psychometric Learning",
+                    "abstract": "Assigning meaning to parts of image data is the goal of semantic image\nsegmentation. Machine learning methods, specifically supervised learning is\ncommonly used in a variety of tasks formulated as semantic segmentation. One of\nthe major challenges in the supervised learning approaches is expressing and\ncollecting the rich knowledge that experts have with respect to the meaning\npresent in the image data. Towards this, typically a fixed set of labels is\nspecified and experts are tasked with annotating the pixels, patches or\nsegments in the images with the given labels. In general, however, the set of\nclasses does not fully capture the rich semantic information present in the\nimages. For example, in medical imaging such as histology images, the different\nparts of cells could be grouped and sub-grouped based on the expertise of the\npathologist.\n  To achieve such a precise semantic representation of the concepts in the\nimage, we need access to the full depth of knowledge of the annotator. In this\nwork, we develop a novel approach to collect segmentation annotations from\nexperts based on psychometric testing. Our method consists of the psychometric\ntesting procedure, active query selection, query enhancement, and a deep metric\nlearning model to achieve a patch-level image embedding that allows for\nsemantic segmentation of images. We show the merits of our method with\nevaluation on the synthetically generated image, aerial image and histology\nimage.",
+                    "authors": "Lu Yin, Vlado Menkovski, Shiwei Liu, Mykola Pechenizkiy",
+                    "published": "2021-07-07",
+                    "updated": "2021-12-16",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.12545v1",
+                    "title": "Unsupervised Domain Adaptation for Semantic Segmentation by Content Transfer",
+                    "abstract": "In this paper, we tackle the unsupervised domain adaptation (UDA) for\nsemantic segmentation, which aims to segment the unlabeled real data using\nlabeled synthetic data. The main problem of UDA for semantic segmentation\nrelies on reducing the domain gap between the real image and synthetic image.\nTo solve this problem, we focused on separating information in an image into\ncontent and style. Here, only the content has cues for semantic segmentation,\nand the style makes the domain gap. Thus, precise separation of content and\nstyle in an image leads to effect as supervision of real data even when\nlearning with synthetic data. To make the best of this effect, we propose a\nzero-style loss. Even though we perfectly extract content for semantic\nsegmentation in the real domain, another main challenge, the class imbalance\nproblem, still exists in UDA for semantic segmentation. We address this problem\nby transferring the contents of tail classes from synthetic to real domain.\nExperimental results show that the proposed method achieves the\nstate-of-the-art performance in semantic segmentation on the major two UDA\nsettings.",
+                    "authors": "Suhyeon Lee, Junhyuk Hyun, Hongje Seong, Euntai Kim",
+                    "published": "2020-12-23",
+                    "updated": "2020-12-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2202.04754v2",
+                    "title": "Wireless Transmission of Images With The Assistance of Multi-level Semantic Information",
+                    "abstract": "Semantic-oriented communication has been considered as a promising to boost\nthe bandwidth efficiency by only transmitting the semantics of the data. In\nthis paper, we propose a multi-level semantic aware communication system for\nwireless image transmission, named MLSC-image, which is based on the deep\nlearning techniques and trained in an end to end manner. In particular, the\nproposed model includes a multilevel semantic feature extractor, that extracts\nboth the highlevel semantic information, such as the text semantics and the\nsegmentation semantics, and the low-level semantic information, such as local\nspatial details of the images. We employ a pretrained image caption to capture\nthe text semantics and a pretrained image segmentation model to obtain the\nsegmentation semantics. These high-level and low-level semantic features are\nthen combined and encoded by a joint semantic and channel encoder into symbols\nto transmit over the physical channel. The numerical results validate the\neffectiveness and efficiency of the proposed semantic communication system,\nespecially under the limited bandwidth condition, which indicates the\nadvantages of the high-level semantics in the compression of images.",
+                    "authors": "Zhenguo Zhang, Qianqian Yang, Shibo He, Mingyang Sun, Jiming Chen",
+                    "published": "2022-02-08",
+                    "updated": "2023-12-08",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.13026v1",
+                    "title": "Weakly-Supervised Semantic Segmentation with Image-Level Labels: from Traditional Models to Foundation Models",
+                    "abstract": "The rapid development of deep learning has driven significant progress in the\nfield of image semantic segmentation - a fundamental task in computer vision.\nSemantic segmentation algorithms often depend on the availability of\npixel-level labels (i.e., masks of objects), which are expensive,\ntime-consuming, and labor-intensive. Weakly-supervised semantic segmentation\n(WSSS) is an effective solution to avoid such labeling. It utilizes only\npartial or incomplete annotations and provides a cost-effective alternative to\nfully-supervised semantic segmentation. In this paper, we focus on the WSSS\nwith image-level labels, which is the most challenging form of WSSS. Our work\nhas two parts. First, we conduct a comprehensive survey on traditional methods,\nprimarily focusing on those presented at premier research conferences. We\ncategorize them into four groups based on where their methods operate:\npixel-wise, image-wise, cross-image, and external data. Second, we investigate\nthe applicability of visual foundation models, such as the Segment Anything\nModel (SAM), in the context of WSSS. We scrutinize SAM in two intriguing\nscenarios: text prompting and zero-shot learning. We provide insights into the\npotential and challenges associated with deploying visual foundational models\nfor WSSS, facilitating future developments in this exciting research area.",
+                    "authors": "Zhaozheng Chen, Qianru Sun",
+                    "published": "2023-10-19",
+                    "updated": "2023-10-19",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.03983v1",
+                    "title": "Weakly Supervised Semantic Segmentation of Satellite Images",
+                    "abstract": "When one wants to train a neural network to perform semantic segmentation,\ncreating pixel-level annotations for each of the images in the database is a\ntedious task. If he works with aerial or satellite images, which are usually\nvery large, it is even worse. With that in mind, we investigate how to use\nimage-level annotations in order to perform semantic segmentation. Image-level\nannotations are much less expensive to acquire than pixel-level annotations,\nbut we lose a lot of information for the training of the model. From the\nannotations of the images, the model must find by itself how to classify the\ndifferent regions of the image. In this work, we use the method proposed by Anh\nand Kwak [1] to produce pixel-level annotation from image level annotation. We\ncompare the overall quality of our generated dataset with the original dataset.\nIn addition, we propose an adaptation of the AffinityNet that allows us to\ndirectly perform a semantic segmentation. Our results show that the generated\nlabels lead to the same performances for the training of several segmentation\nnetworks. Also, the quality of semantic segmentation performed directly by the\nAffinityNet and the Random Walk is close to the one of the best\nfully-supervised approaches.",
+                    "authors": "Adrien Nivaggioli, Hicham Randrianarivo",
+                    "published": "2019-04-08",
+                    "updated": "2019-04-08",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.05840v2",
+                    "title": "Self-Correlation and Cross-Correlation Learning for Few-Shot Remote Sensing Image Semantic Segmentation",
+                    "abstract": "Remote sensing image semantic segmentation is an important problem for remote\nsensing image interpretation. Although remarkable progress has been achieved,\nexisting deep neural network methods suffer from the reliance on massive\ntraining data. Few-shot remote sensing semantic segmentation aims at learning\nto segment target objects from a query image using only a few annotated support\nimages of the target class. Most existing few-shot learning methods stem\nprimarily from their sole focus on extracting information from support images,\nthereby failing to effectively address the large variance in appearance and\nscales of geographic objects. To tackle these challenges, we propose a\nSelf-Correlation and Cross-Correlation Learning Network for the few-shot remote\nsensing image semantic segmentation. Our model enhances the generalization by\nconsidering both self-correlation and cross-correlation between support and\nquery images to make segmentation predictions. To further explore the\nself-correlation with the query image, we propose to adopt a classical spectral\nmethod to produce a class-agnostic segmentation mask based on the basic visual\ninformation of the image. Extensive experiments on two remote sensing image\ndatasets demonstrate the effectiveness and superiority of our model in few-shot\nremote sensing image semantic segmentation. Code and models will be accessed at\nhttps://github.com/linhanwang/SCCNet.",
+                    "authors": "Linhan Wang, Shuo Lei, Jianfeng He, Shengkun Wang, Min Zhang, Chang-Tien Lu",
+                    "published": "2023-09-11",
+                    "updated": "2023-09-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1707.02432v2",
+                    "title": "Deep Semantic Segmentation for Automated Driving: Taxonomy, Roadmap and Challenges",
+                    "abstract": "Semantic segmentation was seen as a challenging computer vision problem few\nyears ago. Due to recent advancements in deep learning, relatively accurate\nsolutions are now possible for its use in automated driving. In this paper, the\nsemantic segmentation problem is explored from the perspective of automated\ndriving. Most of the current semantic segmentation algorithms are designed for\ngeneric images and do not incorporate prior structure and end goal for\nautomated driving. First, the paper begins with a generic taxonomic survey of\nsemantic segmentation algorithms and then discusses how it fits in the context\nof automated driving. Second, the particular challenges of deploying it into a\nsafety system which needs high level of accuracy and robustness are listed.\nThird, different alternatives instead of using an independent semantic\nsegmentation module are explored. Finally, an empirical evaluation of various\nsemantic segmentation architectures was performed on CamVid dataset in terms of\naccuracy and speed. This paper is a preliminary shorter version of a more\ndetailed survey which is work in progress.",
+                    "authors": "Mennatullah Siam, Sara Elkerdawy, Martin Jagersand, Senthil Yogamani",
+                    "published": "2017-07-08",
+                    "updated": "2017-08-03",
+                    "primary_cat": "stat.ML",
+                    "cats": [
+                        "stat.ML",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.00679v3",
+                    "title": "Cooperative Semantic Segmentation and Image Restoration in Adverse Environmental Conditions",
+                    "abstract": "Most state-of-the-art semantic segmentation approaches only achieve high\naccuracy in good conditions. In practically-common but less-discussed adverse\nenvironmental conditions, their performance can decrease enormously. Existing\nstudies usually cast the handling of segmentation in adverse conditions as a\nseparate post-processing step after signal restoration, making the segmentation\nperformance largely depend on the quality of restoration. In this paper, we\npropose a novel deep-learning framework to tackle semantic segmentation and\nimage restoration in adverse environmental conditions in a holistic manner. The\nproposed approach contains two components: Semantically-Guided Adaptation,\nwhich exploits semantic information from degraded images to refine the\nsegmentation; and Exemplar-Guided Synthesis, which restores images from\nsemantic label maps given degraded exemplars as the guidance. Our method\ncooperatively leverages the complementarity and interdependence of low-level\nrestoration and high-level segmentation in adverse environmental conditions.\nExtensive experiments on various datasets demonstrate that our approach can not\nonly improve the accuracy of semantic segmentation with degradation cues, but\nalso boost the perceptual quality and structural similarity of image\nrestoration with semantic guidance.",
+                    "authors": "Weihao Xia, Zhanglin Cheng, Yujiu Yang, Jing-Hao Xue",
+                    "published": "2019-11-02",
+                    "updated": "2020-03-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.12417v2",
+                    "title": "SLAMs: Semantic Learning based Activation Map for Weakly Supervised Semantic Segmentation",
+                    "abstract": "Recent mainstream weakly-supervised semantic segmentation (WSSS) approaches\nmainly relies on image-level classification learning, which has limited\nrepresentation capacity. In this paper, we propose a novel semantic learning\nbased framework, named SLAMs (Semantic Learning based Activation Map), for\nWSSS.",
+                    "authors": "Junliang Chen, Xiaodong Zhao, Minmin Liu, Linlin Shen",
+                    "published": "2022-10-22",
+                    "updated": "2022-11-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.10782v2",
+                    "title": "Three Ways to Improve Semantic Segmentation with Self-Supervised Depth Estimation",
+                    "abstract": "Training deep networks for semantic segmentation requires large amounts of\nlabeled training data, which presents a major challenge in practice, as\nlabeling segmentation masks is a highly labor-intensive process. To address\nthis issue, we present a framework for semi-supervised semantic segmentation,\nwhich is enhanced by self-supervised monocular depth estimation from unlabeled\nimage sequences. In particular, we propose three key contributions: (1) We\ntransfer knowledge from features learned during self-supervised depth\nestimation to semantic segmentation, (2) we implement a strong data\naugmentation by blending images and labels using the geometry of the scene, and\n(3) we utilize the depth feature diversity as well as the level of difficulty\nof learning depth in a student-teacher framework to select the most useful\nsamples to be annotated for semantic segmentation. We validate the proposed\nmodel on the Cityscapes dataset, where all three modules demonstrate\nsignificant performance gains, and we achieve state-of-the-art results for\nsemi-supervised semantic segmentation. The implementation is available at\nhttps://github.com/lhoyer/improving_segmentation_with_selfsupervised_depth.",
+                    "authors": "Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian K\u00f6ring, Suman Saha, Luc Van Gool",
+                    "published": "2020-12-19",
+                    "updated": "2021-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.08465v2",
+                    "title": "DeepAtlas: Joint Semi-Supervised Learning of Image Registration and Segmentation",
+                    "abstract": "Deep convolutional neural networks (CNNs) are state-of-the-art for semantic\nimage segmentation, but typically require many labeled training samples.\nObtaining 3D segmentations of medical images for supervised training is\ndifficult and labor intensive. Motivated by classical approaches for joint\nsegmentation and registration we therefore propose a deep learning framework\nthat jointly learns networks for image registration and image segmentation. In\ncontrast to previous work on deep unsupervised image registration, which showed\nthe benefit of weak supervision via image segmentations, our approach can use\nexisting segmentations when available and computes them via the segmentation\nnetwork otherwise, thereby providing the same registration benefit. Conversely,\nsegmentation network training benefits from the registration, which essentially\nprovides a realistic form of data augmentation. Experiments on knee and brain\n3D magnetic resonance (MR) images show that our approach achieves large\nsimultaneous improvements of segmentation and registration accuracy (over\nindependently trained networks) and allows training high-quality models with\nvery limited training data. Specifically, in a one-shot-scenario (with only one\nmanually labeled image) our approach increases Dice scores (%) over an\nunsupervised registration network by 2.7 and 1.8 on the knee and brain images\nrespectively.",
+                    "authors": "Zhenlin Xu, Marc Niethammer",
+                    "published": "2019-04-17",
+                    "updated": "2019-07-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.13978v3",
+                    "title": "Personalized Image Semantic Segmentation",
+                    "abstract": "Semantic segmentation models trained on public datasets have achieved great\nsuccess in recent years. However, these models didn't consider the\npersonalization issue of segmentation though it is important in practice. In\nthis paper, we address the problem of personalized image segmentation. The\nobjective is to generate more accurate segmentation results on unlabeled\npersonalized images by investigating the data's personalized traits. To open up\nfuture research in this area, we collect a large dataset containing various\nusers' personalized images called PIS (Personalized Image Semantic\nSegmentation). We also survey some recent researches related to this problem\nand report their performance on our dataset. Furthermore, by observing the\ncorrelation among a user's personalized images, we propose a baseline method\nthat incorporates the inter-image context when segmenting certain images.\nExtensive experiments show that our method outperforms the existing methods on\nthe proposed dataset. The code and the PIS dataset will be made publicly\navailable.",
+                    "authors": "Yu Zhang, Chang-Bin Zhang, Peng-Tao Jiang, Ming-Ming Cheng, Feng Mao",
+                    "published": "2021-07-24",
+                    "updated": "2021-09-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2004.10126v3",
+                    "title": "Generative Synthetic Augmentation using Label-to-Image Translation for Nuclei Image Segmentation",
+                    "abstract": "In medical image diagnosis, pathology image analysis using semantic\nsegmentation becomes important for efficient screening as a field of digital\npathology. The spatial augmentation is ordinary used for semantic segmentation.\nTumor images under malignant are rare and to annotate the labels of nuclei\nregion takes much time-consuming. We require an effective use of dataset to\nmaximize the segmentation accuracy. It is expected that some augmentation to\ntransform generalized images influence the segmentation performance. We propose\na synthetic augmentation using label-to-image translation, mapping from a\nsemantic label with the edge structure to a real image. Exactly this paper deal\nwith stain slides of nuclei in tumor. Actually, we demonstrate several\nsegmentation algorithms applied to the initial dataset that contains real\nimages and labels using synthetic augmentation in order to add their\ngeneralized images. We computes and reports that a proposed synthetic\naugmentation procedure improve their accuracy.",
+                    "authors": "Takato Yasuno",
+                    "published": "2020-04-21",
+                    "updated": "2021-03-02",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "eess.IV",
+                        "stat.ML",
+                        "I.4.6; I.2.6"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2007.02361v1",
+                    "title": "Self-supervised Depth Estimation to Regularise Semantic Segmentation in Knee Arthroscopy",
+                    "abstract": "Intra-operative automatic semantic segmentation of knee joint structures can\nassist surgeons during knee arthroscopy in terms of situational awareness.\nHowever, due to poor imaging conditions (e.g., low texture, overexposure,\netc.), automatic semantic segmentation is a challenging scenario, which\njustifies the scarce literature on this topic. In this paper, we propose a\nnovel self-supervised monocular depth estimation to regularise the training of\nthe semantic segmentation in knee arthroscopy. To further regularise the depth\nestimation, we propose the use of clean training images captured by the stereo\narthroscope of routine objects (presenting none of the poor imaging conditions\nand with rich texture information) to pre-train the model. We fine-tune such\nmodel to produce both the semantic segmentation and self-supervised monocular\ndepth using stereo arthroscopic images taken from inside the knee. Using a data\nset containing 3868 arthroscopic images captured during cadaveric knee\narthroscopy with semantic segmentation annotations, 2000 stereo image pairs of\ncadaveric knee arthroscopy, and 2150 stereo image pairs of routine objects, we\nshow that our semantic segmentation regularised by self-supervised depth\nestimation produces a more accurate segmentation than a state-of-the-art\nsemantic segmentation approach modeled exclusively with semantic segmentation\nannotation.",
+                    "authors": "Fengbei Liu, Yaqub Jonmohamadi, Gabriel Maicas, Ajay K. Pandey, Gustavo Carneiro",
+                    "published": "2020-07-05",
+                    "updated": "2020-07-05",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2401.11739v1",
+                    "title": "EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models",
+                    "abstract": "Diffusion models have recently received increasing research attention for\ntheir remarkable transfer abilities in semantic segmentation tasks. However,\ngenerating fine-grained segmentation masks with diffusion models often requires\nadditional training on annotated datasets, leaving it unclear to what extent\npre-trained diffusion models alone understand the semantic relations of their\ngenerated images. To address this question, we leverage the semantic knowledge\nextracted from Stable Diffusion (SD) and aim to develop an image segmentor\ncapable of generating fine-grained segmentation maps without any additional\ntraining. The primary difficulty stems from the fact that semantically\nmeaningful feature maps typically exist only in the spatially lower-dimensional\nlayers, which poses a challenge in directly extracting pixel-level semantic\nrelations from these feature maps. To overcome this issue, our framework\nidentifies semantic correspondences between image pixels and spatial locations\nof low-dimensional feature maps by exploiting SD's generation process and\nutilizes them for constructing image-resolution segmentation maps. In extensive\nexperiments, the produced segmentation maps are demonstrated to be well\ndelineated and capture detailed parts of the images, indicating the existence\nof highly accurate pixel-level semantic knowledge in diffusion models.",
+                    "authors": "Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, Seung Wook Kim",
+                    "published": "2024-01-22",
+                    "updated": "2024-01-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2312.17243v1",
+                    "title": "Unsupervised Universal Image Segmentation",
+                    "abstract": "Several unsupervised image segmentation approaches have been proposed which\neliminate the need for dense manually-annotated segmentation masks; current\nmodels separately handle either semantic segmentation (e.g., STEGO) or\nclass-agnostic instance segmentation (e.g., CutLER), but not both (i.e.,\npanoptic segmentation). We propose an Unsupervised Universal Segmentation model\n(U2Seg) adept at performing various image segmentation tasks -- instance,\nsemantic and panoptic -- using a novel unified framework. U2Seg generates\npseudo semantic labels for these segmentation tasks via leveraging\nself-supervised models followed by clustering; each cluster represents\ndifferent semantic and/or instance membership of pixels. We then self-train the\nmodel on these pseudo semantic labels, yielding substantial performance gains\nover specialized methods tailored to each task: a +2.6 AP$^{\\text{box}}$ boost\nvs. CutLER in unsupervised instance segmentation on COCO and a +7.0 PixelAcc\nincrease (vs. STEGO) in unsupervised semantic segmentation on COCOStuff.\nMoreover, our method sets up a new baseline for unsupervised panoptic\nsegmentation, which has not been previously explored. U2Seg is also a strong\npretrained model for few-shot segmentation, surpassing CutLER by +5.0\nAP$^{\\text{mask}}$ when trained on a low-data regime, e.g., only 1% COCO\nlabels. We hope our simple yet effective method can inspire more research on\nunsupervised universal image segmentation.",
+                    "authors": "Dantong Niu, Xudong Wang, Xinyang Han, Long Lian, Roei Herzig, Trevor Darrell",
+                    "published": "2023-12-28",
+                    "updated": "2023-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.12518v2",
+                    "title": "Segmentation in Style: Unsupervised Semantic Image Segmentation with Stylegan and CLIP",
+                    "abstract": "We introduce a method that allows to automatically segment images into\nsemantically meaningful regions without human supervision. Derived regions are\nconsistent across different images and coincide with human-defined semantic\nclasses on some datasets. In cases where semantic regions might be hard for\nhuman to define and consistently label, our method is still able to find\nmeaningful and consistent semantic classes. In our work, we use pretrained\nStyleGAN2 generative model: clustering in the feature space of the generative\nmodel allows to discover semantic classes. Once classes are discovered, a\nsynthetic dataset with generated images and corresponding segmentation masks\ncan be created. After that a segmentation model is trained on the synthetic\ndataset and is able to generalize to real images. Additionally, by using CLIP\nwe are able to use prompts defined in a natural language to discover some\ndesired semantic classes. We test our method on publicly available datasets and\nshow state-of-the-art results.",
+                    "authors": "Daniil Pakhomov, Sanchit Hira, Narayani Wagle, Kemar E. Green, Nassir Navab",
+                    "published": "2021-07-26",
+                    "updated": "2021-11-18",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.17083v1",
+                    "title": "SegRCDB: Semantic Segmentation via Formula-Driven Supervised Learning",
+                    "abstract": "Pre-training is a strong strategy for enhancing visual models to efficiently\ntrain them with a limited number of labeled images. In semantic segmentation,\ncreating annotation masks requires an intensive amount of labor and time, and\ntherefore, a large-scale pre-training dataset with semantic labels is quite\ndifficult to construct. Moreover, what matters in semantic segmentation\npre-training has not been fully investigated. In this paper, we propose the\nSegmentation Radial Contour DataBase (SegRCDB), which for the first time\napplies formula-driven supervised learning for semantic segmentation. SegRCDB\nenables pre-training for semantic segmentation without real images or any\nmanual semantic labels. SegRCDB is based on insights about what is important in\npre-training for semantic segmentation and allows efficient pre-training.\nPre-training with SegRCDB achieved higher mIoU than the pre-training with\nCOCO-Stuff for fine-tuning on ADE-20k and Cityscapes with the same number of\ntraining images. SegRCDB has a high potential to contribute to semantic\nsegmentation pre-training and investigation by enabling the creation of large\ndatasets without manual annotation. The SegRCDB dataset will be released under\na license that allows research and commercial use. Code is available at:\nhttps://github.com/dahlian00/SegRCDB",
+                    "authors": "Risa Shinoda, Ryo Hayamizu, Kodai Nakashima, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka",
+                    "published": "2023-09-29",
+                    "updated": "2023-09-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2205.13278v1",
+                    "title": "Semantic Segmentation for Thermal Images: A Comparative Survey",
+                    "abstract": "Semantic segmentation is a challenging task since it requires excessively\nmore low-level spatial information of the image compared to other computer\nvision problems. The accuracy of pixel-level classification can be affected by\nmany factors, such as imaging limitations and the ambiguity of object\nboundaries in an image. Conventional methods exploit three-channel RGB images\ncaptured in the visible spectrum with deep neural networks (DNN). Thermal\nimages can significantly contribute during the segmentation since thermal\nimaging cameras are capable of capturing details despite the weather and\nillumination conditions. Using infrared spectrum in semantic segmentation has\nmany real-world use cases, such as autonomous driving, medical imaging,\nagriculture, defense industry, etc. Due to this wide range of use cases,\ndesigning accurate semantic segmentation algorithms with the help of infrared\nspectrum is an important challenge. One approach is to use both visible and\ninfrared spectrum images as inputs. These methods can accomplish higher\naccuracy due to enriched input information, with the cost of extra effort for\nthe alignment and processing of multiple inputs. Another approach is to use\nonly thermal images, enabling less hardware cost for smaller use cases. Even\nthough there are multiple surveys on semantic segmentation methods, the\nliterature lacks a comprehensive survey centered explicitly around semantic\nsegmentation using infrared spectrum. This work aims to fill this gap by\npresenting algorithms in the literature and categorizing them by their input\nimages.",
+                    "authors": "Z\u00fclfiye K\u00fct\u00fck, G\u00f6rkem Algan",
+                    "published": "2022-05-26",
+                    "updated": "2022-05-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.11506v1",
+                    "title": "LCCo: Lending CLIP to Co-Segmentation",
+                    "abstract": "This paper studies co-segmenting the common semantic object in a set of\nimages. Existing works either rely on carefully engineered networks to mine the\nimplicit semantic information in visual features or require extra data (i.e.,\nclassification labels) for training. In this paper, we leverage the contrastive\nlanguage-image pre-training framework (CLIP) for the task. With a backbone\nsegmentation network that independently processes each image from the set, we\nintroduce semantics from CLIP into the backbone features, refining them in a\ncoarse-to-fine manner with three key modules: i) an image set feature\ncorrespondence module, encoding global consistent semantic information of the\nimage set; ii) a CLIP interaction module, using CLIP-mined common semantics of\nthe image set to refine the backbone feature; iii) a CLIP regularization\nmodule, drawing CLIP towards this co-segmentation task, identifying the best\nCLIP semantic and using it to regularize the backbone feature. Experiments on\nfour standard co-segmentation benchmark datasets show that the performance of\nour method outperforms state-of-the-art methods.",
+                    "authors": "Xin Duan, Yan Yang, Liyuan Pan, Xiabi Liu",
+                    "published": "2023-08-22",
+                    "updated": "2023-08-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1610.01706v1",
+                    "title": "Exploiting Depth from Single Monocular Images for Object Detection and Semantic Segmentation",
+                    "abstract": "Augmenting RGB data with measured depth has been shown to improve the\nperformance of a range of tasks in computer vision including object detection\nand semantic segmentation. Although depth sensors such as the Microsoft Kinect\nhave facilitated easy acquisition of such depth information, the vast majority\nof images used in vision tasks do not contain depth information. In this paper,\nwe show that augmenting RGB images with estimated depth can also improve the\naccuracy of both object detection and semantic segmentation. Specifically, we\nfirst exploit the recent success of depth estimation from monocular images and\nlearn a deep depth estimation model. Then we learn deep depth features from the\nestimated depth and combine with RGB features for object detection and semantic\nsegmentation. Additionally, we propose an RGB-D semantic segmentation method\nwhich applies a multi-task training scheme: semantic label prediction and depth\nvalue regression. We test our methods on several datasets and demonstrate that\nincorporating information from estimated depth improves the performance of\nobject detection and semantic segmentation remarkably.",
+                    "authors": "Yuanzhouhan Cao, Chunhua Shen, Heng Tao Shen",
+                    "published": "2016-10-06",
+                    "updated": "2016-10-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.02095v1",
+                    "title": "Content-aware Token Sharing for Efficient Semantic Segmentation with Vision Transformers",
+                    "abstract": "This paper introduces Content-aware Token Sharing (CTS), a token reduction\napproach that improves the computational efficiency of semantic segmentation\nnetworks that use Vision Transformers (ViTs). Existing works have proposed\ntoken reduction approaches to improve the efficiency of ViT-based image\nclassification networks, but these methods are not directly applicable to\nsemantic segmentation, which we address in this work. We observe that, for\nsemantic segmentation, multiple image patches can share a token if they contain\nthe same semantic class, as they contain redundant information. Our approach\nleverages this by employing an efficient, class-agnostic policy network that\npredicts if image patches contain the same semantic class, and lets them share\na token if they do. With experiments, we explore the critical design choices of\nCTS and show its effectiveness on the ADE20K, Pascal Context and Cityscapes\ndatasets, various ViT backbones, and different segmentation decoders. With\nContent-aware Token Sharing, we are able to reduce the number of processed\ntokens by up to 44%, without diminishing the segmentation quality.",
+                    "authors": "Chenyang Lu, Daan de Geus, Gijs Dubbelman",
+                    "published": "2023-06-03",
+                    "updated": "2023-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2112.03241v1",
+                    "title": "Unsupervised Domain Adaptation for Semantic Image Segmentation: a Comprehensive Survey",
+                    "abstract": "Semantic segmentation plays a fundamental role in a broad variety of computer\nvision applications, providing key information for the global understanding of\nan image. Yet, the state-of-the-art models rely on large amount of annotated\nsamples, which are more expensive to obtain than in tasks such as image\nclassification. Since unlabelled data is instead significantly cheaper to\nobtain, it is not surprising that Unsupervised Domain Adaptation reached a\nbroad success within the semantic segmentation community.\n  This survey is an effort to summarize five years of this incredibly rapidly\ngrowing field, which embraces the importance of semantic segmentation itself\nand a critical need of adapting segmentation models to new environments. We\npresent the most important semantic segmentation methods; we provide a\ncomprehensive survey on domain adaptation techniques for semantic segmentation;\nwe unveil newer trends such as multi-domain learning, domain generalization,\ntest-time adaptation or source-free domain adaptation; we conclude this survey\nby describing datasets and benchmarks most widely used in semantic segmentation\nresearch. We hope that this survey will provide researchers across academia and\nindustry with a comprehensive reference guide and will help them in fostering\nnew research directions in the field.",
+                    "authors": "Gabriela Csurka, Riccardo Volpi, Boris Chidlovskii",
+                    "published": "2021-12-06",
+                    "updated": "2021-12-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "I.4.6; I.2"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.08455v3",
+                    "title": "Open-world Semantic Segmentation via Contrasting and Clustering Vision-Language Embedding",
+                    "abstract": "To bridge the gap between supervised semantic segmentation and real-world\napplications that acquires one model to recognize arbitrary new concepts,\nrecent zero-shot segmentation attracts a lot of attention by exploring the\nrelationships between unseen and seen object categories, yet requiring large\namounts of densely-annotated data with diverse base classes. In this paper, we\npropose a new open-world semantic segmentation pipeline that makes the first\nattempt to learn to segment semantic objects of various open-world categories\nwithout any efforts on dense annotations, by purely exploiting the\nimage-caption data that naturally exist on the Internet. Our method,\nVision-language-driven Semantic Segmentation (ViL-Seg), employs an image and a\ntext encoder to generate visual and text embeddings for the image-caption data,\nwith two core components that endow its segmentation ability: First, the image\nencoder is jointly trained with a vision-based contrasting and a cross-modal\ncontrasting, which encourage the visual embeddings to preserve both\nfine-grained semantics and high-level category information that are crucial for\nthe segmentation task. Furthermore, an online clustering head is devised over\nthe image encoder, which allows to dynamically segment the visual embeddings\ninto distinct semantic groups such that they can be classified by comparing\nwith various text embeddings to complete our segmentation pipeline. Experiments\nshow that without using any data with dense annotations, our method can\ndirectly segment objects of arbitrary categories, outperforming zero-shot\nsegmentation methods that require data labeling on three benchmark datasets.",
+                    "authors": "Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, Xiaodan Liang",
+                    "published": "2022-07-18",
+                    "updated": "2022-10-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                }
+            ],
+            [
+                {
+                    "url": "http://arxiv.org/abs/2311.06746v1",
+                    "title": "Two Stream Scene Understanding on Graph Embedding",
+                    "abstract": "The paper presents a novel two-stream network architecture for enhancing\nscene understanding in computer vision. This architecture utilizes a graph\nfeature stream and an image feature stream, aiming to merge the strengths of\nboth modalities for improved performance in image classification and scene\ngraph generation tasks. The graph feature stream network comprises a\nsegmentation structure, scene graph generation, and a graph representation\nmodule. The segmentation structure employs the UPSNet architecture with a\nbackbone that can be a residual network, Vit, or Swin Transformer. The scene\ngraph generation component focuses on extracting object labels and neighborhood\nrelationships from the semantic map to create a scene graph. Graph\nConvolutional Networks (GCN), GraphSAGE, and Graph Attention Networks (GAT) are\nemployed for graph representation, with an emphasis on capturing node features\nand their interconnections. The image feature stream network, on the other\nhand, focuses on image classification through the use of Vision Transformer and\nSwin Transformer models. The two streams are fused using various data fusion\nmethods. This fusion is designed to leverage the complementary strengths of\ngraph-based and image-based features.Experiments conducted on the ADE20K\ndataset demonstrate the effectiveness of the proposed two-stream network in\nimproving image classification accuracy compared to conventional methods. This\nresearch provides a significant contribution to the field of computer vision,\nparticularly in the areas of scene understanding and image classification, by\neffectively combining graph-based and image-based approaches.",
+                    "authors": "Wenkai Yang, Wenyuan Sun, Runxaing Huang",
+                    "published": "2023-11-12",
+                    "updated": "2023-11-12",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "label": "Original Paper",
+                    "paper_cat": "Semantic AND Segmentation AND Image",
+                    "gt": "The scene graph generation has been emerged as a linchpin in enhancing machine understanding of visual scenes by images into structured graphical representations. Xu.[1] was primarily tried on object detection and relationship between every two objects prediction using CNN[2][3] and Region Proposal Network(RPN). This work directly utilize the set of object proposals to predict the object categories, bounding box offsets, and relationship types by Gated Recurrent Unit (GRU)[4] and graph pooling layer. Alejandro Newell et al.[5] proposed a single-stage CNN for object and relationship detection, transitioning from pixel representations to graph structures through two heatmaps. These heatmaps activate at the predicted locations of objects and relationships, providing a useful representation for various computer vision tasks. In 2021, attention mechanisms gained traction in computer vision tasks, exemplified by the advent of the Image Transformer[6]. Empirical evidence suggests that, compared to CNNs, Transformers could exhibit superior performance on large datasets, especially when pre-trained. This theoretical advancement proved beneficial for later multimodal works, such as CLIP (Contrastive Language-Image Pre-training)[7], as both computer vision and text processing tasks could leverage the same architectural underpinning to extract corresponding features. Following this trend, the Swin Transformer was introduced[8], which incorporates shifted window attention and local window attention mechanisms, similiar to the operation of CNNs. This innovation broadened the scope of Transformers, making them applicable to other computer vision tasks like segmentation. The first two stream network has been used in the action recognition in video[9]. Simonyan first propose the two-stream ConvNet architecture which incorporates spatial and temporal networks. The spatial stream network input is the common singe RGB image, while the temporal stream network input is the multi-frame optical flow. To the action video classification, the temporal stream network could highly improve the accuracy and complement the lacking modality information. Actually after the work, there is work[10] combine the LSTM[11] and CNN to sovle the video classification. But there is some difference between theses two works, the Simonyan\u2019s work use two parallel two stream networks, however the the gammulle directly use CNN extract every frame feature and feed them as a sequence into the LSTM network. The Kipf[12] introduce the graph convolutional operator to do Semi-supervised graph classification, and the hamilton[13] utilize inductive framework that leverages node feature information (eg, text attributes) to efficiently generate node embeddings. 3 After the transformer[14] appeared, the velivckovic [15] applied the attention on the graph-structured data. Generally we have tried some different data fusion like late fusion[16], and the different fusion methods have been adapted such as concatenation, cross attention, voting, averaging and Canonical Correlation Analysis (CCA)[17]. The contrastive learning[18] is a powerful self-supervised learning technique. The primary goal is to learn a common representation space where similar items from different modalities are close to each other, while dissimilar items are far apart. In CLIP, this is achieved by training a neural network to map images and text descriptions into a shared embedding space. Utilize a contrastive loss function, such as the Noise Contrastive Estimation (NCE)[19] loss or InfoNCE[20] loss, to train the model. The loss function encourages the model to minimize the distance between positive pairs and maximize the distance between negative pairs in the shared embedding space. In our case, the positive pair is the same image of graph embedding and pixel embedding, and the negative pair is the different image of graph embedding and pixel embedding.",
+                    "pre_questions": [],
+                    "main_content": "Introduction Scene understanding is a complex field within computer vision that aims to enable machines to interpret and understand visual data in a manner similar to human perception. The objective is to create systems that can identify and localize individual objects within an image while also comprehending the relationships and functionalities between these objects. Scene graph generation is instrumental for tasks related to scene understanding. The generation of scene graphs often relies on scene segmentation, a long-standing and challenging problem in computer vision with numerous downstream applications such as augmented reality, autonomous driving, human-machine interaction, and video content analysis. The primary motivation for our work in extracting scene graphs from images or videos lies in the graph structure\u2019s significant utility for scene understanding. Regardless of image resizing or cropping, the fundamental objects and their relationships within the scene are preserved, making the graph a robust form of representation. To combine features from the graph representation and common image features obtainable through Convolutional Neural Networks (CNN) or Vision Transformer frameworks, we will employ multi-modal methods for data fusion. Our architecture utilizes a two-stream model for this multi-modal computer vision task. One stream leverages either CNN or Transformer frameworks as the backbone for segmentation, and uses scene graph generation techniques to convert the semantic map into a simplified graph, which is then represented through graph embeddings. The other stream employs conventional strategies, using either the Swin Transformer or ViT (Vision Transformer) for image classification. The fusion methods for combining features from these two streams can be either concatenation or cross-attention mechanisms. The relationship between objects is crucial for scene understanding, as it is not explicitly captured by existing frameworks. While humans can readily identify the main objects and their relationships, image classification models like CNN or ViT primarily focus on recognizing patterns and pixel-level features in images, rather than understanding the scene as a whole. The two-stream network offers a solution by helping the model comprehend different objects, their positions, and spatial relationships within a scene. In this paper, we introduce a novel two-stream network architecture leveraging graph representation and Swin Transformer to enhance both image classification and scene graph generation tasks. This advanced scheme transcends conventional image classification approaches that predominantly rely on a single modality, such as CNN or attention mechanisms. The primary contributions of our work are outlined below: 2 \u2022 We employ a two-stream network, training each stream independently using the same loss function to predict either the graph or image label. This design facilitates a seamless fusion of different modalities. \u2022 a seamless fusion of different modalities. \u2022 In transitioning from segmentation to graph, we have incorporated scene graph generation techniques to elucidate the relationships between neighboring objects effectively. \u2022 effectively. \u2022 We have implemented diverse data fusion strategies to ensure coherent alignment of graph and image features. 3.1 Graph Feature Stream Network 3.1.1 The Segmentation Structure In segmentation part, we use the Upsnet[21] as whole architecture, which is a unified panoptic segmentation network for tackling the newly proposed segmentation task. The backbone could be residual network, Vision Transformer(Vit)[22] and Swin Transformer. The deformable convolution based semantic segmentation head and a Mask R-CNN style instance segmentation head have been used to solve these two subtasks(Semantic Segmentation and Instance Segmentation) simultaneously. We adopt the original Mask R-CNN[23] backbone and Swin Transformer as our feature extraction network. The Mask R-CNN backbone exploits a deep residual network[24] (ResNet) with a feature pyramid network[25] (FPN). The architecture is in Fig.1 Fig. 1 The architecture of Segmentation 4 The goal of the semantic segmentation head is to segment all semantic classes, Our semantic head consists of the multi-scale feature from FPN as input. In particular, we use P2, P3, P4 and P5 feature maps of FPN which contain 256 channels and are 1/4, 1/8, 1/16 and 1/32 of the original scale respectively. These feature maps first go through a deformable convolution network[26] independently and are subsequently upsampled to the 1/4 scale. We then concatenate them and apply 1 \u00d7 1 convolutions with softmax to predict the semantic class. The architecture is shown in Fig. 2. Semantic segmentation head is associated with the regular pixel-wise cross entropy loss. To put more emphasis on the foreground objects such as pedestrians, the RoI loss is also incorporated. Fig. 2 The architecture of semantic segmentation head Pixel-wise cross entropy is often used for semantic segmentation tasks where each pixel of an image is classified into one of C classes. The loss for each pixel is computed as the cross entropy between the predicted probabilities and the true class. Given that pc(i, j) is the predicted probability of pixel at position (i, j) belonging to class c and yc(i, j) is the ground truth label (1 if the pixel belongs to class c, 0 otherwise), the pixel-wise cross entropy loss L can be given as: L = \u2212 X i,j C X c=1 yc(i, j) log(pc(i, j)) (1) We use the Swin Transformer as the backbone in UPSNet. Swin Transformer divides an input image into non-overlapping patches, linearly embeds them, and then feeds them into a series of Transformer layers. Unlike conventional Transformers, the Swin Transformer uses shifted windows to allow local features to be aggregated over layers. This results in multiple levels of feature resolutions throughout its layers, similar to the hierarchical features in a CNN. The swin architecture with FPN is shown in Fig. 3. 5 Fig. 3 The architecture of Swin Transformer 3.1.2 The Scene Graph Generation The success of segmentation or detection has surged interest in examining the detailed structures of a visual scene, especially in the form of object relationships. Scene graph offers a platform to explicitly model objects and their relationships. In short, a scene graph is a visually-grounded graph over the object instances in an image, where the edges depict their pairwise relationships, see example in Fig.4. In this work, we address the problem of scene graph generation, where the goal is to generate a visually-grounded scene graph from an image segmentation. Fig. 4 The example of Scene Graph Generation From the semantic map, object labels can be extracted and according to the boundary between different labels, a neighborhood relationship between objects can be established. Therefore, we need to detect the boundary of pixels in the semantic 6 map row by row and column by column, so as to get the possible neighborhood relationship between different objects of the original image, as shown in Fig. 5. Depending on the object category labels and the neighborhood relationship between different objects in original image, we can obtain the scene graph. First, we convert the labels of Fig. 5 The illustration of generating scene graph objects, which is provided by the semantic map consisting of C classes into the nodes in scene graph. The node feature could be represented by the vector(C length) via one hot encoding. The neighborhood relationship is determined simply by the presence or absence of pixel intersection and could be regraded as the edge(the pair of two nodes) in scene graph. 3.1.3 The Graph Representation Graph Neural Network(GNN) recently has received a lot of attention due to its ability to analyze graph structural data. A set of objects, and the connections between them, are naturally expressed as a graph. There are three general types of prediction tasks on graphs: graph-level, node-level, and edge-level. In a graph-level task, we predict a single property for a whole graph. For a node-level task, we predict some property for each node in a graph. For an edge-level task, we want to predict the property or presence of edges in a graph. In this work, we will predict the label of graph means the corresponding image, because we intend to get the feature of graph. Graph Convolutional Network(GCN)[12] layer is one of the pioneering methods in GNNs. The key idea of a GCN layer is to update the representation of a node by aggregating information from its neighbors. Given: 7 \u2022 A: adjacency matrix of the graph \u2022 X: node features matrix \u2022 H(l) : node representations at layer l \u2022 W (l): weight matrix at layer l The basic update rule for a GCN layer is: H(l+1) = \u03c3 \u0010 \u02dc D\u22121 2 \u02dc A \u02dc D\u22121 2 H(l)W (l)\u0011 (2) Where: \u2022 \u02dc A = A + IN is the adjacency matrix of the undirected graph G with added selfconnections, IN is the identity matrix. \u2022 \u02dc Dii = P j \u02dc Aij and W (l) is a layer-specific trainable weight matrix. \u2022 \u03c3 is a non-linear activation function, e.g. ReLU. GCN uses a fixed-weight averaging scheme based on the adjacency matrix of the graph. GraphSAGE (Graph Sample and Aggregation)[13] is another popular GNN method, which is designed to generate embeddings by sampling and aggregating features from a node\u2019s neighbors. The key steps in a GraphSAGE layer are: \u2022 Neighbor Sampling: Sample a fixed-size set of neighbors for each node. \u2022 Feature Aggregation: Aggregate the features of the sampled neighbors. \u2022 Combination: Combine a node\u2019s current features with the aggregated neighborhood features. The aggregation function can be mean, LSTM, pooling, etc. For the mean aggregator, the update rule can be given as: h(l) neigh(v) = MEAN \u0010 {h(l\u22121) u : u \u2208N(v)} \u0011 (3) h(l) v = \u03c3 \u0010 W \u00b7 CONCAT \u0010 h(l\u22121) v , h(l) neigh(v) \u0011\u0011 (4) Where: \u2022 h(l\u22121) v is the representation of node v at layer l. \u2022 N(v) is the set of neighbors of node v. GraphSAGE uses neighborhood sampling and various aggregation functions to update node embeddings. GAT (Graph Attention Network)[15] introduces the concept of attention mechanisms to the graph domain. The main idea is to assign different attention weights to different neighbors of a node, indicating the importance of each neighbor\u2019s information. Given: \u2022 W: weight matrix \u2022 a: attention mechanism\u2019s weight \u2022 hi and hj : node feature representations 8 The attention coefficient between two nodes ii and jj can be computed as: eij = LeakyReLU \u0000aT [Whi||Whj] \u0001 (5) Where \u2225denotes concatenation. These coefficients are then normalized across all choices of j using a softmax: \u03b1ij = exp(eij) P k\u2208N(i) exp(eik) (6) The new node representation can then be computed as a weighted sum of its neighbors: h\u2032 i = \u03c3 \uf8eb \uf8edX j\u2208N(i) \u03b1ijWhj \uf8f6 \uf8f8 (7) GAT employs attention mechanisms to weigh the importance of each neighbor during aggregation. The graph will pass three GNN Layers where each layer performs node feature propagation aggregating information from its neighbor nodes and follow a Readout Layer which could aggregate information from all nodes in a graph to produce a single vector representation of the entire graph, shown in Fig. 6. Common readout functions include Sum, Mean and Max-pooling. Fig. 6 The architecture of generating scene graph 3.2 Image Feature Stream Network 3.2.1 The Classification Net The image classification is a common computer vision task which could be solved by Conv or Transformer. In this section, we will try to use Vision Transformer shown in Fig.7 and Swin Transformer shown in Fig.3 to do image classification. Vit utilize self-attention mechanisms, originally used in NLP tasks, to capture global dependencies within images. The Vision Transformer treats an image as a 9 sequence of fixed-size patches and processes these patches through a series of transformer blocks. Divide the image into fixed-size patches (e.g., 16x16 pixels). Flatten each patch and linearly embed each of them with a trainable linear projection to create patch embeddings. Optionally add positional embeddings to retain positional information. Pass the sequence of patch embeddings through multiple layers of transformer encoders. Each encoder layer includes multi-head self-attention and MLP blocks. Swin introduces shifted windows, which limit self-attention computation to nonoverlapping local windows while also allowing for cross-window connections. The preprocessing is same as ViT, but with potentially different patch sizes. The image is divided into non-overlapping windows. Within each window, self-attention is computed (W-MSA Window based Multi-head Self Attention). Every alternate block shifts the windows to allow for cross-window connection (SW-MSA Shifted Window based Multi-head Self Attention). Adjacent patches are merged progressively to reduce the number of tokens and increase the receptive field. The final layer\u2019s feature map is used for classification, often through a global average pooling layer followed by a fully connected layer (classifier). In this part we will use the pre-trained model on ImageNet-21k[27] and full parameter fine-tuning on our ADE20K dataset. Fig. 7 The architecture of Vision Transformer The image feature we need is the cls token after the Transformer Encoder in Vision Transformer or the feature map after stage 4 Swin Transformer Block. 3.3 The Two Stream Network The two-stream network operates on separate training pathways, where the objective of the loss function is to accurately predict the image\u2019s label. The graph network 10 is designed to distill object features and their interrelations within the image, effectively mapping out a graph-like representation of the scene. Concurrently, the image classification network applies conventional methods to categorize the image. Our aim is to extract features that remain consistent regardless of image transformations such as resizing, cropping, or rotation\u2014mirroring the human capability to interpret a scene by discerning the constituent objects and their spatial relationships. We posit that these features closely align with a spatial graph representation. By combining graph-based features with those derived from image analysis, we anticipate a bolstered performance in image classification and related tasks. This fusion strategy leverages the complementary strengths of both feature sets to enhance the model\u2019s robustness and interpretive accuracy.The whole two stream network shown in Fig.8 Fig. 8 The architecture of Two Stream Network 3.3.1 The Features Fusion The common data fusion methods are concatenation, sum, average, product and voting. In this paper we utilize the cross attention mechanism[28][29] to do the late fusion. The attention mechanism is often use in the multi-modility work. Cross-attention is a mechanism often found in transformer architectures that allows the model to weigh the importance of different parts of input data from one source when processing another. It\u2019s a type of attention that enables the model to attend across different modalities or feature sets, making it particularly useful for tasks that involve multiple types of data (e.g., text and images, audio and video). Cross-attention can be employed to allow one modality to influence the processing of another. When you have two different modality vectors, such as an image vector and a graph representation vector, cross-attention can be used to fuse these vectors by allowing each modality to attend to and influence the representation of the other. Implement a cross-attention layer that takes the encoded vectors from both modalities. In this layer, one modality serves as the \u201dquery\u201d while the other serves as the \u201dkey\u201d and \u201dvalue\u201d (following the standard transformer nomenclature), shown in Fig. 8. Queries, keys, and values are projected into a shared space where the compatibility 11 of the query with each key is computed, often using a dot-product attention mechanism. The output is a weighted sum of the value vectors, with weights corresponding to the computed attention scores. After cross-attention, the updated vectors that now contain information from both modalities can be used in the corresponding head layer in different tasks. In classification task, we will use a fully connected layer to do projection from the cross-attention output dimension to the classes size. 4 Experiments 4.1 Dataset The ADE20K[30] dataset is a widely used benchmark in the field of computer vision, particularly in the context of semantic segmentation. In deep learning research, it is widely used in segmentation task. ADE20K is an openly available dataset, making it convenient for comparing and evaluating the performance of various algorithms and models. This dataset comprises a diverse collection of images with pixel-level annotations, scene or image label, covering a broad spectrum of scenes and objects. With over 20,000 images and more than 150 object categories(grass, sky, building etc.), and 1055 scecne categories(airport, basketball court etc.), ADE20K provides a rich and varied source of data for training and testing deep learning algorithms, shown in Fig. 9. The annotated images cover the scene categories from the SUN and Places database. Here there are some examples showing the images, object segmentations, and parts segmentations. It enable you to explore and develop advanced models to achieve stateof-the-art results and potentially contribute to the publication of impressive research papers in this domain. Fig. 9 The overview of ADE20k Dataset 12 4.2 Model Settings The Segmentation Net: We conduct experiments on the ADE20K dataset using Swin-L, Swin-B, Convnext-L and Convnext-B which are pre-trained on the ImageNet21k. The image augmentations have been applied, including normalizing, resize and crop, which could let the image has standard size(512, 512). More details on Table. 2 Optimizer Learning Rate Weight Decay Batch Size Hardware Epochs SGD 1e-6 5e-7 6 RTX 3090 20 Adam 6e-5 1e-6 6 RTX 3090 20 RMSprop 5e-4 0 6 RTX 3090 20 Table 1 Segmentation model training settings Graph Representation: First we pack the scene graph dataset, and try the different graph neural net to compare the accuracy of graph(image) classification. Training details on Optimizer Learning Rate Weight Decay Batch Size Hardware Epochs SGD 1e-4 5e-5 256 RTX 3090 150 Adam 1e-4 5e-5 256 RTX 3090 150 Table 2 GNN model training settings 4.3 Result The Segmentation Net: Table.3 show the different backbone segmentation model metrics on the ADE20K dataset. Backbone Validation mean iou Overall pixel accuracy Mean pixel accuracy Parameters Swin-L 48% 89% 74% 197M Swin-B 45% 82% 69% 88M Convnext-L 51% 92% 81% 198M Convnext-B 48% 90% 77% 89M Table 3 Segmentation model results Graph Representation: Table.4 presents the metrics for various GNN models applied to our graph dataset. The results from the graph representation show a significant improvement in scene understanding. This confirms the essential role of graph modality in enabling the network to recognize scenes. Additionally, the object neighbor feature enhances the performance of two stream networks and facilitates data fusion. 13 Network Train accuracy Test accuracy GCNConv 63% 48% SAGE conv graph 48% 46% Graph Attention 43% 43% Table 4 Graph Representation model results Two stream network: Our two-stream network task focuses on image classification. As shown in Table.5, there is a notable improvement when compared to conventional image classification methods. Our experiments demonstrate that incorporating scene graphs and graph representations can enhance these common methods. From another perspective, the two-stream network architecture effectively fuses two different modalities or features through independent training. Network Accuracy Parameters Swin-L 69.3% 197M Vit-L 66.8% 304M Our Two Stream Network(cross-attention fusion) 71.2% 278M Our Two Stream Network(concatenation fusion) 70.2% 278M Table 5 Two stream network results 5 Conclusion In this paper, we introduce a novel two-stream network capable of generating a scene graph from a segmentation map and classifying images based on a fused graph representation. Our approach involves training a transformer-based scene segmentation model to categorize pixel objects. Concurrently, we develop an automated algorithm to determine neighborhood relationships and create the corresponding scene graph. We employ various Graph Neural Networks (GNN) for graph classification, aimed at deriving graph representation features. The second stream of our network is a standard image classification model, which incorporates transformer encoding and a classifier head. The key of our approach lies in merging features extracted from both streams: the graph features and the image features. We experimented with several data fusion methods, including cross-attention, concatenation, sum, average, product, and voting. Our experimental results, particularly on the ADE20K dataset, demonstrate significant improvements with our two-stream network. This indicates that our scene graph representation and integration methods can revolutionize traditional image classification approaches. Furthermore, in considering the graph feature as an intermediary form of information, we open new possibilities for image encoders. Looking ahead, we plan to extend our two-stream network to other applications and datasets, such as object detection and classification of object relationships. 14"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1706.07365v2",
+                    "title": "Pixels to Graphs by Associative Embedding",
+                    "abstract": "Graphs are a useful abstraction of image content. Not only can graphs\nrepresent details about individual objects in a scene but they can capture the\ninteractions between pairs of objects. We present a method for training a\nconvolutional neural network such that it takes in an input image and produces\na full graph definition. This is done end-to-end in a single stage with the use\nof associative embeddings. The network learns to simultaneously identify all of\nthe elements that make up a graph and piece them together. We benchmark on the\nVisual Genome dataset, and demonstrate state-of-the-art performance on the\nchallenging task of scene graph generation.",
+                    "authors": "Alejandro Newell, Jia Deng",
+                    "published": "2017-06-22",
+                    "updated": "2018-03-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2010.11929v2",
+                    "title": "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale",
+                    "abstract": "While the Transformer architecture has become the de-facto standard for\nnatural language processing tasks, its applications to computer vision remain\nlimited. In vision, attention is either applied in conjunction with\nconvolutional networks, or used to replace certain components of convolutional\nnetworks while keeping their overall structure in place. We show that this\nreliance on CNNs is not necessary and a pure transformer applied directly to\nsequences of image patches can perform very well on image classification tasks.\nWhen pre-trained on large amounts of data and transferred to multiple mid-sized\nor small image recognition benchmarks (ImageNet, CIFAR-100, VTAB, etc.), Vision\nTransformer (ViT) attains excellent results compared to state-of-the-art\nconvolutional networks while requiring substantially fewer computational\nresources to train.",
+                    "authors": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby",
+                    "published": "2020-10-22",
+                    "updated": "2021-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1704.01194v1",
+                    "title": "Two Stream LSTM: A Deep Fusion Framework for Human Action Recognition",
+                    "abstract": "In this paper we address the problem of human action recognition from video\nsequences. Inspired by the exemplary results obtained via automatic feature\nlearning and deep learning approaches in computer vision, we focus our\nattention towards learning salient spatial features via a convolutional neural\nnetwork (CNN) and then map their temporal relationship with the aid of\nLong-Short-Term-Memory (LSTM) networks. Our contribution in this paper is a\ndeep fusion framework that more effectively exploits spatial features from CNNs\nwith temporal features from LSTM models. We also extensively evaluate their\nstrengths and weaknesses. We find that by combining both the sets of features,\nthe fully connected features effectively act as an attention mechanism to\ndirect the LSTM to interesting parts of the convolutional feature sequence. The\nsignificance of our fusion method is its simplicity and effectiveness compared\nto other state-of-the-art methods. The evaluation results demonstrate that this\nhierarchical multi stream fusion method has higher performance compared to\nsingle stream mapping methods allowing it to achieve high accuracy\noutperforming current state-of-the-art methods in three widely used databases:\nUCF11, UCFSports, jHMDB.",
+                    "authors": "Harshala Gammulle, Simon Denman, Sridha Sridharan, Clinton Fookes",
+                    "published": "2017-04-04",
+                    "updated": "2017-04-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2103.14030v2",
+                    "title": "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows",
+                    "abstract": "This paper presents a new vision Transformer, called Swin Transformer, that\ncapably serves as a general-purpose backbone for computer vision. Challenges in\nadapting Transformer from language to vision arise from differences between the\ntwo domains, such as large variations in the scale of visual entities and the\nhigh resolution of pixels in images compared to words in text. To address these\ndifferences, we propose a hierarchical Transformer whose representation is\ncomputed with \\textbf{S}hifted \\textbf{win}dows. The shifted windowing scheme\nbrings greater efficiency by limiting self-attention computation to\nnon-overlapping local windows while also allowing for cross-window connection.\nThis hierarchical architecture has the flexibility to model at various scales\nand has linear computational complexity with respect to image size. These\nqualities of Swin Transformer make it compatible with a broad range of vision\ntasks, including image classification (87.3 top-1 accuracy on ImageNet-1K) and\ndense prediction tasks such as object detection (58.7 box AP and 51.1 mask AP\non COCO test-dev) and semantic segmentation (53.5 mIoU on ADE20K val). Its\nperformance surpasses the previous state-of-the-art by a large margin of +2.7\nbox AP and +2.6 mask AP on COCO, and +3.2 mIoU on ADE20K, demonstrating the\npotential of Transformer-based models as vision backbones. The hierarchical\ndesign and the shifted window approach also prove beneficial for all-MLP\narchitectures. The code and models are publicly available\nat~\\url{https://github.com/microsoft/Swin-Transformer}.",
+                    "authors": "Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo",
+                    "published": "2021-03-25",
+                    "updated": "2021-08-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1706.03762v7",
+                    "title": "Attention Is All You Need",
+                    "abstract": "The dominant sequence transduction models are based on complex recurrent or\nconvolutional neural networks in an encoder-decoder configuration. The best\nperforming models also connect the encoder and decoder through an attention\nmechanism. We propose a new simple network architecture, the Transformer, based\nsolely on attention mechanisms, dispensing with recurrence and convolutions\nentirely. Experiments on two machine translation tasks show these models to be\nsuperior in quality while being more parallelizable and requiring significantly\nless time to train. Our model achieves 28.4 BLEU on the WMT 2014\nEnglish-to-German translation task, improving over the existing best results,\nincluding ensembles by over 2 BLEU. On the WMT 2014 English-to-French\ntranslation task, our model establishes a new single-model state-of-the-art\nBLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction\nof the training costs of the best models from the literature. We show that the\nTransformer generalizes well to other tasks by applying it successfully to\nEnglish constituency parsing both with large and limited training data.",
+                    "authors": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin",
+                    "published": "2017-06-12",
+                    "updated": "2023-08-02",
+                    "primary_cat": "cs.CL",
+                    "cats": [
+                        "cs.CL",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1609.02907v4",
+                    "title": "Semi-Supervised Classification with Graph Convolutional Networks",
+                    "abstract": "We present a scalable approach for semi-supervised learning on\ngraph-structured data that is based on an efficient variant of convolutional\nneural networks which operate directly on graphs. We motivate the choice of our\nconvolutional architecture via a localized first-order approximation of\nspectral graph convolutions. Our model scales linearly in the number of graph\nedges and learns hidden layer representations that encode both local graph\nstructure and features of nodes. In a number of experiments on citation\nnetworks and on a knowledge graph dataset we demonstrate that our approach\noutperforms related methods by a significant margin.",
+                    "authors": "Thomas N. Kipf, Max Welling",
+                    "published": "2016-09-09",
+                    "updated": "2017-02-22",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2002.05709v3",
+                    "title": "A Simple Framework for Contrastive Learning of Visual Representations",
+                    "abstract": "This paper presents SimCLR: a simple framework for contrastive learning of\nvisual representations. We simplify recently proposed contrastive\nself-supervised learning algorithms without requiring specialized architectures\nor a memory bank. In order to understand what enables the contrastive\nprediction tasks to learn useful representations, we systematically study the\nmajor components of our framework. We show that (1) composition of data\naugmentations plays a critical role in defining effective predictive tasks, (2)\nintroducing a learnable nonlinear transformation between the representation and\nthe contrastive loss substantially improves the quality of the learned\nrepresentations, and (3) contrastive learning benefits from larger batch sizes\nand more training steps compared to supervised learning. By combining these\nfindings, we are able to considerably outperform previous methods for\nself-supervised and semi-supervised learning on ImageNet. A linear classifier\ntrained on self-supervised representations learned by SimCLR achieves 76.5%\ntop-1 accuracy, which is a 7% relative improvement over previous\nstate-of-the-art, matching the performance of a supervised ResNet-50. When\nfine-tuned on only 1% of the labels, we achieve 85.8% top-5 accuracy,\noutperforming AlexNet with 100X fewer labels.",
+                    "authors": "Ting Chen, Simon Kornblith, Mohammad Norouzi, Geoffrey Hinton",
+                    "published": "2020-02-13",
+                    "updated": "2020-07-01",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.CV",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1710.10903v3",
+                    "title": "Graph Attention Networks",
+                    "abstract": "We present graph attention networks (GATs), novel neural network\narchitectures that operate on graph-structured data, leveraging masked\nself-attentional layers to address the shortcomings of prior methods based on\ngraph convolutions or their approximations. By stacking layers in which nodes\nare able to attend over their neighborhoods' features, we enable (implicitly)\nspecifying different weights to different nodes in a neighborhood, without\nrequiring any kind of costly matrix operation (such as inversion) or depending\non knowing the graph structure upfront. In this way, we address several key\nchallenges of spectral-based graph neural networks simultaneously, and make our\nmodel readily applicable to inductive as well as transductive problems. Our GAT\nmodels have achieved or matched state-of-the-art results across four\nestablished transductive and inductive graph benchmarks: the Cora, Citeseer and\nPubmed citation network datasets, as well as a protein-protein interaction\ndataset (wherein test graphs remain unseen during training).",
+                    "authors": "Petar Veli\u010dkovi\u0107, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li\u00f2, Yoshua Bengio",
+                    "published": "2017-10-30",
+                    "updated": "2018-02-04",
+                    "primary_cat": "stat.ML",
+                    "cats": [
+                        "stat.ML",
+                        "cs.AI",
+                        "cs.LG",
+                        "cs.SI"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.03748v2",
+                    "title": "Representation Learning with Contrastive Predictive Coding",
+                    "abstract": "While supervised learning has enabled great progress in many applications,\nunsupervised learning has not seen such widespread adoption, and remains an\nimportant and challenging endeavor for artificial intelligence. In this work,\nwe propose a universal unsupervised learning approach to extract useful\nrepresentations from high-dimensional data, which we call Contrastive\nPredictive Coding. The key insight of our model is to learn such\nrepresentations by predicting the future in latent space by using powerful\nautoregressive models. We use a probabilistic contrastive loss which induces\nthe latent space to capture information that is maximally useful to predict\nfuture samples. It also makes the model tractable by using negative sampling.\nWhile most prior work has focused on evaluating representations for a\nparticular modality, we demonstrate that our approach is able to learn useful\nrepresentations achieving strong performance on four distinct domains: speech,\nimages, text and reinforcement learning in 3D environments.",
+                    "authors": "Aaron van den Oord, Yazhe Li, Oriol Vinyals",
+                    "published": "2018-07-10",
+                    "updated": "2019-01-22",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1406.2199v2",
+                    "title": "Two-Stream Convolutional Networks for Action Recognition in Videos",
+                    "abstract": "We investigate architectures of discriminatively trained deep Convolutional\nNetworks (ConvNets) for action recognition in video. The challenge is to\ncapture the complementary information on appearance from still frames and\nmotion between frames. We also aim to generalise the best performing\nhand-crafted features within a data-driven learning framework.\n  Our contribution is three-fold. First, we propose a two-stream ConvNet\narchitecture which incorporates spatial and temporal networks. Second, we\ndemonstrate that a ConvNet trained on multi-frame dense optical flow is able to\nachieve very good performance in spite of limited training data. Finally, we\nshow that multi-task learning, applied to two different action classification\ndatasets, can be used to increase the amount of training data and improve the\nperformance on both.\n  Our architecture is trained and evaluated on the standard video actions\nbenchmarks of UCF-101 and HMDB-51, where it is competitive with the state of\nthe art. It also exceeds by a large margin previous attempts to use deep nets\nfor video classification.",
+                    "authors": "Karen Simonyan, Andrew Zisserman",
+                    "published": "2014-06-09",
+                    "updated": "2014-11-12",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1406.1078v3",
+                    "title": "Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation",
+                    "abstract": "In this paper, we propose a novel neural network model called RNN\nEncoder-Decoder that consists of two recurrent neural networks (RNN). One RNN\nencodes a sequence of symbols into a fixed-length vector representation, and\nthe other decodes the representation into another sequence of symbols. The\nencoder and decoder of the proposed model are jointly trained to maximize the\nconditional probability of a target sequence given a source sequence. The\nperformance of a statistical machine translation system is empirically found to\nimprove by using the conditional probabilities of phrase pairs computed by the\nRNN Encoder-Decoder as an additional feature in the existing log-linear model.\nQualitatively, we show that the proposed model learns a semantically and\nsyntactically meaningful representation of linguistic phrases.",
+                    "authors": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, Yoshua Bengio",
+                    "published": "2014-06-03",
+                    "updated": "2014-09-03",
+                    "primary_cat": "cs.CL",
+                    "cats": [
+                        "cs.CL",
+                        "cs.LG",
+                        "cs.NE",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1701.02426v2",
+                    "title": "Scene Graph Generation by Iterative Message Passing",
+                    "abstract": "Understanding a visual scene goes beyond recognizing individual objects in\nisolation. Relationships between objects also constitute rich semantic\ninformation about the scene. In this work, we explicitly model the objects and\ntheir relationships using scene graphs, a visually-grounded graphical structure\nof an image. We propose a novel end-to-end model that generates such structured\nscene representation from an input image. The model solves the scene graph\ninference problem using standard RNNs and learns to iteratively improves its\npredictions via message passing. Our joint inference model can take advantage\nof contextual cues to make better predictions on objects and their\nrelationships. The experiments show that our model significantly outperforms\nprevious methods for generating scene graphs using Visual Genome dataset and\ninferring support relations with NYU Depth v2 dataset.",
+                    "authors": "Danfei Xu, Yuke Zhu, Christopher B. Choy, Li Fei-Fei",
+                    "published": "2017-01-10",
+                    "updated": "2017-04-12",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1706.02216v4",
+                    "title": "Inductive Representation Learning on Large Graphs",
+                    "abstract": "Low-dimensional embeddings of nodes in large graphs have proved extremely\nuseful in a variety of prediction tasks, from content recommendation to\nidentifying protein functions. However, most existing approaches require that\nall nodes in the graph are present during training of the embeddings; these\nprevious approaches are inherently transductive and do not naturally generalize\nto unseen nodes. Here we present GraphSAGE, a general, inductive framework that\nleverages node feature information (e.g., text attributes) to efficiently\ngenerate node embeddings for previously unseen data. Instead of training\nindividual embeddings for each node, we learn a function that generates\nembeddings by sampling and aggregating features from a node's local\nneighborhood. Our algorithm outperforms strong baselines on three inductive\nnode-classification benchmarks: we classify the category of unseen nodes in\nevolving information graphs based on citation and Reddit post data, and we show\nthat our algorithm generalizes to completely unseen graphs using a multi-graph\ndataset of protein-protein interactions.",
+                    "authors": "William L. Hamilton, Rex Ying, Jure Leskovec",
+                    "published": "2017-06-07",
+                    "updated": "2018-09-10",
+                    "primary_cat": "cs.SI",
+                    "cats": [
+                        "cs.SI",
+                        "cs.LG",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2103.00020v1",
+                    "title": "Learning Transferable Visual Models From Natural Language Supervision",
+                    "abstract": "State-of-the-art computer vision systems are trained to predict a fixed set\nof predetermined object categories. This restricted form of supervision limits\ntheir generality and usability since additional labeled data is needed to\nspecify any other visual concept. Learning directly from raw text about images\nis a promising alternative which leverages a much broader source of\nsupervision. We demonstrate that the simple pre-training task of predicting\nwhich caption goes with which image is an efficient and scalable way to learn\nSOTA image representations from scratch on a dataset of 400 million (image,\ntext) pairs collected from the internet. After pre-training, natural language\nis used to reference learned visual concepts (or describe new ones) enabling\nzero-shot transfer of the model to downstream tasks. We study the performance\nof this approach by benchmarking on over 30 different existing computer vision\ndatasets, spanning tasks such as OCR, action recognition in videos,\ngeo-localization, and many types of fine-grained object classification. The\nmodel transfers non-trivially to most tasks and is often competitive with a\nfully supervised baseline without the need for any dataset specific training.\nFor instance, we match the accuracy of the original ResNet-50 on ImageNet\nzero-shot without needing to use any of the 1.28 million training examples it\nwas trained on. We release our code and pre-trained model weights at\nhttps://github.com/OpenAI/CLIP.",
+                    "authors": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever",
+                    "published": "2021-02-26",
+                    "updated": "2021-02-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.17091v1",
+                    "title": "SSSegmenation: An Open Source Supervised Semantic Segmentation Toolbox Based on PyTorch",
+                    "abstract": "This paper presents SSSegmenation, which is an open source supervised\nsemantic image segmentation toolbox based on PyTorch. The design of this\ntoolbox is motivated by MMSegmentation while it is easier to use because of\nfewer dependencies and achieves superior segmentation performance under a\ncomparable training and testing setup. Moreover, the toolbox also provides\nplenty of trained weights for popular and contemporary semantic segmentation\nmethods, including Deeplab, PSPNet, OCRNet, MaskFormer, \\emph{etc}. We expect\nthat this toolbox can contribute to the future development of semantic\nsegmentation. Codes and model zoos are available at\n\\href{https://github.com/SegmentationBLWX/sssegmentation/}{SSSegmenation}.",
+                    "authors": "Zhenchao Jin",
+                    "published": "2023-05-26",
+                    "updated": "2023-05-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2401.11739v1",
+                    "title": "EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models",
+                    "abstract": "Diffusion models have recently received increasing research attention for\ntheir remarkable transfer abilities in semantic segmentation tasks. However,\ngenerating fine-grained segmentation masks with diffusion models often requires\nadditional training on annotated datasets, leaving it unclear to what extent\npre-trained diffusion models alone understand the semantic relations of their\ngenerated images. To address this question, we leverage the semantic knowledge\nextracted from Stable Diffusion (SD) and aim to develop an image segmentor\ncapable of generating fine-grained segmentation maps without any additional\ntraining. The primary difficulty stems from the fact that semantically\nmeaningful feature maps typically exist only in the spatially lower-dimensional\nlayers, which poses a challenge in directly extracting pixel-level semantic\nrelations from these feature maps. To overcome this issue, our framework\nidentifies semantic correspondences between image pixels and spatial locations\nof low-dimensional feature maps by exploiting SD's generation process and\nutilizes them for constructing image-resolution segmentation maps. In extensive\nexperiments, the produced segmentation maps are demonstrated to be well\ndelineated and capture detailed parts of the images, indicating the existence\nof highly accurate pixel-level semantic knowledge in diffusion models.",
+                    "authors": "Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, Seung Wook Kim",
+                    "published": "2024-01-22",
+                    "updated": "2024-01-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.00242v1",
+                    "title": "3D Guided Weakly Supervised Semantic Segmentation",
+                    "abstract": "Pixel-wise clean annotation is necessary for fully-supervised semantic\nsegmentation, which is laborious and expensive to obtain. In this paper, we\npropose a weakly supervised 2D semantic segmentation model by incorporating\nsparse bounding box labels with available 3D information, which is much easier\nto obtain with advanced sensors. We manually labeled a subset of the 2D-3D\nSemantics(2D-3D-S) dataset with bounding boxes, and introduce our 2D-3D\ninference module to generate accurate pixel-wise segment proposal masks. Guided\nby 3D information, we first generate a point cloud of objects and calculate\nobjectness probability score for each point. Then we project the point cloud\nwith objectness probabilities back to 2D images followed by a refinement step\nto obtain segment proposals, which are treated as pseudo labels to train a\nsemantic segmentation network. Our method works in a recursive manner to\ngradually refine the above-mentioned segment proposals. Extensive experimental\nresults on the 2D-3D-S dataset show that the proposed method can generate\naccurate segment proposals when bounding box labels are available on only a\nsmall subset of training images. Performance comparison with recent\nstate-of-the-art methods further illustrates the effectiveness of our method.",
+                    "authors": "Weixuan Sun, Jing Zhang, Nick Barnes",
+                    "published": "2020-12-01",
+                    "updated": "2020-12-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.02094v1",
+                    "title": "Segment Anything Meets Semantic Communication",
+                    "abstract": "In light of the diminishing returns of traditional methods for enhancing\ntransmission rates, the domain of semantic communication presents promising new\nfrontiers. Focusing on image transmission, this paper explores the application\nof foundation models, particularly the Segment Anything Model (SAM) developed\nby Meta AI Research, to improve semantic communication. SAM is a promptable\nimage segmentation model that has gained attention for its ability to perform\nzero-shot segmentation tasks without explicit training or domain-specific\nknowledge. By employing SAM's segmentation capability and lightweight neural\nnetwork architecture for semantic coding, we propose a practical approach to\nsemantic communication. We demonstrate that this approach retains critical\nsemantic features, achieving higher image reconstruction quality and reducing\ncommunication overhead. This practical solution eliminates the\nresource-intensive stage of training a segmentation model and can be applied to\nany semantic coding architecture, paving the way for real-world applications.",
+                    "authors": "Shehbaz Tariq, Brian Estadimas Arfeto, Chaoning Zhang, Hyundong Shin",
+                    "published": "2023-06-03",
+                    "updated": "2023-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.04297v1",
+                    "title": "SHDM-NET: Heat Map Detail Guidance with Image Matting for Industrial Weld Semantic Segmentation Network",
+                    "abstract": "In actual industrial production, the assessment of the steel plate welding\neffect is an important task, and the segmentation of the weld section is the\nbasis of the assessment. This paper proposes an industrial weld segmentation\nnetwork based on a deep learning semantic segmentation algorithm fused with\nheatmap detail guidance and Image Matting to solve the automatic segmentation\nproblem of weld regions. In the existing semantic segmentation networks, the\nboundary information can be preserved by fusing the features of both high-level\nand low-level layers. However, this method can lead to insufficient expression\nof the spatial information in the low-level layer, resulting in inaccurate\nsegmentation boundary positioning. We propose a detailed guidance module based\non heatmaps to fully express the segmented region boundary information in the\nlow-level network to address this problem. Specifically, the expression of\nboundary information can be enhanced by adding a detailed branch to predict\nsegmented boundary and then matching it with the boundary heat map generated by\nmask labels to calculate the mean square error loss. In addition, although deep\nlearning has achieved great success in the field of semantic segmentation, the\nprecision of the segmentation boundary region is not high due to the loss of\ndetailed information caused by the classical segmentation network in the\nprocess of encoding and decoding process. This paper introduces a matting\nalgorithm to calibrate the boundary of the segmentation region of the semantic\nsegmentation network to solve this problem. Through many experiments on\nindustrial weld data sets, the effectiveness of our method is demonstrated, and\nthe MIOU reaches 97.93%. It is worth noting that this performance is comparable\nto human manual segmentation ( MIOU 97.96%).",
+                    "authors": "Qi Wang, Jingwu Mei",
+                    "published": "2022-07-09",
+                    "updated": "2022-07-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2111.02444v2",
+                    "title": "Panoptic 3D Scene Reconstruction From a Single RGB Image",
+                    "abstract": "Understanding 3D scenes from a single image is fundamental to a wide variety\nof tasks, such as for robotics, motion planning, or augmented reality. Existing\nworks in 3D perception from a single RGB image tend to focus on geometric\nreconstruction only, or geometric reconstruction with semantic segmentation or\ninstance segmentation. Inspired by 2D panoptic segmentation, we propose to\nunify the tasks of geometric reconstruction, 3D semantic segmentation, and 3D\ninstance segmentation into the task of panoptic 3D scene reconstruction - from\na single RGB image, predicting the complete geometric reconstruction of the\nscene in the camera frustum of the image, along with semantic and instance\nsegmentations. We thus propose a new approach for holistic 3D scene\nunderstanding from a single RGB image which learns to lift and propagate 2D\nfeatures from an input image to a 3D volumetric scene representation. We\ndemonstrate that this holistic view of joint scene reconstruction, semantic,\nand instance segmentation is beneficial over treating the tasks independently,\nthus outperforming alternative approaches.",
+                    "authors": "Manuel Dahnert, Ji Hou, Matthias Nie\u00dfner, Angela Dai",
+                    "published": "2021-11-03",
+                    "updated": "2022-05-16",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.03273v1",
+                    "title": "Semantic Segmentation using Vision Transformers: A survey",
+                    "abstract": "Semantic segmentation has a broad range of applications in a variety of\ndomains including land coverage analysis, autonomous driving, and medical image\nanalysis. Convolutional neural networks (CNN) and Vision Transformers (ViTs)\nprovide the architecture models for semantic segmentation. Even though ViTs\nhave proven success in image classification, they cannot be directly applied to\ndense prediction tasks such as image segmentation and object detection since\nViT is not a general purpose backbone due to its patch partitioning scheme. In\nthis survey, we discuss some of the different ViT architectures that can be\nused for semantic segmentation and how their evolution managed the above-stated\nchallenge. The rise of ViT and its performance with a high success rate\nmotivated the community to slowly replace the traditional convolutional neural\nnetworks in various computer vision tasks. This survey aims to review and\ncompare the performances of ViT architectures designed for semantic\nsegmentation using benchmarking datasets. This will be worthwhile for the\ncommunity to yield knowledge regarding the implementations carried out in\nsemantic segmentation and to discover more efficient methodologies using ViTs.",
+                    "authors": "Hans Thisanke, Chamli Deshan, Kavindu Chamith, Sachith Seneviratne, Rajith Vidanaarachchi, Damayanthi Herath",
+                    "published": "2023-05-05",
+                    "updated": "2023-05-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.03212v2",
+                    "title": "Hierarchical Semantic Segmentation using Psychometric Learning",
+                    "abstract": "Assigning meaning to parts of image data is the goal of semantic image\nsegmentation. Machine learning methods, specifically supervised learning is\ncommonly used in a variety of tasks formulated as semantic segmentation. One of\nthe major challenges in the supervised learning approaches is expressing and\ncollecting the rich knowledge that experts have with respect to the meaning\npresent in the image data. Towards this, typically a fixed set of labels is\nspecified and experts are tasked with annotating the pixels, patches or\nsegments in the images with the given labels. In general, however, the set of\nclasses does not fully capture the rich semantic information present in the\nimages. For example, in medical imaging such as histology images, the different\nparts of cells could be grouped and sub-grouped based on the expertise of the\npathologist.\n  To achieve such a precise semantic representation of the concepts in the\nimage, we need access to the full depth of knowledge of the annotator. In this\nwork, we develop a novel approach to collect segmentation annotations from\nexperts based on psychometric testing. Our method consists of the psychometric\ntesting procedure, active query selection, query enhancement, and a deep metric\nlearning model to achieve a patch-level image embedding that allows for\nsemantic segmentation of images. We show the merits of our method with\nevaluation on the synthetically generated image, aerial image and histology\nimage.",
+                    "authors": "Lu Yin, Vlado Menkovski, Shiwei Liu, Mykola Pechenizkiy",
+                    "published": "2021-07-07",
+                    "updated": "2021-12-16",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2312.17243v1",
+                    "title": "Unsupervised Universal Image Segmentation",
+                    "abstract": "Several unsupervised image segmentation approaches have been proposed which\neliminate the need for dense manually-annotated segmentation masks; current\nmodels separately handle either semantic segmentation (e.g., STEGO) or\nclass-agnostic instance segmentation (e.g., CutLER), but not both (i.e.,\npanoptic segmentation). We propose an Unsupervised Universal Segmentation model\n(U2Seg) adept at performing various image segmentation tasks -- instance,\nsemantic and panoptic -- using a novel unified framework. U2Seg generates\npseudo semantic labels for these segmentation tasks via leveraging\nself-supervised models followed by clustering; each cluster represents\ndifferent semantic and/or instance membership of pixels. We then self-train the\nmodel on these pseudo semantic labels, yielding substantial performance gains\nover specialized methods tailored to each task: a +2.6 AP$^{\\text{box}}$ boost\nvs. CutLER in unsupervised instance segmentation on COCO and a +7.0 PixelAcc\nincrease (vs. STEGO) in unsupervised semantic segmentation on COCOStuff.\nMoreover, our method sets up a new baseline for unsupervised panoptic\nsegmentation, which has not been previously explored. U2Seg is also a strong\npretrained model for few-shot segmentation, surpassing CutLER by +5.0\nAP$^{\\text{mask}}$ when trained on a low-data regime, e.g., only 1% COCO\nlabels. We hope our simple yet effective method can inspire more research on\nunsupervised universal image segmentation.",
+                    "authors": "Dantong Niu, Xudong Wang, Xinyang Han, Long Lian, Roei Herzig, Trevor Darrell",
+                    "published": "2023-12-28",
+                    "updated": "2023-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.05457v1",
+                    "title": "Instance Segmentation based Semantic Matting for Compositing Applications",
+                    "abstract": "Image compositing is a key step in film making and image editing that aims to\nsegment a foreground object and combine it with a new background. Automatic\nimage compositing can be done easily in a studio using chroma-keying when the\nbackground is pure blue or green. However, image compositing in natural scenes\nwith complex backgrounds remains a tedious task, requiring experienced artists\nto hand-segment. In order to achieve automatic compositing in natural scenes,\nwe propose a fully automated method that integrates instance segmentation and\nimage matting processes to generate high-quality semantic mattes that can be\nused for image editing task. Our approach can be seen both as a refinement of\nexisting instance segmentation algorithms and as a fully automated semantic\nimage matting method. It extends automatic image compositing techniques such as\nchroma-keying to scenes with complex natural backgrounds without the need for\nany kind of user interaction. The output of our approach can be considered as\nboth refined instance segmentations and alpha mattes with semantic meanings. We\nprovide experimental results which show improved performance results as\ncompared to existing approaches.",
+                    "authors": "Guanqing Hu, James J. Clark",
+                    "published": "2019-04-10",
+                    "updated": "2019-04-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.00487v1",
+                    "title": "Linear Semantics in Generative Adversarial Networks",
+                    "abstract": "Generative Adversarial Networks (GANs) are able to generate high-quality\nimages, but it remains difficult to explicitly specify the semantics of\nsynthesized images. In this work, we aim to better understand the semantic\nrepresentation of GANs, and thereby enable semantic control in GAN's generation\nprocess. Interestingly, we find that a well-trained GAN encodes image semantics\nin its internal feature maps in a surprisingly simple way: a linear\ntransformation of feature maps suffices to extract the generated image\nsemantics. To verify this simplicity, we conduct extensive experiments on\nvarious GANs and datasets; and thanks to this simplicity, we are able to learn\na semantic segmentation model for a trained GAN from a small number (e.g., 8)\nof labeled images. Last but not least, leveraging our findings, we propose two\nfew-shot image editing approaches, namely Semantic-Conditional Sampling and\nSemantic Image Editing. Given a trained GAN and as few as eight semantic\nannotations, the user is able to generate diverse images subject to a\nuser-provided semantic layout, and control the synthesized image semantics. We\nhave made the code publicly available.",
+                    "authors": "Jianjin Xu, Changxi Zheng",
+                    "published": "2021-04-01",
+                    "updated": "2021-04-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1611.08408v1",
+                    "title": "Semantic Segmentation using Adversarial Networks",
+                    "abstract": "Adversarial training has been shown to produce state of the art results for\ngenerative image modeling. In this paper we propose an adversarial training\napproach to train semantic segmentation models. We train a convolutional\nsemantic segmentation network along with an adversarial network that\ndiscriminates segmentation maps coming either from the ground truth or from the\nsegmentation network. The motivation for our approach is that it can detect and\ncorrect higher-order inconsistencies between ground truth segmentation maps and\nthe ones produced by the segmentation net. Our experiments show that our\nadversarial training approach leads to improved accuracy on the Stanford\nBackground and PASCAL VOC 2012 datasets.",
+                    "authors": "Pauline Luc, Camille Couprie, Soumith Chintala, Jakob Verbeek",
+                    "published": "2016-11-25",
+                    "updated": "2016-11-25",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.08465v2",
+                    "title": "DeepAtlas: Joint Semi-Supervised Learning of Image Registration and Segmentation",
+                    "abstract": "Deep convolutional neural networks (CNNs) are state-of-the-art for semantic\nimage segmentation, but typically require many labeled training samples.\nObtaining 3D segmentations of medical images for supervised training is\ndifficult and labor intensive. Motivated by classical approaches for joint\nsegmentation and registration we therefore propose a deep learning framework\nthat jointly learns networks for image registration and image segmentation. In\ncontrast to previous work on deep unsupervised image registration, which showed\nthe benefit of weak supervision via image segmentations, our approach can use\nexisting segmentations when available and computes them via the segmentation\nnetwork otherwise, thereby providing the same registration benefit. Conversely,\nsegmentation network training benefits from the registration, which essentially\nprovides a realistic form of data augmentation. Experiments on knee and brain\n3D magnetic resonance (MR) images show that our approach achieves large\nsimultaneous improvements of segmentation and registration accuracy (over\nindependently trained networks) and allows training high-quality models with\nvery limited training data. Specifically, in a one-shot-scenario (with only one\nmanually labeled image) our approach increases Dice scores (%) over an\nunsupervised registration network by 2.7 and 1.8 on the knee and brain images\nrespectively.",
+                    "authors": "Zhenlin Xu, Marc Niethammer",
+                    "published": "2019-04-17",
+                    "updated": "2019-07-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.12545v1",
+                    "title": "Unsupervised Domain Adaptation for Semantic Segmentation by Content Transfer",
+                    "abstract": "In this paper, we tackle the unsupervised domain adaptation (UDA) for\nsemantic segmentation, which aims to segment the unlabeled real data using\nlabeled synthetic data. The main problem of UDA for semantic segmentation\nrelies on reducing the domain gap between the real image and synthetic image.\nTo solve this problem, we focused on separating information in an image into\ncontent and style. Here, only the content has cues for semantic segmentation,\nand the style makes the domain gap. Thus, precise separation of content and\nstyle in an image leads to effect as supervision of real data even when\nlearning with synthetic data. To make the best of this effect, we propose a\nzero-style loss. Even though we perfectly extract content for semantic\nsegmentation in the real domain, another main challenge, the class imbalance\nproblem, still exists in UDA for semantic segmentation. We address this problem\nby transferring the contents of tail classes from synthetic to real domain.\nExperimental results show that the proposed method achieves the\nstate-of-the-art performance in semantic segmentation on the major two UDA\nsettings.",
+                    "authors": "Suhyeon Lee, Junhyuk Hyun, Hongje Seong, Euntai Kim",
+                    "published": "2020-12-23",
+                    "updated": "2020-12-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.02270v1",
+                    "title": "Weak-shot Semantic Segmentation via Dual Similarity Transfer",
+                    "abstract": "Semantic segmentation is an important and prevalent task, but severely\nsuffers from the high cost of pixel-level annotations when extending to more\nclasses in wider applications. To this end, we focus on the problem named\nweak-shot semantic segmentation, where the novel classes are learnt from\ncheaper image-level labels with the support of base classes having\noff-the-shelf pixel-level labels. To tackle this problem, we propose SimFormer,\nwhich performs dual similarity transfer upon MaskFormer. Specifically,\nMaskFormer disentangles the semantic segmentation task into two sub-tasks:\nproposal classification and proposal segmentation for each proposal. Proposal\nsegmentation allows proposal-pixel similarity transfer from base classes to\nnovel classes, which enables the mask learning of novel classes. We also learn\npixel-pixel similarity from base classes and distill such class-agnostic\nsemantic similarity to the semantic masks of novel classes, which regularizes\nthe segmentation model with pixel-level semantic relationship across images. In\naddition, we propose a complementary loss to facilitate the learning of novel\nclasses. Comprehensive experiments on the challenging COCO-Stuff-10K and ADE20K\ndatasets demonstrate the effectiveness of our method. Codes are available at\nhttps://github.com/bcmi/SimFormer-Weak-Shot-Semantic-Segmentation.",
+                    "authors": "Junjie Chen, Li Niu, Siyuan Zhou, Jianlou Si, Chen Qian, Liqing Zhang",
+                    "published": "2022-10-05",
+                    "updated": "2022-10-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.04829v2",
+                    "title": "MixReorg: Cross-Modal Mixed Patch Reorganization is a Good Mask Learner for Open-World Semantic Segmentation",
+                    "abstract": "Recently, semantic segmentation models trained with image-level text\nsupervision have shown promising results in challenging open-world scenarios.\nHowever, these models still face difficulties in learning fine-grained semantic\nalignment at the pixel level and predicting accurate object masks. To address\nthis issue, we propose MixReorg, a novel and straightforward pre-training\nparadigm for semantic segmentation that enhances a model's ability to\nreorganize patches mixed across images, exploring both local visual relevance\nand global semantic coherence. Our approach involves generating fine-grained\npatch-text pairs data by mixing image patches while preserving the\ncorrespondence between patches and text. The model is then trained to minimize\nthe segmentation loss of the mixed images and the two contrastive losses of the\noriginal and restored features. With MixReorg as a mask learner, conventional\ntext-supervised semantic segmentation models can achieve highly generalizable\npixel-semantic alignment ability, which is crucial for open-world segmentation.\nAfter training with large-scale image-text data, MixReorg models can be applied\ndirectly to segment visual objects of arbitrary categories, without the need\nfor further fine-tuning. Our proposed framework demonstrates strong performance\non popular zero-shot semantic segmentation benchmarks, outperforming GroupViT\nby significant margins of 5.0%, 6.2%, 2.5%, and 3.4% mIoU on PASCAL VOC2012,\nPASCAL Context, MS COCO, and ADE20K, respectively.",
+                    "authors": "Kaixin Cai, Pengzhen Ren, Yi Zhu, Hang Xu, Jianzhuang Liu, Changlin Li, Guangrun Wang, Xiaodan Liang",
+                    "published": "2023-08-09",
+                    "updated": "2024-03-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2203.15943v1",
+                    "title": "Self-Supervised Leaf Segmentation under Complex Lighting Conditions",
+                    "abstract": "As an essential prerequisite task in image-based plant phenotyping, leaf\nsegmentation has garnered increasing attention in recent years. While\nself-supervised learning is emerging as an effective alternative to various\ncomputer vision tasks, its adaptation for image-based plant phenotyping remains\nrather unexplored. In this work, we present a self-supervised leaf segmentation\nframework consisting of a self-supervised semantic segmentation model, a\ncolor-based leaf segmentation algorithm, and a self-supervised color correction\nmodel. The self-supervised semantic segmentation model groups the semantically\nsimilar pixels by iteratively referring to the self-contained information,\nallowing the pixels of the same semantic object to be jointly considered by the\ncolor-based leaf segmentation algorithm for identifying the leaf regions.\nAdditionally, we propose to use a self-supervised color correction model for\nimages taken under complex illumination conditions. Experimental results on\ndatasets of different plant species demonstrate the potential of the proposed\nself-supervised framework in achieving effective and generalizable leaf\nsegmentation.",
+                    "authors": "Xufeng Lin, Chang-Tsun Li, Scott Adams, Abbas Kouzani, Richard Jiang, Ligang He, Yongjian Hu, Michael Vernon, Egan Doeven, Lawrence Webb, Todd Mcclellan, Adam Guskic",
+                    "published": "2022-03-29",
+                    "updated": "2022-03-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.05840v2",
+                    "title": "Self-Correlation and Cross-Correlation Learning for Few-Shot Remote Sensing Image Semantic Segmentation",
+                    "abstract": "Remote sensing image semantic segmentation is an important problem for remote\nsensing image interpretation. Although remarkable progress has been achieved,\nexisting deep neural network methods suffer from the reliance on massive\ntraining data. Few-shot remote sensing semantic segmentation aims at learning\nto segment target objects from a query image using only a few annotated support\nimages of the target class. Most existing few-shot learning methods stem\nprimarily from their sole focus on extracting information from support images,\nthereby failing to effectively address the large variance in appearance and\nscales of geographic objects. To tackle these challenges, we propose a\nSelf-Correlation and Cross-Correlation Learning Network for the few-shot remote\nsensing image semantic segmentation. Our model enhances the generalization by\nconsidering both self-correlation and cross-correlation between support and\nquery images to make segmentation predictions. To further explore the\nself-correlation with the query image, we propose to adopt a classical spectral\nmethod to produce a class-agnostic segmentation mask based on the basic visual\ninformation of the image. Extensive experiments on two remote sensing image\ndatasets demonstrate the effectiveness and superiority of our model in few-shot\nremote sensing image semantic segmentation. Code and models will be accessed at\nhttps://github.com/linhanwang/SCCNet.",
+                    "authors": "Linhan Wang, Shuo Lei, Jianfeng He, Shengkun Wang, Min Zhang, Chang-Tien Lu",
+                    "published": "2023-09-11",
+                    "updated": "2023-09-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.13697v1",
+                    "title": "Generalizable Semantic Vision Query Generation for Zero-shot Panoptic and Semantic Segmentation",
+                    "abstract": "Zero-shot Panoptic Segmentation (ZPS) aims to recognize foreground instances\nand background stuff without images containing unseen categories in training.\nDue to the visual data sparsity and the difficulty of generalizing from seen to\nunseen categories, this task remains challenging. To better generalize to\nunseen classes, we propose Conditional tOken aligNment and Cycle trAnsiTion\n(CONCAT), to produce generalizable semantic vision queries. First, a feature\nextractor is trained by CON to link the vision and semantics for providing\ntarget queries. Formally, CON is proposed to align the semantic queries with\nthe CLIP visual CLS token extracted from complete and masked images. To address\nthe lack of unseen categories, a generator is required. However, one of the\ngaps in synthesizing pseudo vision queries, ie, vision queries for unseen\ncategories, is describing fine-grained visual details through semantic\nembeddings. Therefore, we approach CAT to train the generator in\nsemantic-vision and vision-semantic manners. In semantic-vision, visual query\ncontrast is proposed to model the high granularity of vision by pulling the\npseudo vision queries with the corresponding targets containing segments while\npushing those without segments away. To ensure the generated queries retain\nsemantic information, in vision-semantic, the pseudo vision queries are mapped\nback to semantic and supervised by real semantic embeddings. Experiments on ZPS\nachieve a 5.2% hPQ increase surpassing SOTA. We also examine inductive ZPS and\nopen-vocabulary semantic segmentation and obtain comparative results while\nbeing 2 times faster in testing.",
+                    "authors": "Jialei Chen, Daisuke Deguchi, Chenkai Zhang, Hiroshi Murase",
+                    "published": "2024-02-21",
+                    "updated": "2024-02-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2211.08352v1",
+                    "title": "Visual Semantic Segmentation Based on Few/Zero-Shot Learning: An Overview",
+                    "abstract": "Visual semantic segmentation aims at separating a visual sample into diverse\nblocks with specific semantic attributes and identifying the category for each\nblock, and it plays a crucial role in environmental perception. Conventional\nlearning-based visual semantic segmentation approaches count heavily on\nlarge-scale training data with dense annotations and consistently fail to\nestimate accurate semantic labels for unseen categories. This obstruction spurs\na craze for studying visual semantic segmentation with the assistance of\nfew/zero-shot learning. The emergence and rapid progress of few/zero-shot\nvisual semantic segmentation make it possible to learn unseen-category from a\nfew labeled or zero-labeled samples, which advances the extension to practical\napplications. Therefore, this paper focuses on the recently published\nfew/zero-shot visual semantic segmentation methods varying from 2D to 3D space\nand explores the commonalities and discrepancies of technical settlements under\ndifferent segmentation circumstances. Specifically, the preliminaries on\nfew/zero-shot visual semantic segmentation, including the problem definitions,\ntypical datasets, and technical remedies, are briefly reviewed and discussed.\nMoreover, three typical instantiations are involved to uncover the interactions\nof few/zero-shot learning with visual semantic segmentation, including image\nsemantic segmentation, video object segmentation, and 3D segmentation. Finally,\nthe future challenges of few/zero-shot visual semantic segmentation are\ndiscussed.",
+                    "authors": "Wenqi Ren, Yang Tang, Qiyu Sun, Chaoqiang Zhao, Qing-Long Han",
+                    "published": "2022-11-13",
+                    "updated": "2022-11-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2011.00674v1",
+                    "title": "Highway Driving Dataset for Semantic Video Segmentation",
+                    "abstract": "Scene understanding is an essential technique in semantic segmentation.\nAlthough there exist several datasets that can be used for semantic\nsegmentation, they are mainly focused on semantic image segmentation with large\ndeep neural networks. Therefore, these networks are not useful for real time\napplications, especially in autonomous driving systems. In order to solve this\nproblem, we make two contributions to semantic segmentation task. The first\ncontribution is that we introduce the semantic video dataset, the Highway\nDriving dataset, which is a densely annotated benchmark for a semantic video\nsegmentation task. The Highway Driving dataset consists of 20 video sequences\nhaving a 30Hz frame rate, and every frame is densely annotated. Secondly, we\npropose a baseline algorithm that utilizes a temporal correlation. Together\nwith our attempt to analyze the temporal correlation, we expect the Highway\nDriving dataset to encourage research on semantic video segmentation.",
+                    "authors": "Byungju Kim, Junho Yim, Junmo Kim",
+                    "published": "2020-11-02",
+                    "updated": "2020-11-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1804.04882v2",
+                    "title": "Learning to Exploit the Prior Network Knowledge for Weakly-Supervised Semantic Segmentation",
+                    "abstract": "Training a Convolutional Neural Network (CNN) for semantic segmentation\ntypically requires to collect a large amount of accurate pixel-level\nannotations, a hard and expensive task. In contrast, simple image tags are\neasier to gather. With this paper we introduce a novel weakly-supervised\nsemantic segmentation model able to learn from image labels, and just image\nlabels. Our model uses the prior knowledge of a network trained for image\nrecognition, employing these image annotations as an attention mechanism to\nidentify semantic regions in the images. We then present a methodology that\nbuilds accurate class-specific segmentation masks from these regions, where\nneither external objectness nor saliency algorithms are required. We describe\nhow to incorporate this mask generation strategy into a fully end-to-end\ntrainable process where the network jointly learns to classify and segment\nimages. Our experiments on PASCAL VOC 2012 dataset show that exploiting these\ngenerated class-specific masks in conjunction with our novel end-to-end\nlearning process outperforms several recent weakly-supervised semantic\nsegmentation methods that use image tags only, and even some models that\nleverage additional supervision or training data.",
+                    "authors": "Carolina Redondo-Cabrera, Marcos Baptista-R\u00edos, Roberto J. L\u00f3pez-Sastre",
+                    "published": "2018-04-13",
+                    "updated": "2019-02-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2405.04913v1",
+                    "title": "Weakly-supervised Semantic Segmentation via Dual-stream Contrastive Learning of Cross-image Contextual Information",
+                    "abstract": "Weakly supervised semantic segmentation (WSSS) aims at learning a semantic\nsegmentation model with only image-level tags. Despite intensive research on\ndeep learning approaches over a decade, there is still a significant\nperformance gap between WSSS and full semantic segmentation. Most current WSSS\nmethods always focus on a limited single image (pixel-wise) information while\nignoring the valuable inter-image (semantic-wise) information. From this\nperspective, a novel end-to-end WSSS framework called DSCNet is developed along\nwith two innovations: i) pixel-wise group contrast and semantic-wise graph\ncontrast are proposed and introduced into the WSSS framework; ii) a novel\ndual-stream contrastive learning (DSCL) mechanism is designed to jointly handle\npixel-wise and semantic-wise context information for better WSSS performance.\nSpecifically, the pixel-wise group contrast learning (PGCL) and semantic-wise\ngraph contrast learning (SGCL) tasks form a more comprehensive solution.\nExtensive experiments on PASCAL VOC and MS COCO benchmarks verify the\nsuperiority of DSCNet over SOTA approaches and baseline models.",
+                    "authors": "Qi Lai, Chi-Man Vong",
+                    "published": "2024-05-08",
+                    "updated": "2024-05-08",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2101.08418v2",
+                    "title": "Rethinking Semantic Segmentation Evaluation for Explainability and Model Selection",
+                    "abstract": "Semantic segmentation aims to robustly predict coherent class labels for\nentire regions of an image. It is a scene understanding task that powers\nreal-world applications (e.g., autonomous navigation). One important\napplication, the use of imagery for automated semantic understanding of\npedestrian environments, provides remote mapping of accessibility features in\nstreet environments. This application (and others like it) require detailed\ngeometric information of geographical objects. Semantic segmentation is a\nprerequisite for this task since it maps contiguous regions of the same class\nas single entities. Importantly, semantic segmentation uses like ours are not\npixel-wise outcomes; however, most of their quantitative evaluation metrics\n(e.g., mean Intersection Over Union) are based on pixel-wise similarities to a\nground-truth, which fails to emphasize over- and under-segmentation properties\nof a segmentation model. Here, we introduce a new metric to assess region-based\nover- and under-segmentation. We analyze and compare it to other metrics,\ndemonstrating that the use of our metric lends greater explainability to\nsemantic segmentation model performance in real-world applications.",
+                    "authors": "Yuxiang Zhang, Sachin Mehta, Anat Caspi",
+                    "published": "2021-01-21",
+                    "updated": "2023-02-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.13118v1",
+                    "title": "Automatic Preprocessing and Ensemble Learning for Low Quality Cell Image Segmentation",
+                    "abstract": "We propose an automatic preprocessing and ensemble learning for segmentation\nof cell images with low quality. It is difficult to capture cells with strong\nlight. Therefore, the microscopic images of cells tend to have low image\nquality but these images are not good for semantic segmentation. Here we\npropose a method to translate an input image to the images that are easy to\nrecognize by deep learning. The proposed method consists of two deep neural\nnetworks. The first network is the usual training for semantic segmentation,\nand penultimate feature maps of the first network are used as filters to\ntranslate an input image to the images that emphasize each class. This is the\nautomatic preprocessing and translated cell images are easily classified. The\ninput cell image with low quality is translated by the feature maps in the\nfirst network, and the translated images are fed into the second network for\nsemantic segmentation. Since the outputs of the second network are multiple\nsegmentation results, we conduct the weighted ensemble of those segmentation\nimages. Two networks are trained by end-to-end manner, and we do not need to\nprepare images with high quality for the translation. We confirmed that our\nproposed method can translate cell images with low quality to the images that\nare easy to segment, and segmentation accuracy has improved using the weighted\nensemble learning.",
+                    "authors": "Sota Kato, Kazuhiro Hotta",
+                    "published": "2021-08-30",
+                    "updated": "2021-08-30",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2304.11332v2",
+                    "title": "Input Augmentation with SAM: Boosting Medical Image Segmentation with Segmentation Foundation Model",
+                    "abstract": "The Segment Anything Model (SAM) is a recently developed large model for\ngeneral-purpose segmentation for computer vision tasks. SAM was trained using\n11 million images with over 1 billion masks and can produce segmentation\nresults for a wide range of objects in natural scene images. SAM can be viewed\nas a general perception model for segmentation (partitioning images into\nsemantically meaningful regions). Thus, how to utilize such a large foundation\nmodel for medical image segmentation is an emerging research target. This paper\nshows that although SAM does not immediately give high-quality segmentation for\nmedical image data, its generated masks, features, and stability scores are\nuseful for building and training better medical image segmentation models. In\nparticular, we demonstrate how to use SAM to augment image input for\ncommonly-used medical image segmentation models (e.g., U-Net). Experiments on\nthree segmentation tasks show the effectiveness of our proposed SAMAug method.\nThe code is available at \\url{https://github.com/yizhezhang2000/SAMAug}.",
+                    "authors": "Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Danny Z. Chen",
+                    "published": "2023-04-22",
+                    "updated": "2023-06-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2007.02361v1",
+                    "title": "Self-supervised Depth Estimation to Regularise Semantic Segmentation in Knee Arthroscopy",
+                    "abstract": "Intra-operative automatic semantic segmentation of knee joint structures can\nassist surgeons during knee arthroscopy in terms of situational awareness.\nHowever, due to poor imaging conditions (e.g., low texture, overexposure,\netc.), automatic semantic segmentation is a challenging scenario, which\njustifies the scarce literature on this topic. In this paper, we propose a\nnovel self-supervised monocular depth estimation to regularise the training of\nthe semantic segmentation in knee arthroscopy. To further regularise the depth\nestimation, we propose the use of clean training images captured by the stereo\narthroscope of routine objects (presenting none of the poor imaging conditions\nand with rich texture information) to pre-train the model. We fine-tune such\nmodel to produce both the semantic segmentation and self-supervised monocular\ndepth using stereo arthroscopic images taken from inside the knee. Using a data\nset containing 3868 arthroscopic images captured during cadaveric knee\narthroscopy with semantic segmentation annotations, 2000 stereo image pairs of\ncadaveric knee arthroscopy, and 2150 stereo image pairs of routine objects, we\nshow that our semantic segmentation regularised by self-supervised depth\nestimation produces a more accurate segmentation than a state-of-the-art\nsemantic segmentation approach modeled exclusively with semantic segmentation\nannotation.",
+                    "authors": "Fengbei Liu, Yaqub Jonmohamadi, Gabriel Maicas, Ajay K. Pandey, Gustavo Carneiro",
+                    "published": "2020-07-05",
+                    "updated": "2020-07-05",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.12903v1",
+                    "title": "Land Cover Change Detection via Semantic Segmentation",
+                    "abstract": "This paper presents a change detection method that identifies land cover\nchanges from aerial imagery, using semantic segmentation, a machine learning\napproach. We present a land cover classification training pipeline with Deeplab\nv3+, state-of-the-art semantic segmentation technology, including data\npreparation, model training for seven land cover types, and model exporting\nmodules. In the land cover change detection system, the inputs are images\nretrieved from Google Earth at the same location but from different times. The\nsystem then predicts semantic segmentation results on these images using the\ntrained model and calculates the land cover class percentage for each input\nimage. We see an improvement in the accuracy of the land cover semantic\nsegmentation model, with a mean IoU of 0.756 compared to 0.433, as reported in\nthe DeepGlobe land cover classification challenge. The land cover change\ndetection system that leverages the state-of-the-art semantic segmentation\ntechnology is proposed and can be used for deforestation analysis, land\nmanagement, and urban planning.",
+                    "authors": "Renee Su, Rong Chen",
+                    "published": "2019-11-28",
+                    "updated": "2019-11-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2212.07623v1",
+                    "title": "SBSS: Stacking-Based Semantic Segmentation Framework for Very High Resolution Remote Sensing Image",
+                    "abstract": "Semantic segmentation of Very High Resolution (VHR) remote sensing images is\na fundamental task for many applications. However, large variations in the\nscales of objects in those VHR images pose a challenge for performing accurate\nsemantic segmentation. Existing semantic segmentation networks are able to\nanalyse an input image at up to four resizing scales, but this may be\ninsufficient given the diversity of object scales. Therefore, Multi Scale (MS)\ntest-time data augmentation is often used in practice to obtain more accurate\nsegmentation results, which makes equal use of the segmentation results\nobtained at the different resizing scales. However, it was found in this study\nthat different classes of objects had their preferred resizing scale for more\naccurate semantic segmentation. Based on this behaviour, a Stacking-Based\nSemantic Segmentation (SBSS) framework is proposed to improve the segmentation\nresults by learning this behaviour, which contains a learnable Error Correction\nModule (ECM) for segmentation result fusion and an Error Correction Scheme\n(ECS) for computational complexity control. Two ECS, i.e., ECS-MS and ECS-SS,\nare proposed and investigated in this study. The Floating-point operations\n(Flops) required for ECS-MS and ECS-SS are similar to the commonly used MS test\nand the Single-Scale (SS) test, respectively. Extensive experiments on four\ndatasets (i.e., Cityscapes, UAVid, LoveDA and Potsdam) show that SBSS is an\neffective and flexible framework. It achieved higher accuracy than MS when\nusing ECS-MS, and similar accuracy as SS with a quarter of the memory footprint\nwhen using ECS-SS.",
+                    "authors": "Yuanzhi Cai, Lei Fan, Yuan Fang",
+                    "published": "2022-12-15",
+                    "updated": "2022-12-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.11857v1",
+                    "title": "Joint Learning of Intrinsic Images and Semantic Segmentation",
+                    "abstract": "Semantic segmentation of outdoor scenes is problematic when there are\nvariations in imaging conditions. It is known that albedo (reflectance) is\ninvariant to all kinds of illumination effects. Thus, using reflectance images\nfor semantic segmentation task can be favorable. Additionally, not only\nsegmentation may benefit from reflectance, but also segmentation may be useful\nfor reflectance computation. Therefore, in this paper, the tasks of semantic\nsegmentation and intrinsic image decomposition are considered as a combined\nprocess by exploring their mutual relationship in a joint fashion. To that end,\nwe propose a supervised end-to-end CNN architecture to jointly learn intrinsic\nimage decomposition and semantic segmentation. We analyze the gains of\naddressing those two problems jointly. Moreover, new cascade CNN architectures\nfor intrinsic-for-segmentation and segmentation-for-intrinsic are proposed as\nsingle tasks. Furthermore, a dataset of 35K synthetic images of natural\nenvironments is created with corresponding albedo and shading (intrinsics), as\nwell as semantic labels (segmentation) assigned to each object/scene. The\nexperiments show that joint learning of intrinsic image decomposition and\nsemantic segmentation is beneficial for both tasks for natural scenes. Dataset\nand models are available at: https://ivi.fnwi.uva.nl/cv/intrinseg",
+                    "authors": "Anil S. Baslamisli, Thomas T. Groenestege, Partha Das, Hoang-An Le, Sezer Karaoglu, Theo Gevers",
+                    "published": "2018-07-31",
+                    "updated": "2018-07-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.13395v3",
+                    "title": "ACDC: The Adverse Conditions Dataset with Correspondences for Semantic Driving Scene Understanding",
+                    "abstract": "Level 5 autonomy for self-driving cars requires a robust visual perception\nsystem that can parse input images under any visual condition. However,\nexisting semantic segmentation datasets are either dominated by images captured\nunder normal conditions or are small in scale. To address this, we introduce\nACDC, the Adverse Conditions Dataset with Correspondences for training and\ntesting semantic segmentation methods on adverse visual conditions. ACDC\nconsists of a large set of 4006 images which are equally distributed between\nfour common adverse conditions: fog, nighttime, rain, and snow. Each\nadverse-condition image comes with a high-quality fine pixel-level semantic\nannotation, a corresponding image of the same scene taken under normal\nconditions, and a binary mask that distinguishes between intra-image regions of\nclear and uncertain semantic content. Thus, ACDC supports both standard\nsemantic segmentation and the newly introduced uncertainty-aware semantic\nsegmentation. A detailed empirical study demonstrates the challenges that the\nadverse domains of ACDC pose to state-of-the-art supervised and unsupervised\napproaches and indicates the value of our dataset in steering future progress\nin the field. Our dataset and benchmark are publicly available.",
+                    "authors": "Christos Sakaridis, Dengxin Dai, Luc Van Gool",
+                    "published": "2021-04-27",
+                    "updated": "2021-09-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.13297v5",
+                    "title": "GPS-GLASS: Learning Nighttime Semantic Segmentation Using Daytime Video and GPS data",
+                    "abstract": "Semantic segmentation for autonomous driving should be robust against various\nin-the-wild environments. Nighttime semantic segmentation is especially\nchallenging due to a lack of annotated nighttime images and a large domain gap\nfrom daytime images with sufficient annotation. In this paper, we propose a\nnovel GPS-based training framework for nighttime semantic segmentation. Given\nGPS-aligned pairs of daytime and nighttime images, we perform cross-domain\ncorrespondence matching to obtain pixel-level pseudo supervision. Moreover, we\nconduct flow estimation between daytime video frames and apply GPS-based\nscaling to acquire another pixel-level pseudo supervision. Using these pseudo\nsupervisions with a confidence map, we train a nighttime semantic segmentation\nnetwork without any annotation from nighttime images. Experimental results\ndemonstrate the effectiveness of the proposed method on several nighttime\nsemantic segmentation datasets. Our source code is available at\nhttps://github.com/jimmy9704/GPS-GLASS.",
+                    "authors": "Hongjae Lee, Changwoo Han, Jun-Sang Yoo, Seung-Won Jung",
+                    "published": "2022-07-27",
+                    "updated": "2023-08-18",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.08988v1",
+                    "title": "Heterogeneous Feature Distillation Network for SAR Image Semantic Segmentation",
+                    "abstract": "Semantic segmentation for SAR (Synthetic Aperture Radar) images has attracted\nincreasing attention in the remote sensing community recently, due to SAR's\nall-time and all-weather imaging capability. However, SAR images are generally\nmore difficult to be segmented than their EO (Electro-Optical) counterparts,\nsince speckle noises and layovers are inevitably involved in SAR images. To\naddress this problem, we investigate how to introduce EO features to assist the\ntraining of a SAR-segmentation model, and propose a heterogeneous feature\ndistillation network for segmenting SAR images, called HFD-Net, where a\nSAR-segmentation student model gains knowledge from a pre-trained\nEO-segmentation teacher model. In the proposed HFD-Net, both the student and\nteacher models employ an identical architecture but different parameter\nconfigurations, and a heterogeneous feature distillation model is explored for\ntransferring latent EO features from the teacher model to the student model and\nthen enhancing the ability of the student model for SAR image segmentation. In\naddition, a heterogeneous feature alignment module is explored to aggregate\nmulti-scale features for segmentation in each of the student model and teacher\nmodel. Extensive experimental results on two public datasets demonstrate that\nthe proposed HFD-Net outperforms seven state-of-the-art SAR image semantic\nsegmentation methods.",
+                    "authors": "Gao Mengyu, Dong Qiulei",
+                    "published": "2022-10-17",
+                    "updated": "2022-10-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.01369v2",
+                    "title": "Exploring Limits of Diffusion-Synthetic Training with Weakly Supervised Semantic Segmentation",
+                    "abstract": "The advance of generative models for images has inspired various training\ntechniques for image recognition utilizing synthetic images. In semantic\nsegmentation, one promising approach is extracting pseudo-masks from attention\nmaps in text-to-image diffusion models, which enables\nreal-image-and-annotation-free training. However, the pioneering training\nmethod using the diffusion-synthetic images and pseudo-masks, i.e., DiffuMask\nhas limitations in terms of mask quality, scalability, and ranges of applicable\ndomains. To overcome these limitations, this work introduces three techniques\nfor diffusion-synthetic semantic segmentation training. First,\nreliability-aware robust training, originally used in weakly supervised\nlearning, helps segmentation with insufficient synthetic mask quality. %Second,\nlarge-scale pretraining of whole segmentation models, not only backbones, on\nsynthetic ImageNet-1k-class images with pixel-labels benefits downstream\nsegmentation tasks. Second, we introduce prompt augmentation, data augmentation\nto the prompt text set to scale up and diversify training images with a limited\ntext resources. Finally, LoRA-based adaptation of Stable Diffusion enables the\ntransfer to a distant domain, e.g., auto-driving images. Experiments in PASCAL\nVOC, ImageNet-S, and Cityscapes show that our method effectively closes gap\nbetween real and synthetic training in semantic segmentation.",
+                    "authors": "Ryota Yoshihashi, Yuya Otsuka, Kenji Doi, Tomohiro Tanaka, Hirokatsu Kataoka",
+                    "published": "2023-09-04",
+                    "updated": "2024-04-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1709.01956v1",
+                    "title": "Learning Dilation Factors for Semantic Segmentation of Street Scenes",
+                    "abstract": "Contextual information is crucial for semantic segmentation. However, finding\nthe optimal trade-off between keeping desired fine details and at the same time\nproviding sufficiently large receptive fields is non trivial. This is even more\nso, when objects or classes present in an image significantly vary in size.\nDilated convolutions have proven valuable for semantic segmentation, because\nthey allow to increase the size of the receptive field without sacrificing\nimage resolution. However, in current state-of-the-art methods, dilation\nparameters are hand-tuned and fixed. In this paper, we present an approach for\nlearning dilation parameters adaptively per channel, consistently improving\nsemantic segmentation results on street-scene datasets like Cityscapes and\nCamvid.",
+                    "authors": "Yang He, Margret Keuper, Bernt Schiele, Mario Fritz",
+                    "published": "2017-09-06",
+                    "updated": "2017-09-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2401.07589v1",
+                    "title": "Semantic Scene Segmentation for Robotics",
+                    "abstract": "Comprehensive scene understanding is a critical enabler of robot autonomy.\nSemantic segmentation is one of the key scene understanding tasks which is\npivotal for several robotics applications including autonomous driving,\ndomestic service robotics, last mile delivery, amongst many others. Semantic\nsegmentation is a dense prediction task that aims to provide a scene\nrepresentation in which each pixel of an image is assigned a semantic class\nlabel. Therefore, semantic segmentation considers the full scene context,\nincorporating the object category, location, and shape of all the scene\nelements, including the background. Numerous algorithms have been proposed for\nsemantic segmentation over the years. However, the recent advances in deep\nlearning combined with the boost in the computational capacity and the\navailability of large-scale labeled datasets have led to significant advances\nin semantic segmentation. In this chapter, we introduce the task of semantic\nsegmentation and present the deep learning techniques that have been proposed\nto address this task over the years. We first define the task of semantic\nsegmentation and contrast it with other closely related scene understanding\nproblems. We detail different algorithms and architectures for semantic\nsegmentation and the commonly employed loss functions. Furthermore, we present\nan overview of datasets, benchmarks, and metrics that are used in semantic\nsegmentation. We conclude the chapter with a discussion of challenges and\nopportunities for further research in this area.",
+                    "authors": "Juana Valeria Hurtado, Abhinav Valada",
+                    "published": "2024-01-15",
+                    "updated": "2024-01-15",
+                    "primary_cat": "cs.RO",
+                    "cats": [
+                        "cs.RO"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.00949v3",
+                    "title": "Synthetic Instance Segmentation from Semantic Image Segmentation Masks",
+                    "abstract": "In recent years, the development of instance segmentation has garnered\nsignificant attention in a wide range of applications. However, the training of\na fully-supervised instance segmentation model requires costly both\ninstance-level and pixel-level annotations. In contrast, weakly-supervised\ninstance segmentation methods (i.e., with image-level class labels or point\nlabels) struggle to satisfy the accuracy and recall requirements of practical\nscenarios. In this paper, we propose a novel paradigm called synthetic instance\nsegmentation (SISeg), which achieves Instance Segmentation results from image\nmasks predicted using off-the-shelf semantic segmentation models. SISeg does\nnot require training a semantic or/and instance segmentation model and avoids\nthe need for instance-level image annotations. Therefore, it is highly\nefficient. Specifically, we first obtain a semantic segmentation mask of the\ninput image via a trained semantic segmentation model. Then, we calculate a\ndisplacement field vector for each pixel based on the segmentation mask, which\ncan indicate representations belonging to the same class but different\ninstances, i.e., obtaining the instance-level object information. Finally,\ninstance segmentation results are obtained after being refined by a learnable\ncategory-agnostic object boundary branch. Extensive experimental results on two\nchallenging datasets and representative semantic segmentation baselines\n(including CNNs and Transformers) demonstrate that SISeg can achieve\ncompetitive results compared to the state-of-the-art fully-supervised instance\nsegmentation methods without the need for additional human resources or\nincreased computational costs. The code is available at: SISeg",
+                    "authors": "Yuchen Shen, Dong Zhang, Yuhui Zheng, Zechao Li, Liyong Fu, Qiaolin Ye",
+                    "published": "2023-08-02",
+                    "updated": "2023-10-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.08455v3",
+                    "title": "Open-world Semantic Segmentation via Contrasting and Clustering Vision-Language Embedding",
+                    "abstract": "To bridge the gap between supervised semantic segmentation and real-world\napplications that acquires one model to recognize arbitrary new concepts,\nrecent zero-shot segmentation attracts a lot of attention by exploring the\nrelationships between unseen and seen object categories, yet requiring large\namounts of densely-annotated data with diverse base classes. In this paper, we\npropose a new open-world semantic segmentation pipeline that makes the first\nattempt to learn to segment semantic objects of various open-world categories\nwithout any efforts on dense annotations, by purely exploiting the\nimage-caption data that naturally exist on the Internet. Our method,\nVision-language-driven Semantic Segmentation (ViL-Seg), employs an image and a\ntext encoder to generate visual and text embeddings for the image-caption data,\nwith two core components that endow its segmentation ability: First, the image\nencoder is jointly trained with a vision-based contrasting and a cross-modal\ncontrasting, which encourage the visual embeddings to preserve both\nfine-grained semantics and high-level category information that are crucial for\nthe segmentation task. Furthermore, an online clustering head is devised over\nthe image encoder, which allows to dynamically segment the visual embeddings\ninto distinct semantic groups such that they can be classified by comparing\nwith various text embeddings to complete our segmentation pipeline. Experiments\nshow that without using any data with dense annotations, our method can\ndirectly segment objects of arbitrary categories, outperforming zero-shot\nsegmentation methods that require data labeling on three benchmark datasets.",
+                    "authors": "Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, Xiaodan Liang",
+                    "published": "2022-07-18",
+                    "updated": "2022-10-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2101.09642v2",
+                    "title": "Image Compression with Encoder-Decoder Matched Semantic Segmentation",
+                    "abstract": "In recent years, layered image compression is demonstrated to be a promising\ndirection, which encodes a compact representation of the input image and apply\nan up-sampling network to reconstruct the image. To further improve the quality\nof the reconstructed image, some works transmit the semantic segment together\nwith the compressed image data. Consequently, the compression ratio is also\ndecreased because extra bits are required for transmitting the semantic\nsegment. To solve this problem, we propose a new layered image compression\nframework with encoder-decoder matched semantic segmentation (EDMS). And then,\nfollowed by the semantic segmentation, a special convolution neural network is\nused to enhance the inaccurate semantic segment. As a result, the accurate\nsemantic segment can be obtained in the decoder without requiring extra bits.\nThe experimental results show that the proposed EDMS framework can get up to\n35.31% BD-rate reduction over the HEVC-based (BPG) codec, 5% bitrate, and 24%\nencoding time saving compare to the state-of-the-art semantic-based image\ncodec.",
+                    "authors": "Trinh Man Hoang, Jinjia Zhou, Yibo Fan",
+                    "published": "2021-01-24",
+                    "updated": "2021-01-30",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV",
+                        "cs.MM"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2004.10126v3",
+                    "title": "Generative Synthetic Augmentation using Label-to-Image Translation for Nuclei Image Segmentation",
+                    "abstract": "In medical image diagnosis, pathology image analysis using semantic\nsegmentation becomes important for efficient screening as a field of digital\npathology. The spatial augmentation is ordinary used for semantic segmentation.\nTumor images under malignant are rare and to annotate the labels of nuclei\nregion takes much time-consuming. We require an effective use of dataset to\nmaximize the segmentation accuracy. It is expected that some augmentation to\ntransform generalized images influence the segmentation performance. We propose\na synthetic augmentation using label-to-image translation, mapping from a\nsemantic label with the edge structure to a real image. Exactly this paper deal\nwith stain slides of nuclei in tumor. Actually, we demonstrate several\nsegmentation algorithms applied to the initial dataset that contains real\nimages and labels using synthetic augmentation in order to add their\ngeneralized images. We computes and reports that a proposed synthetic\naugmentation procedure improve their accuracy.",
+                    "authors": "Takato Yasuno",
+                    "published": "2020-04-21",
+                    "updated": "2021-03-02",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "eess.IV",
+                        "stat.ML",
+                        "I.4.6; I.2.6"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2112.03241v1",
+                    "title": "Unsupervised Domain Adaptation for Semantic Image Segmentation: a Comprehensive Survey",
+                    "abstract": "Semantic segmentation plays a fundamental role in a broad variety of computer\nvision applications, providing key information for the global understanding of\nan image. Yet, the state-of-the-art models rely on large amount of annotated\nsamples, which are more expensive to obtain than in tasks such as image\nclassification. Since unlabelled data is instead significantly cheaper to\nobtain, it is not surprising that Unsupervised Domain Adaptation reached a\nbroad success within the semantic segmentation community.\n  This survey is an effort to summarize five years of this incredibly rapidly\ngrowing field, which embraces the importance of semantic segmentation itself\nand a critical need of adapting segmentation models to new environments. We\npresent the most important semantic segmentation methods; we provide a\ncomprehensive survey on domain adaptation techniques for semantic segmentation;\nwe unveil newer trends such as multi-domain learning, domain generalization,\ntest-time adaptation or source-free domain adaptation; we conclude this survey\nby describing datasets and benchmarks most widely used in semantic segmentation\nresearch. We hope that this survey will provide researchers across academia and\nindustry with a comprehensive reference guide and will help them in fostering\nnew research directions in the field.",
+                    "authors": "Gabriela Csurka, Riccardo Volpi, Boris Chidlovskii",
+                    "published": "2021-12-06",
+                    "updated": "2021-12-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "I.4.6; I.2"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1606.01481v1",
+                    "title": "Better Image Segmentation by Exploiting Dense Semantic Predictions",
+                    "abstract": "It is well accepted that image segmentation can benefit from utilizing\nmultilevel cues. The paper focuses on utilizing the FCNN-based dense semantic\npredictions in the bottom-up image segmentation, arguing to take semantic cues\ninto account from the very beginning. By this we can avoid merging regions of\nsimilar appearance but distinct semantic categories as possible. The semantic\ninefficiency problem is handled. We also propose a straightforward way to use\nthe contour cues to suppress the noise in multilevel cues, thus to improve the\nsegmentation robustness. The evaluation on the BSDS500 shows that we obtain the\ncompetitive region and boundary performance. Furthermore, since all individual\nregions can be assigned with appropriate semantic labels during the\ncomputation, we are capable of extracting the adjusted semantic segmentations.\nThe experiment on Pascal VOC 2012 shows our improvement to the original\nsemantic segmentations which derives directly from the dense predictions.",
+                    "authors": "Qiyang Zhao, Lewis D Griffin",
+                    "published": "2016-06-05",
+                    "updated": "2016-06-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.02840v1",
+                    "title": "Attention-based fusion of semantic boundary and non-boundary information to improve semantic segmentation",
+                    "abstract": "This paper introduces a method for image semantic segmentation grounded on a\nnovel fusion scheme, which takes place inside a deep convolutional neural\nnetwork. The main goal of our proposal is to explore object boundary\ninformation to improve the overall segmentation performance. Unlike previous\nworks that combine boundary and segmentation features, or those that use\nboundary information to regularize semantic segmentation, we instead propose a\nnovel approach that embodies boundary information onto segmentation. For that,\nour semantic segmentation method uses two streams, which are combined through\nan attention gate, forming an end-to-end Y-model. To the best of our knowledge,\nours is the first work to show that boundary detection can improve semantic\nsegmentation when fused through a semantic fusion gate (attention model). We\nperformed an extensive evaluation of our method over public data sets. We found\ncompetitive results on all data sets after comparing our proposed model with\nother twelve state-of-the-art segmenters, considering the same training\nconditions. Our proposed model achieved the best mIoU on the CityScapes,\nCamVid, and Pascal Context data sets, and the second best on Mapillary Vistas.",
+                    "authors": "Jefferson Fontinele, Gabriel Lefundes, Luciano Oliveira",
+                    "published": "2021-08-05",
+                    "updated": "2021-08-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.03983v1",
+                    "title": "Weakly Supervised Semantic Segmentation of Satellite Images",
+                    "abstract": "When one wants to train a neural network to perform semantic segmentation,\ncreating pixel-level annotations for each of the images in the database is a\ntedious task. If he works with aerial or satellite images, which are usually\nvery large, it is even worse. With that in mind, we investigate how to use\nimage-level annotations in order to perform semantic segmentation. Image-level\nannotations are much less expensive to acquire than pixel-level annotations,\nbut we lose a lot of information for the training of the model. From the\nannotations of the images, the model must find by itself how to classify the\ndifferent regions of the image. In this work, we use the method proposed by Anh\nand Kwak [1] to produce pixel-level annotation from image level annotation. We\ncompare the overall quality of our generated dataset with the original dataset.\nIn addition, we propose an adaptation of the AffinityNet that allows us to\ndirectly perform a semantic segmentation. Our results show that the generated\nlabels lead to the same performances for the training of several segmentation\nnetworks. Also, the quality of semantic segmentation performed directly by the\nAffinityNet and the Random Walk is close to the one of the best\nfully-supervised approaches.",
+                    "authors": "Adrien Nivaggioli, Hicham Randrianarivo",
+                    "published": "2019-04-08",
+                    "updated": "2019-04-08",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1509.02441v1",
+                    "title": "Semantic Video Segmentation : Exploring Inference Efficiency",
+                    "abstract": "We explore the efficiency of the CRF inference beyond image level semantic\nsegmentation and perform joint inference in video frames. The key idea is to\ncombine best of two worlds: semantic co-labeling and more expressive models.\nOur formulation enables us to perform inference over ten thousand images within\nseconds and makes the system amenable to perform video semantic segmentation\nmost effectively. On CamVid dataset, with TextonBoost unaries, our proposed\nmethod achieves up to 8% improvement in accuracy over individual semantic image\nsegmentation without additional time overhead. The source code is available at\nhttps://github.com/subtri/video_inference",
+                    "authors": "Subarna Tripathi, Serge Belongie, Youngbae Hwang, Truong Nguyen",
+                    "published": "2015-09-04",
+                    "updated": "2015-09-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1811.00174v4",
+                    "title": "Pixel Level Data Augmentation for Semantic Image Segmentation using Generative Adversarial Networks",
+                    "abstract": "Semantic segmentation is one of the basic topics in computer vision, it aims\nto assign semantic labels to every pixel of an image. Unbalanced semantic label\ndistribution could have a negative influence on segmentation accuracy. In this\npaper, we investigate using data augmentation approach to balance the semantic\nlabel distribution in order to improve segmentation performance. We propose\nusing generative adversarial networks (GANs) to generate realistic images for\nimproving the performance of semantic segmentation networks. Experimental\nresults show that the proposed method can not only improve segmentation\nperformance on those classes with low accuracy, but also obtain 1.3% to 2.1%\nincrease in average segmentation accuracy. It shows that this augmentation\nmethod can boost accuracy and be easily applicable to any other segmentation\nmodels.",
+                    "authors": "Shuangting Liu, Jiaqi Zhang, Yuxin Chen, Yifan Liu, Zengchang Qin, Tao Wan",
+                    "published": "2018-11-01",
+                    "updated": "2019-11-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2006.07601v1",
+                    "title": "NoPeopleAllowed: The Three-Step Approach to Weakly Supervised Semantic Segmentation",
+                    "abstract": "We propose a novel approach to weakly supervised semantic segmentation, which\nconsists of three consecutive steps. The first two steps extract high-quality\npseudo masks from image-level annotated data, which are then used to train a\nsegmentation model on the third step. The presented approach also addresses two\nproblems in the data: class imbalance and missing labels. Using only\nimage-level annotations as supervision, our method is capable of segmenting\nvarious classes and complex objects. It achieves 37.34 mean IoU on the test\nset, placing 3rd at the LID Challenge in the task of weakly supervised semantic\nsegmentation.",
+                    "authors": "Mariia Dobko, Ostap Viniavskyi, Oles Dobosevych",
+                    "published": "2020-06-13",
+                    "updated": "2020-06-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2001.00335v1",
+                    "title": "Graph-FCN for image semantic segmentation",
+                    "abstract": "Semantic segmentation with deep learning has achieved great progress in\nclassifying the pixels in the image. However, the local location information is\nusually ignored in the high-level feature extraction by the deep learning,\nwhich is important for image semantic segmentation. To avoid this problem, we\npropose a graph model initialized by a fully convolutional network (FCN) named\nGraph-FCN for image semantic segmentation. Firstly, the image grid data is\nextended to graph structure data by a convolutional network, which transforms\nthe semantic segmentation problem into a graph node classification problem.\nThen we apply graph convolutional network to solve this graph node\nclassification problem. As far as we know, it is the first time that we apply\nthe graph convolutional network in image semantic segmentation. Our method\nachieves competitive performance in mean intersection over union (mIOU) on the\nVOC dataset(about 1.34% improvement), compared to the original FCN model.",
+                    "authors": "Yi Lu, Yaran Chen, Dongbin Zhao, Jianxin Chen",
+                    "published": "2020-01-02",
+                    "updated": "2020-01-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1812.10885v1",
+                    "title": "Coarse-to-fine Semantic Segmentation from Image-level Labels",
+                    "abstract": "Deep neural network-based semantic segmentation generally requires\nlarge-scale cost extensive annotations for training to obtain better\nperformance. To avoid pixel-wise segmentation annotations which are needed for\nmost methods, recently some researchers attempted to use object-level labels\n(e.g. bounding boxes) or image-level labels (e.g. image categories). In this\npaper, we propose a novel recursive coarse-to-fine semantic segmentation\nframework based on only image-level category labels. For each image, an initial\ncoarse mask is first generated by a convolutional neural network-based\nunsupervised foreground segmentation model and then is enhanced by a graph\nmodel. The enhanced coarse mask is fed to a fully convolutional neural network\nto be recursively refined. Unlike existing image-level label-based semantic\nsegmentation methods which require to label all categories for images contain\nmultiple types of objects, our framework only needs one label for each image\nand can handle images contains multi-category objects. With only trained on\nImageNet, our framework achieves comparable performance on PASCAL VOC dataset\nas other image-level label-based state-of-the-arts of semantic segmentation.\nFurthermore, our framework can be easily extended to foreground object\nsegmentation task and achieves comparable performance with the state-of-the-art\nsupervised methods on the Internet Object dataset.",
+                    "authors": "Longlong Jing, Yucheng Chen, Yingli Tian",
+                    "published": "2018-12-28",
+                    "updated": "2018-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.13026v1",
+                    "title": "Weakly-Supervised Semantic Segmentation with Image-Level Labels: from Traditional Models to Foundation Models",
+                    "abstract": "The rapid development of deep learning has driven significant progress in the\nfield of image semantic segmentation - a fundamental task in computer vision.\nSemantic segmentation algorithms often depend on the availability of\npixel-level labels (i.e., masks of objects), which are expensive,\ntime-consuming, and labor-intensive. Weakly-supervised semantic segmentation\n(WSSS) is an effective solution to avoid such labeling. It utilizes only\npartial or incomplete annotations and provides a cost-effective alternative to\nfully-supervised semantic segmentation. In this paper, we focus on the WSSS\nwith image-level labels, which is the most challenging form of WSSS. Our work\nhas two parts. First, we conduct a comprehensive survey on traditional methods,\nprimarily focusing on those presented at premier research conferences. We\ncategorize them into four groups based on where their methods operate:\npixel-wise, image-wise, cross-image, and external data. Second, we investigate\nthe applicability of visual foundation models, such as the Segment Anything\nModel (SAM), in the context of WSSS. We scrutinize SAM in two intriguing\nscenarios: text prompting and zero-shot learning. We provide insights into the\npotential and challenges associated with deploying visual foundational models\nfor WSSS, facilitating future developments in this exciting research area.",
+                    "authors": "Zhaozheng Chen, Qianru Sun",
+                    "published": "2023-10-19",
+                    "updated": "2023-10-19",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.15608v1",
+                    "title": "Semantic Segmentation by Semantic Proportions",
+                    "abstract": "Semantic segmentation is a critical task in computer vision that aims to\nidentify and classify individual pixels in an image, with numerous applications\nfor example autonomous driving and medical image analysis. However, semantic\nsegmentation can be super challenging particularly due to the need for large\namounts of annotated data. Annotating images is a time-consuming and costly\nprocess, often requiring expert knowledge and significant effort. In this\npaper, we propose a novel approach for semantic segmentation by eliminating the\nneed of ground-truth segmentation maps. Instead, our approach requires only the\nrough information of individual semantic class proportions, shortened as\nsemantic proportions. It greatly simplifies the data annotation process and\nthus will significantly reduce the annotation time and cost, making it more\nfeasible for large-scale applications. Moreover, it opens up new possibilities\nfor semantic segmentation tasks where obtaining the full ground-truth\nsegmentation maps may not be feasible or practical. Extensive experimental\nresults demonstrate that our approach can achieve comparable and sometimes even\nbetter performance against the benchmark method that relies on the ground-truth\nsegmentation maps. Utilising semantic proportions suggested in this work offers\na promising direction for future research in the field of semantic\nsegmentation.",
+                    "authors": "Halil Ibrahim Aysel, Xiaohao Cai, Adam Pr\u00fcgel-Bennett",
+                    "published": "2023-05-24",
+                    "updated": "2023-05-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.00679v3",
+                    "title": "Cooperative Semantic Segmentation and Image Restoration in Adverse Environmental Conditions",
+                    "abstract": "Most state-of-the-art semantic segmentation approaches only achieve high\naccuracy in good conditions. In practically-common but less-discussed adverse\nenvironmental conditions, their performance can decrease enormously. Existing\nstudies usually cast the handling of segmentation in adverse conditions as a\nseparate post-processing step after signal restoration, making the segmentation\nperformance largely depend on the quality of restoration. In this paper, we\npropose a novel deep-learning framework to tackle semantic segmentation and\nimage restoration in adverse environmental conditions in a holistic manner. The\nproposed approach contains two components: Semantically-Guided Adaptation,\nwhich exploits semantic information from degraded images to refine the\nsegmentation; and Exemplar-Guided Synthesis, which restores images from\nsemantic label maps given degraded exemplars as the guidance. Our method\ncooperatively leverages the complementarity and interdependence of low-level\nrestoration and high-level segmentation in adverse environmental conditions.\nExtensive experiments on various datasets demonstrate that our approach can not\nonly improve the accuracy of semantic segmentation with degradation cues, but\nalso boost the perceptual quality and structural similarity of image\nrestoration with semantic guidance.",
+                    "authors": "Weihao Xia, Zhanglin Cheng, Yujiu Yang, Jing-Hao Xue",
+                    "published": "2019-11-02",
+                    "updated": "2020-03-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1605.06885v1",
+                    "title": "Bridging Category-level and Instance-level Semantic Image Segmentation",
+                    "abstract": "We propose an approach to instance-level image segmentation that is built on\ntop of category-level segmentation. Specifically, for each pixel in a semantic\ncategory mask, its corresponding instance bounding box is predicted using a\ndeep fully convolutional regression network. Thus it follows a different\npipeline to the popular detect-then-segment approaches that first predict\ninstances' bounding boxes, which are the current state-of-the-art in instance\nsegmentation. We show that, by leveraging the strength of our state-of-the-art\nsemantic segmentation models, the proposed method can achieve comparable or\neven better results to detect-then-segment approaches. We make the following\ncontributions. (i) First, we propose a simple yet effective approach to\nsemantic instance segmentation. (ii) Second, we propose an online bootstrapping\nmethod during training, which is critically important for achieving good\nperformance for both semantic category segmentation and instance-level\nsegmentation. (iii) As the performance of semantic category segmentation has a\nsignificant impact on the instance-level segmentation, which is the second step\nof our approach, we train fully convolutional residual networks to achieve the\nbest semantic category segmentation accuracy. On the PASCAL VOC 2012 dataset,\nwe obtain the currently best mean intersection-over-union score of 79.1%. (iv)\nWe also achieve state-of-the-art results for instance-level segmentation.",
+                    "authors": "Zifeng Wu, Chunhua Shen, Anton van den Hengel",
+                    "published": "2016-05-23",
+                    "updated": "2016-05-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2106.04108v3",
+                    "title": "Fully Transformer Networks for Semantic Image Segmentation",
+                    "abstract": "Transformers have shown impressive performance in various natural language\nprocessing and computer vision tasks, due to the capability of modeling\nlong-range dependencies. Recent progress has demonstrated that combining such\nTransformers with CNN-based semantic image segmentation models is very\npromising. However, it is not well studied yet on how well a pure Transformer\nbased approach can achieve for image segmentation. In this work, we explore a\nnovel framework for semantic image segmentation, which is encoder-decoder based\nFully Transformer Networks (FTN). Specifically, we first propose a Pyramid\nGroup Transformer (PGT) as the encoder for progressively learning hierarchical\nfeatures, meanwhile reducing the computation complexity of the standard Visual\nTransformer (ViT). Then, we propose a Feature Pyramid Transformer (FPT) to fuse\nsemantic-level and spatial-level information from multiple levels of the PGT\nencoder for semantic image segmentation. Surprisingly, this simple baseline can\nachieve better results on multiple challenging semantic segmentation and face\nparsing benchmarks, including PASCAL Context, ADE20K, COCOStuff, and\nCelebAMask-HQ. The source code will be released on\nhttps://github.com/BR-IDL/PaddleViT.",
+                    "authors": "Sitong Wu, Tianyi Wu, Fangjian Lin, Shengwei Tian, Guodong Guo",
+                    "published": "2021-06-08",
+                    "updated": "2021-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2404.04231v1",
+                    "title": "Image-Text Co-Decomposition for Text-Supervised Semantic Segmentation",
+                    "abstract": "This paper addresses text-supervised semantic segmentation, aiming to learn a\nmodel capable of segmenting arbitrary visual concepts within images by using\nonly image-text pairs without dense annotations. Existing methods have\ndemonstrated that contrastive learning on image-text pairs effectively aligns\nvisual segments with the meanings of texts. We notice that there is a\ndiscrepancy between text alignment and semantic segmentation: A text often\nconsists of multiple semantic concepts, whereas semantic segmentation strives\nto create semantically homogeneous segments. To address this issue, we propose\na novel framework, Image-Text Co-Decomposition (CoDe), where the paired image\nand text are jointly decomposed into a set of image regions and a set of word\nsegments, respectively, and contrastive learning is developed to enforce\nregion-word alignment. To work with a vision-language model, we present a\nprompt learning mechanism that derives an extra representation to highlight an\nimage segment or a word segment of interest, with which more effective features\ncan be extracted from that segment. Comprehensive experimental results\ndemonstrate that our method performs favorably against existing text-supervised\nsemantic segmentation methods on six benchmark datasets.",
+                    "authors": "Ji-Jia Wu, Andy Chia-Hao Chang, Chieh-Yu Chuang, Chun-Pei Chen, Yu-Lun Liu, Min-Hung Chen, Hou-Ning Hu, Yung-Yu Chuang, Yen-Yu Lin",
+                    "published": "2024-04-05",
+                    "updated": "2024-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.03138v1",
+                    "title": "Quantity beats quality for semantic segmentation of corrosion in images",
+                    "abstract": "Dataset creation is typically one of the first steps when applying Artificial\nIntelligence methods to a new task; and the real world performance of models\nhinges on the quality and quantity of data available. Producing an image\ndataset for semantic segmentation is resource intensive, particularly for\nspecialist subjects where class segmentation is not able to be effectively\nfarmed out. The benefit of producing a large, but poorly labelled, dataset\nversus a small, expertly segmented dataset for semantic segmentation is an open\nquestion. Here we show that a large, noisy dataset outperforms a small,\nexpertly segmented dataset for training a Fully Convolutional Network model for\nsemantic segmentation of corrosion in images. A large dataset of 250 images\nwith segmentations labelled by undergraduates and a second dataset of just 10\nimages, with segmentations labelled by subject matter experts were produced.\nThe mean Intersection over Union and micro F-score metrics were compared after\ntraining for 50,000 epochs. This work is illustrative for researchers setting\nout to develop deep learning models for detection and location of specialist\nfeatures.",
+                    "authors": "Will Nash, Tom Drummond, Nick Birbilis",
+                    "published": "2018-06-30",
+                    "updated": "2018-06-30",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.05321v1",
+                    "title": "Image Segmentation Semantic Communication over Internet of Vehicles",
+                    "abstract": "In this paper, the problem of semantic-based efficient image transmission is\nstudied over the Internet of Vehicles (IoV). In the considered model, a vehicle\nshares massive amount of visual data perceived by its visual sensors to assist\nother vehicles in making driving decisions. However, it is hard to maintain a\nhigh reliable visual data transmission due to the limited spectrum resources.\nTo tackle this problem, a semantic communication approach is introduced to\nreduce the transmission data amount while ensuring the semantic-level accuracy.\nParticularly, an image segmentation semantic communication (ISSC) system is\nproposed, which can extract the semantic features from the perceived images and\ntransmit the features to the receiving vehicle that reconstructs the image\nsegmentations. The ISSC system consists of an encoder and a decoder at the\ntransmitter and the receiver, respectively. To accurately extract the image\nsemantic features, the ISSC system encoder employs a Swin Transformer based\nmulti-scale semantic feature extractor. Then, to resist the wireless noise and\nreconstruct the image segmentation, a semantic feature decoder and a\nreconstructor are designed at the receiver. Simulation results show that the\nproposed ISSC system can reconstruct the image segmentation accurately with a\nhigh compression ratio, and can achieve robust transmission performance against\nchannel noise, especially at the low signal-to-noise ratio (SNR). In terms of\nmean Intersection over Union (mIoU), the ISSC system can achieve an increase by\n75%, compared to the baselines using traditional coding method",
+                    "authors": "Qiang Pan, Haonan Tong, Jie Lv, Tao Luo, Zhilong Zhang, Changchuan Yin, Jianfeng Li",
+                    "published": "2022-10-11",
+                    "updated": "2022-10-11",
+                    "primary_cat": "cs.NI",
+                    "cats": [
+                        "cs.NI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1803.10464v2",
+                    "title": "Learning Pixel-level Semantic Affinity with Image-level Supervision for Weakly Supervised Semantic Segmentation",
+                    "abstract": "The deficiency of segmentation labels is one of the main obstacles to\nsemantic segmentation in the wild. To alleviate this issue, we present a novel\nframework that generates segmentation labels of images given their image-level\nclass labels. In this weakly supervised setting, trained models have been known\nto segment local discriminative parts rather than the entire object area. Our\nsolution is to propagate such local responses to nearby areas which belong to\nthe same semantic entity. To this end, we propose a Deep Neural Network (DNN)\ncalled AffinityNet that predicts semantic affinity between a pair of adjacent\nimage coordinates. The semantic propagation is then realized by random walk\nwith the affinities predicted by AffinityNet. More importantly, the supervision\nemployed to train AffinityNet is given by the initial discriminative part\nsegmentation, which is incomplete as a segmentation annotation but sufficient\nfor learning semantic affinities within small image areas. Thus the entire\nframework relies only on image-level class labels and does not require any\nextra data or annotations. On the PASCAL VOC 2012 dataset, a DNN learned with\nsegmentation labels generated by our method outperforms previous models trained\nwith the same level of supervision, and is even as competitive as those relying\non stronger supervision.",
+                    "authors": "Jiwoon Ahn, Suha Kwak",
+                    "published": "2018-03-28",
+                    "updated": "2018-04-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1905.08748v3",
+                    "title": "RIU-Net: Embarrassingly simple semantic segmentation of 3D LiDAR point cloud",
+                    "abstract": "This paper proposes RIU-Net (for Range-Image U-Net), the adaptation of a\npopular semantic segmentation network for the semantic segmentation of a 3D\nLiDAR point cloud. The point cloud is turned into a 2D range-image by\nexploiting the topology of the sensor. This image is then used as input to a\nU-net. This architecture has already proved its efficiency for the task of\nsemantic segmentation of medical images. We demonstrate how it can also be used\nfor the accurate semantic segmentation of a 3D LiDAR point cloud and how it\nrepresents a valid bridge between image processing and 3D point cloud\nprocessing. Our model is trained on range-images built from KITTI 3D object\ndetection dataset. Experiments show that RIU-Net, despite being very simple,\noffers results that are comparable to the state-of-the-art of range-image based\nmethods. Finally, we demonstrate that this architecture is able to operate at\n90fps on a single GPU, which enables deployment for real-time segmentation.",
+                    "authors": "Pierre Biasutti, Aur\u00e9lie Bugeau, Jean-Fran\u00e7ois Aujol, Mathieu Br\u00e9dif",
+                    "published": "2019-05-21",
+                    "updated": "2019-06-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.06974v1",
+                    "title": "A One Stop 3D Target Reconstruction and multilevel Segmentation Method",
+                    "abstract": "3D object reconstruction and multilevel segmentation are fundamental to\ncomputer vision research. Existing algorithms usually perform 3D scene\nreconstruction and target objects segmentation independently, and the\nperformance is not fully guaranteed due to the challenge of the 3D\nsegmentation. Here we propose an open-source one stop 3D target reconstruction\nand multilevel segmentation framework (OSTRA), which performs segmentation on\n2D images, tracks multiple instances with segmentation labels in the image\nsequence, and then reconstructs labelled 3D objects or multiple parts with\nMulti-View Stereo (MVS) or RGBD-based 3D reconstruction methods. We extend\nobject tracking and 3D reconstruction algorithms to support continuous\nsegmentation labels to leverage the advances in the 2D image segmentation,\nespecially the Segment-Anything Model (SAM) which uses the pretrained neural\nnetwork without additional training for new scenes, for 3D object segmentation.\nOSTRA supports most popular 3D object models including point cloud, mesh and\nvoxel, and achieves high performance for semantic segmentation, instance\nsegmentation and part segmentation on several 3D datasets. It even surpasses\nthe manual segmentation in scenes with complex structures and occlusions. Our\nmethod opens up a new avenue for reconstructing 3D targets embedded with rich\nmulti-scale segmentation information in complex scenes. OSTRA is available from\nhttps://github.com/ganlab/OSTRA.",
+                    "authors": "Jiexiong Xu, Weikun Zhao, Zhiyan Tang, Xiangchao Gan",
+                    "published": "2023-08-14",
+                    "updated": "2023-08-14",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.10122v2",
+                    "title": "Weakly-supervised Semantic Segmentation in Cityscape via Hyperspectral Image",
+                    "abstract": "High-resolution hyperspectral images (HSIs) contain the response of each\npixel in different spectral bands, which can be used to effectively distinguish\nvarious objects in complex scenes. While HSI cameras have become low cost,\nalgorithms based on it have not been well exploited. In this paper, we focus on\na novel topic, weakly-supervised semantic segmentation in cityscape via HSIs.\nIt is based on the idea that high-resolution HSIs in city scenes contain rich\nspectral information, which can be easily associated to semantics without\nmanual labeling. Therefore, it enables low cost, highly reliable semantic\nsegmentation in complex scenes. Specifically, in this paper, we theoretically\nanalyze the HSIs and introduce a weakly-supervised HSI semantic segmentation\nframework, which utilizes spectral information to improve the coarse labels to\na finer degree. The experimental results show that our method can obtain highly\ncompetitive labels and even have higher edge fineness than artificial fine\nlabels in some classes. At the same time, the results also show that the\nrefined labels can effectively improve the effect of semantic segmentation. The\ncombination of HSIs and semantic segmentation proves that HSIs have great\npotential in high-level visual tasks.",
+                    "authors": "Yuxing Huang, Shaodi You, Ying Fu, Qiu Shen",
+                    "published": "2020-12-18",
+                    "updated": "2021-07-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.06753v1",
+                    "title": "3rd Place Solution for PVUW Challenge 2023: Video Panoptic Segmentation",
+                    "abstract": "In order to deal with the task of video panoptic segmentation in the wild, we\npropose a robust integrated video panoptic segmentation solution. In our\nsolution, we regard the video panoptic segmentation task as a segmentation\ntarget querying task, represent both semantic and instance targets as a set of\nqueries, and then combine these queries with video features extracted by neural\nnetworks to predict segmentation masks. In order to improve the learning\naccuracy and convergence speed of the solution, we add additional tasks of\nvideo semantic segmentation and video instance segmentation for joint training.\nIn addition, we also add an additional image semantic segmentation model to\nfurther improve the performance of semantic classes. In addition, we also add\nsome additional operations to improve the robustness of the model. Extensive\nexperiments on the VIPSeg dataset show that the proposed solution achieves\nstate-of-the-art performance with 50.04\\% VPQ on the VIPSeg test set, which is\n3rd place on the video panoptic segmentation track of the PVUW Challenge 2023.",
+                    "authors": "Jinming Su, Wangwang Yang, Junfeng Luo, Xiaolin Wei",
+                    "published": "2023-06-11",
+                    "updated": "2023-06-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2202.04754v2",
+                    "title": "Wireless Transmission of Images With The Assistance of Multi-level Semantic Information",
+                    "abstract": "Semantic-oriented communication has been considered as a promising to boost\nthe bandwidth efficiency by only transmitting the semantics of the data. In\nthis paper, we propose a multi-level semantic aware communication system for\nwireless image transmission, named MLSC-image, which is based on the deep\nlearning techniques and trained in an end to end manner. In particular, the\nproposed model includes a multilevel semantic feature extractor, that extracts\nboth the highlevel semantic information, such as the text semantics and the\nsegmentation semantics, and the low-level semantic information, such as local\nspatial details of the images. We employ a pretrained image caption to capture\nthe text semantics and a pretrained image segmentation model to obtain the\nsegmentation semantics. These high-level and low-level semantic features are\nthen combined and encoded by a joint semantic and channel encoder into symbols\nto transmit over the physical channel. The numerical results validate the\neffectiveness and efficiency of the proposed semantic communication system,\nespecially under the limited bandwidth condition, which indicates the\nadvantages of the high-level semantics in the compression of images.",
+                    "authors": "Zhenguo Zhang, Qianqian Yang, Shibo He, Mingyang Sun, Jiming Chen",
+                    "published": "2022-02-08",
+                    "updated": "2023-12-08",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.17874v1",
+                    "title": "SmooSeg: Smoothness Prior for Unsupervised Semantic Segmentation",
+                    "abstract": "Unsupervised semantic segmentation is a challenging task that segments images\ninto semantic groups without manual annotation. Prior works have primarily\nfocused on leveraging prior knowledge of semantic consistency or priori\nconcepts from self-supervised learning methods, which often overlook the\ncoherence property of image segments. In this paper, we demonstrate that the\nsmoothness prior, asserting that close features in a metric space share the\nsame semantics, can significantly simplify segmentation by casting unsupervised\nsemantic segmentation as an energy minimization problem. Under this paradigm,\nwe propose a novel approach called SmooSeg that harnesses self-supervised\nlearning methods to model the closeness relationships among observations as\nsmoothness signals. To effectively discover coherent semantic segments, we\nintroduce a novel smoothness loss that promotes piecewise smoothness within\nsegments while preserving discontinuities across different segments.\nAdditionally, to further enhance segmentation quality, we design an asymmetric\nteacher-student style predictor that generates smoothly updated pseudo labels,\nfacilitating an optimal fit between observations and labeling outputs. Thanks\nto the rich supervision cues of the smoothness prior, our SmooSeg significantly\noutperforms STEGO in terms of pixel accuracy on three datasets: COCOStuff\n(+14.9%), Cityscapes (+13.0%), and Potsdam-3 (+5.7%).",
+                    "authors": "Mengcheng Lan, Xinjiang Wang, Yiping Ke, Jiaxing Xu, Litong Feng, Wayne Zhang",
+                    "published": "2023-10-27",
+                    "updated": "2023-10-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.07898v5",
+                    "title": "ISLE: A Framework for Image Level Semantic Segmentation Ensemble",
+                    "abstract": "One key bottleneck of employing state-of-the-art semantic segmentation\nnetworks in the real world is the availability of training labels. Conventional\nsemantic segmentation networks require massive pixel-wise annotated labels to\nreach state-of-the-art prediction quality. Hence, several works focus on\nsemantic segmentation networks trained with only image-level annotations.\nHowever, when scrutinizing the results of state-of-the-art in more detail, we\nnotice that they are remarkably close to each other on average prediction\nquality, different approaches perform better in different classes while\nproviding low quality in others. To address this problem, we propose a novel\nframework, ISLE, which employs an ensemble of the \"pseudo-labels\" for a given\nset of different semantic segmentation techniques on a class-wise level.\nPseudo-labels are the pixel-wise predictions of the image-level semantic\nsegmentation frameworks used to train the final segmentation model. Our\npseudo-labels seamlessly combine the strong points of multiple segmentation\ntechniques approaches to reach superior prediction quality. We reach up to 2.4%\nimprovement over ISLE's individual components. An exhaustive analysis was\nperformed to demonstrate ISLE's effectiveness over state-of-the-art frameworks\nfor image-level semantic segmentation.",
+                    "authors": "Erik Ostrowski, Muhammad Shafique",
+                    "published": "2023-03-14",
+                    "updated": "2023-09-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1502.04983v1",
+                    "title": "Context Tricks for Cheap Semantic Segmentation",
+                    "abstract": "Accurate semantic labeling of image pixels is difficult because intra-class\nvariability is often greater than inter-class variability. In turn, fast\nsemantic segmentation is hard because accurate models are usually too\ncomplicated to also run quickly at test-time. Our experience with building and\nrunning semantic segmentation systems has also shown a reasonably obvious\nbottleneck on model complexity, imposed by small training datasets. We\ntherefore propose two simple complementary strategies that leverage context to\ngive better semantic segmentation, while scaling up or down to train on\ndifferent-sized datasets.\n  As easy modifications for existing semantic segmentation algorithms, we\nintroduce Decorrelated Semantic Texton Forests, and the Context Sensitive Image\nLevel Prior. The proposed modifications are tested using a Semantic Texton\nForest (STF) system, and the modifications are validated on two standard\nbenchmark datasets, MSRC-21 and PascalVOC-2010. In Python based comparisons,\nour system is insignificantly slower than STF at test-time, yet produces\nsuperior semantic segmentations overall, with just push-button training.",
+                    "authors": "Thanapong Intharah, Gabriel J. Brostow",
+                    "published": "2015-02-17",
+                    "updated": "2015-02-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2110.04487v1",
+                    "title": "Colour augmentation for improved semi-supervised semantic segmentation",
+                    "abstract": "Consistency regularization describes a class of approaches that have yielded\nstate-of-the-art results for semi-supervised classification. While\nsemi-supervised semantic segmentation proved to be more challenging, a number\nof successful approaches have been recently proposed. Recent work explored the\nchallenges involved in using consistency regularization for segmentation\nproblems. In their self-supervised work Chen et al. found that colour\naugmentation prevents a classification network from using image colour\nstatistics as a short-cut for self-supervised learning via instance\ndiscrimination. Drawing inspiration from this we find that a similar problem\nimpedes semi-supervised semantic segmentation and offer colour augmentation as\na solution, improving semi-supervised semantic segmentation performance on\nchallenging photographic imagery.",
+                    "authors": "Geoff French, Michal Mackiewicz",
+                    "published": "2021-10-09",
+                    "updated": "2021-10-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2102.12095v1",
+                    "title": "Synergy Between Semantic Segmentation and Image Denoising via Alternate Boosting",
+                    "abstract": "The capability of image semantic segmentation may be deteriorated due to\nnoisy input image, where image denoising prior to segmentation helps. Both\nimage denoising and semantic segmentation have been developed significantly\nwith the advance of deep learning. Thus, we are interested in the synergy\nbetween them by using a holistic deep model. We observe that not only denoising\nhelps combat the drop of segmentation accuracy due to noise, but also\npixel-wise semantic information boosts the capability of denoising. We then\npropose a boosting network to perform denoising and segmentation alternately.\nThe proposed network is composed of multiple segmentation and denoising blocks\n(SDBs), each of which estimates semantic map then uses the map to regularize\ndenoising. Experimental results show that the denoised image quality is\nimproved substantially and the segmentation accuracy is improved to close to\nthat of clean images. Our code and models will be made publicly available.",
+                    "authors": "Shunxin Xu, Ke Sun, Dong Liu, Zhiwei Xiong, Zheng-Jun Zha",
+                    "published": "2021-02-24",
+                    "updated": "2021-02-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.12417v2",
+                    "title": "SLAMs: Semantic Learning based Activation Map for Weakly Supervised Semantic Segmentation",
+                    "abstract": "Recent mainstream weakly-supervised semantic segmentation (WSSS) approaches\nmainly relies on image-level classification learning, which has limited\nrepresentation capacity. In this paper, we propose a novel semantic learning\nbased framework, named SLAMs (Semantic Learning based Activation Map), for\nWSSS.",
+                    "authors": "Junliang Chen, Xiaodong Zhao, Minmin Liu, Linlin Shen",
+                    "published": "2022-10-22",
+                    "updated": "2022-11-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.11506v1",
+                    "title": "LCCo: Lending CLIP to Co-Segmentation",
+                    "abstract": "This paper studies co-segmenting the common semantic object in a set of\nimages. Existing works either rely on carefully engineered networks to mine the\nimplicit semantic information in visual features or require extra data (i.e.,\nclassification labels) for training. In this paper, we leverage the contrastive\nlanguage-image pre-training framework (CLIP) for the task. With a backbone\nsegmentation network that independently processes each image from the set, we\nintroduce semantics from CLIP into the backbone features, refining them in a\ncoarse-to-fine manner with three key modules: i) an image set feature\ncorrespondence module, encoding global consistent semantic information of the\nimage set; ii) a CLIP interaction module, using CLIP-mined common semantics of\nthe image set to refine the backbone feature; iii) a CLIP regularization\nmodule, drawing CLIP towards this co-segmentation task, identifying the best\nCLIP semantic and using it to regularize the backbone feature. Experiments on\nfour standard co-segmentation benchmark datasets show that the performance of\nour method outperforms state-of-the-art methods.",
+                    "authors": "Xin Duan, Yan Yang, Liyuan Pan, Xiabi Liu",
+                    "published": "2023-08-22",
+                    "updated": "2023-08-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2404.04807v1",
+                    "title": "D2SL: Decouple Defogging and Semantic Learning for Foggy Domain-Adaptive Segmentation",
+                    "abstract": "We investigated domain adaptive semantic segmentation in foggy weather\nscenarios, which aims to enhance the utilization of unlabeled foggy data and\nimprove the model's adaptability to foggy conditions. Current methods rely on\nclear images as references, jointly learning defogging and segmentation for\nfoggy images. Despite making some progress, there are still two main drawbacks:\n(1) the coupling of segmentation and defogging feature representations,\nresulting in a decrease in semantic representation capability, and (2) the\nfailure to leverage real fog priors in unlabeled foggy data, leading to\ninsufficient model generalization ability. To address these issues, we propose\na novel training framework, Decouple Defogging and Semantic learning, called\nD2SL, aiming to alleviate the adverse impact of defogging tasks on the final\nsegmentation task. In this framework, we introduce a domain-consistent transfer\nstrategy to establish a connection between defogging and segmentation tasks.\nFurthermore, we design a real fog transfer strategy to improve defogging\neffects by fully leveraging the fog priors from real foggy images. Our approach\nenhances the semantic representations required for segmentation during the\ndefogging learning process and maximizes the representation capability of fog\ninvariance by effectively utilizing real fog data. Comprehensive experiments\nvalidate the effectiveness of the proposed method.",
+                    "authors": "Xuan Sun, Zhanfu An, Yuyu Liu",
+                    "published": "2024-04-07",
+                    "updated": "2024-04-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.MM"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.02095v1",
+                    "title": "Content-aware Token Sharing for Efficient Semantic Segmentation with Vision Transformers",
+                    "abstract": "This paper introduces Content-aware Token Sharing (CTS), a token reduction\napproach that improves the computational efficiency of semantic segmentation\nnetworks that use Vision Transformers (ViTs). Existing works have proposed\ntoken reduction approaches to improve the efficiency of ViT-based image\nclassification networks, but these methods are not directly applicable to\nsemantic segmentation, which we address in this work. We observe that, for\nsemantic segmentation, multiple image patches can share a token if they contain\nthe same semantic class, as they contain redundant information. Our approach\nleverages this by employing an efficient, class-agnostic policy network that\npredicts if image patches contain the same semantic class, and lets them share\na token if they do. With experiments, we explore the critical design choices of\nCTS and show its effectiveness on the ADE20K, Pascal Context and Cityscapes\ndatasets, various ViT backbones, and different segmentation decoders. With\nContent-aware Token Sharing, we are able to reduce the number of processed\ntokens by up to 44%, without diminishing the segmentation quality.",
+                    "authors": "Chenyang Lu, Daan de Geus, Gijs Dubbelman",
+                    "published": "2023-06-03",
+                    "updated": "2023-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1705.09052v3",
+                    "title": "Weakly Supervised Semantic Segmentation Based on Web Image Co-segmentation",
+                    "abstract": "Training a Fully Convolutional Network (FCN) for semantic segmentation\nrequires a large number of masks with pixel level labelling, which involves a\nlarge amount of human labour and time for annotation. In contrast, web images\nand their image-level labels are much easier and cheaper to obtain. In this\nwork, we propose a novel method for weakly supervised semantic segmentation\nwith only image-level labels. The method utilizes the internet to retrieve a\nlarge number of images and uses a large scale co-segmentation framework to\ngenerate masks for the retrieved images. We first retrieve images from search\nengines, e.g. Flickr and Google, using semantic class names as queries, e.g.\nclass names in the dataset PASCAL VOC 2012. We then use high quality masks\nproduced by co-segmentation on the retrieved images as well as the target\ndataset images with image level labels to train segmentation networks. We\nobtain an IoU score of 56.9 on test set of PASCAL VOC 2012, which reaches the\nstate-of-the-art performance.",
+                    "authors": "Tong Shen, Guosheng Lin, Lingqiao Liu, Chunhua Shen, Ian Reid",
+                    "published": "2017-05-25",
+                    "updated": "2017-08-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.10782v2",
+                    "title": "Three Ways to Improve Semantic Segmentation with Self-Supervised Depth Estimation",
+                    "abstract": "Training deep networks for semantic segmentation requires large amounts of\nlabeled training data, which presents a major challenge in practice, as\nlabeling segmentation masks is a highly labor-intensive process. To address\nthis issue, we present a framework for semi-supervised semantic segmentation,\nwhich is enhanced by self-supervised monocular depth estimation from unlabeled\nimage sequences. In particular, we propose three key contributions: (1) We\ntransfer knowledge from features learned during self-supervised depth\nestimation to semantic segmentation, (2) we implement a strong data\naugmentation by blending images and labels using the geometry of the scene, and\n(3) we utilize the depth feature diversity as well as the level of difficulty\nof learning depth in a student-teacher framework to select the most useful\nsamples to be annotated for semantic segmentation. We validate the proposed\nmodel on the Cityscapes dataset, where all three modules demonstrate\nsignificant performance gains, and we achieve state-of-the-art results for\nsemi-supervised semantic segmentation. The implementation is available at\nhttps://github.com/lhoyer/improving_segmentation_with_selfsupervised_depth.",
+                    "authors": "Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian K\u00f6ring, Suman Saha, Luc Van Gool",
+                    "published": "2020-12-19",
+                    "updated": "2021-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2307.13215v1",
+                    "title": "Image Segmentation Keras : Implementation of Segnet, FCN, UNet, PSPNet and other models in Keras",
+                    "abstract": "Semantic segmentation plays a vital role in computer vision tasks, enabling\nprecise pixel-level understanding of images. In this paper, we present a\ncomprehensive library for semantic segmentation, which contains implementations\nof popular segmentation models like SegNet, FCN, UNet, and PSPNet. We also\nevaluate and compare these models on several datasets, offering researchers and\npractitioners a powerful toolset for tackling diverse segmentation challenges.",
+                    "authors": "Divam Gupta",
+                    "published": "2023-07-25",
+                    "updated": "2023-07-25",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1809.10198v1",
+                    "title": "Recent progress in semantic image segmentation",
+                    "abstract": "Semantic image segmentation, which becomes one of the key applications in\nimage processing and computer vision domain, has been used in multiple domains\nsuch as medical area and intelligent transportation. Lots of benchmark datasets\nare released for researchers to verify their algorithms. Semantic segmentation\nhas been studied for many years. Since the emergence of Deep Neural Network\n(DNN), segmentation has made a tremendous progress. In this paper, we divide\nsemantic image segmentation methods into two categories: traditional and recent\nDNN method. Firstly, we briefly summarize the traditional method as well as\ndatasets released for segmentation, then we comprehensively investigate recent\nmethods based on DNN which are described in the eight aspects: fully\nconvolutional network, upsample ways, FCN joint with CRF methods, dilated\nconvolution approaches, progresses in backbone network, pyramid methods,\nMulti-level feature and multi-stage method, supervised, weakly-supervised and\nunsupervised methods. Finally, a conclusion in this area is drawn.",
+                    "authors": "Xiaolong Liu, Zhidong Deng, Yuhan Yang",
+                    "published": "2018-09-20",
+                    "updated": "2018-09-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2003.04404v1",
+                    "title": "FusionLane: Multi-Sensor Fusion for Lane Marking Semantic Segmentation Using Deep Neural Networks",
+                    "abstract": "It is a crucial step to achieve effective semantic segmentation of lane\nmarking during the construction of the lane level high-precision map. In recent\nyears, many image semantic segmentation methods have been proposed. These\nmethods mainly focus on the image from camera, due to the limitation of the\nsensor itself, the accurate three-dimensional spatial position of the lane\nmarking cannot be obtained, so the demand for the lane level high-precision map\nconstruction cannot be met. This paper proposes a lane marking semantic\nsegmentation method based on LIDAR and camera fusion deep neural network.\nDifferent from other methods, in order to obtain accurate position information\nof the segmentation results, the semantic segmentation object of this paper is\na bird's eye view converted from a LIDAR points cloud instead of an image\ncaptured by a camera. This method first uses the deeplabv3+ [\\ref{ref:1}]\nnetwork to segment the image captured by the camera, and the segmentation\nresult is merged with the point clouds collected by the LIDAR as the input of\nthe proposed network. In this neural network, we also add a long short-term\nmemory (LSTM) structure to assist the network for semantic segmentation of lane\nmarkings by using the the time series information. The experiments on more than\n14,000 image datasets which we have manually labeled and expanded have shown\nthe proposed method has better performance on the semantic segmentation of the\npoints cloud bird's eye view. Therefore, the automation of high-precision map\nconstruction can be significantly improved. Our code is available at\nhttps://github.com/rolandying/FusionLane.",
+                    "authors": "Ruochen Yin, Biao Yu, Huapeng Wu, Yutao Song, Runxin Niu",
+                    "published": "2020-03-09",
+                    "updated": "2020-03-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1809.10245v1",
+                    "title": "Cylindrical Transform: 3D Semantic Segmentation of Kidneys With Limited Annotated Images",
+                    "abstract": "In this paper, we propose a novel technique for sampling sequential images\nusing a cylindrical transform in a cylindrical coordinate system for kidney\nsemantic segmentation in abdominal computed tomography (CT). The images\ngenerated from a cylindrical transform augment a limited annotated set of\nimages in three dimensions. This approach enables us to train contemporary\nclassification deep convolutional neural networks (DCNNs) instead of fully\nconvolutional networks (FCNs) for semantic segmentation. Typical semantic\nsegmentation models segment a sequential set of images (e.g. CT or video) by\nsegmenting each image independently. However, the proposed method not only\nconsiders the spatial dependency in the x-y plane, but also the spatial\nsequential dependency along the z-axis. The results show that classification\nDCNNs, trained on cylindrical transformed images, can achieve a higher\nsegmentation performance value than FCNs using a limited number of annotated\nimages.",
+                    "authors": "Hojjat Salehinejad, Sumeya Naqvi, Errol Colak, Joseph Barfett, Shahrokh Valaee",
+                    "published": "2018-09-24",
+                    "updated": "2018-09-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.NE"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1707.02432v2",
+                    "title": "Deep Semantic Segmentation for Automated Driving: Taxonomy, Roadmap and Challenges",
+                    "abstract": "Semantic segmentation was seen as a challenging computer vision problem few\nyears ago. Due to recent advancements in deep learning, relatively accurate\nsolutions are now possible for its use in automated driving. In this paper, the\nsemantic segmentation problem is explored from the perspective of automated\ndriving. Most of the current semantic segmentation algorithms are designed for\ngeneric images and do not incorporate prior structure and end goal for\nautomated driving. First, the paper begins with a generic taxonomic survey of\nsemantic segmentation algorithms and then discusses how it fits in the context\nof automated driving. Second, the particular challenges of deploying it into a\nsafety system which needs high level of accuracy and robustness are listed.\nThird, different alternatives instead of using an independent semantic\nsegmentation module are explored. Finally, an empirical evaluation of various\nsemantic segmentation architectures was performed on CamVid dataset in terms of\naccuracy and speed. This paper is a preliminary shorter version of a more\ndetailed survey which is work in progress.",
+                    "authors": "Mennatullah Siam, Sara Elkerdawy, Martin Jagersand, Senthil Yogamani",
+                    "published": "2017-07-08",
+                    "updated": "2017-08-03",
+                    "primary_cat": "stat.ML",
+                    "cats": [
+                        "stat.ML",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2007.13138v1",
+                    "title": "Virtual Multi-view Fusion for 3D Semantic Segmentation",
+                    "abstract": "Semantic segmentation of 3D meshes is an important problem for 3D scene\nunderstanding. In this paper we revisit the classic multiview representation of\n3D meshes and study several techniques that make them effective for 3D semantic\nsegmentation of meshes. Given a 3D mesh reconstructed from RGBD sensors, our\nmethod effectively chooses different virtual views of the 3D mesh and renders\nmultiple 2D channels for training an effective 2D semantic segmentation model.\nFeatures from multiple per view predictions are finally fused on 3D mesh\nvertices to predict mesh semantic segmentation labels. Using the large scale\nindoor 3D semantic segmentation benchmark of ScanNet, we show that our virtual\nviews enable more effective training of 2D semantic segmentation networks than\nprevious multiview approaches. When the 2D per pixel predictions are aggregated\non 3D surfaces, our virtual multiview fusion method is able to achieve\nsignificantly better 3D semantic segmentation results compared to all prior\nmultiview approaches and competitive with recent 3D convolution approaches.",
+                    "authors": "Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, Caroline Pantofaru",
+                    "published": "2020-07-26",
+                    "updated": "2020-07-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.17083v1",
+                    "title": "SegRCDB: Semantic Segmentation via Formula-Driven Supervised Learning",
+                    "abstract": "Pre-training is a strong strategy for enhancing visual models to efficiently\ntrain them with a limited number of labeled images. In semantic segmentation,\ncreating annotation masks requires an intensive amount of labor and time, and\ntherefore, a large-scale pre-training dataset with semantic labels is quite\ndifficult to construct. Moreover, what matters in semantic segmentation\npre-training has not been fully investigated. In this paper, we propose the\nSegmentation Radial Contour DataBase (SegRCDB), which for the first time\napplies formula-driven supervised learning for semantic segmentation. SegRCDB\nenables pre-training for semantic segmentation without real images or any\nmanual semantic labels. SegRCDB is based on insights about what is important in\npre-training for semantic segmentation and allows efficient pre-training.\nPre-training with SegRCDB achieved higher mIoU than the pre-training with\nCOCO-Stuff for fine-tuning on ADE-20k and Cityscapes with the same number of\ntraining images. SegRCDB has a high potential to contribute to semantic\nsegmentation pre-training and investigation by enabling the creation of large\ndatasets without manual annotation. The SegRCDB dataset will be released under\na license that allows research and commercial use. Code is available at:\nhttps://github.com/dahlian00/SegRCDB",
+                    "authors": "Risa Shinoda, Ryo Hayamizu, Kodai Nakashima, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka",
+                    "published": "2023-09-29",
+                    "updated": "2023-09-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.11316v2",
+                    "title": "Generative Semantic Segmentation",
+                    "abstract": "We present Generative Semantic Segmentation (GSS), a generative learning\napproach for semantic segmentation. Uniquely, we cast semantic segmentation as\nan image-conditioned mask generation problem. This is achieved by replacing the\nconventional per-pixel discriminative learning with a latent prior learning\nprocess. Specifically, we model the variational posterior distribution of\nlatent variables given the segmentation mask. To that end, the segmentation\nmask is expressed with a special type of image (dubbed as maskige). This\nposterior distribution allows to generate segmentation masks unconditionally.\nTo achieve semantic segmentation on a given image, we further introduce a\nconditioning network. It is optimized by minimizing the divergence between the\nposterior distribution of maskige (i.e., segmentation masks) and the latent\nprior distribution of input training images. Extensive experiments on standard\nbenchmarks show that our GSS can perform competitively to prior art\nalternatives in the standard semantic segmentation setting, whilst achieving a\nnew state of the art in the more challenging cross-domain setting.",
+                    "authors": "Jiaqi Chen, Jiachen Lu, Xiatian Zhu, Li Zhang",
+                    "published": "2023-03-20",
+                    "updated": "2023-08-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2403.01482v4",
+                    "title": "EAGLE: Eigen Aggregation Learning for Object-Centric Unsupervised Semantic Segmentation",
+                    "abstract": "Semantic segmentation has innately relied on extensive pixel-level annotated\ndata, leading to the emergence of unsupervised methodologies. Among them,\nleveraging self-supervised Vision Transformers for unsupervised semantic\nsegmentation (USS) has been making steady progress with expressive deep\nfeatures. Yet, for semantically segmenting images with complex objects, a\npredominant challenge remains: the lack of explicit object-level semantic\nencoding in patch-level features. This technical limitation often leads to\ninadequate segmentation of complex objects with diverse structures. To address\nthis gap, we present a novel approach, EAGLE, which emphasizes object-centric\nrepresentation learning for unsupervised semantic segmentation. Specifically,\nwe introduce EiCue, a spectral technique providing semantic and structural cues\nthrough an eigenbasis derived from the semantic similarity matrix of deep image\nfeatures and color affinity from an image. Further, by incorporating our\nobject-centric contrastive loss with EiCue, we guide our model to learn\nobject-level representations with intra- and inter-image object-feature\nconsistency, thereby enhancing semantic accuracy. Extensive experiments on\nCOCO-Stuff, Cityscapes, and Potsdam-3 datasets demonstrate the state-of-the-art\nUSS results of EAGLE with accurate and consistent semantic segmentation across\ncomplex scenes.",
+                    "authors": "Chanyoung Kim, Woojung Han, Dayun Ju, Seong Jae Hwang",
+                    "published": "2024-03-03",
+                    "updated": "2024-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.04618v1",
+                    "title": "Multi-Scale Semantic Segmentation with Modified MBConv Blocks",
+                    "abstract": "Recently, MBConv blocks, initially designed for efficiency in\nresource-limited settings and later adapted for cutting-edge image\nclassification performances, have demonstrated significant potential in image\nclassification tasks. Despite their success, their application in semantic\nsegmentation has remained relatively unexplored. This paper introduces a novel\nadaptation of MBConv blocks specifically tailored for semantic segmentation.\nOur modification stems from the insight that semantic segmentation requires the\nextraction of more detailed spatial information than image classification. We\nargue that to effectively perform multi-scale semantic segmentation, each\nbranch of a U-Net architecture, regardless of its resolution, should possess\nequivalent segmentation capabilities. By implementing these changes, our\napproach achieves impressive mean Intersection over Union (IoU) scores of 84.5%\nand 84.0% on the Cityscapes test and validation datasets, respectively,\ndemonstrating the efficacy of our proposed modifications in enhancing semantic\nsegmentation performance.",
+                    "authors": "Xi Chen, Yang Cai, Yuan Wu, Bo Xiong, Taesung Park",
+                    "published": "2024-02-07",
+                    "updated": "2024-02-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.13978v3",
+                    "title": "Personalized Image Semantic Segmentation",
+                    "abstract": "Semantic segmentation models trained on public datasets have achieved great\nsuccess in recent years. However, these models didn't consider the\npersonalization issue of segmentation though it is important in practice. In\nthis paper, we address the problem of personalized image segmentation. The\nobjective is to generate more accurate segmentation results on unlabeled\npersonalized images by investigating the data's personalized traits. To open up\nfuture research in this area, we collect a large dataset containing various\nusers' personalized images called PIS (Personalized Image Semantic\nSegmentation). We also survey some recent researches related to this problem\nand report their performance on our dataset. Furthermore, by observing the\ncorrelation among a user's personalized images, we propose a baseline method\nthat incorporates the inter-image context when segmenting certain images.\nExtensive experiments show that our method outperforms the existing methods on\nthe proposed dataset. The code and the PIS dataset will be made publicly\navailable.",
+                    "authors": "Yu Zhang, Chang-Bin Zhang, Peng-Tao Jiang, Ming-Ming Cheng, Feng Mao",
+                    "published": "2021-07-24",
+                    "updated": "2021-09-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2009.12232v4",
+                    "title": "From Pixel to Patch: Synthesize Context-aware Features for Zero-shot Semantic Segmentation",
+                    "abstract": "Zero-shot learning has been actively studied for image classification task to\nrelieve the burden of annotating image labels. Interestingly, semantic\nsegmentation task requires more labor-intensive pixel-wise annotation, but\nzero-shot semantic segmentation has only attracted limited research interest.\nThus, we focus on zero-shot semantic segmentation, which aims to segment unseen\nobjects with only category-level semantic representations provided for unseen\ncategories. In this paper, we propose a novel Context-aware feature Generation\nNetwork (CaGNet), which can synthesize context-aware pixel-wise visual features\nfor unseen categories based on category-level semantic representations and\npixel-wise contextual information. The synthesized features are used to\nfinetune the classifier to enable segmenting unseen objects. Furthermore, we\nextend pixel-wise feature generation and finetuning to patch-wise feature\ngeneration and finetuning, which additionally considers inter-pixel\nrelationship. Experimental results on Pascal-VOC, Pascal-Context, and\nCOCO-stuff show that our method significantly outperforms the existing\nzero-shot semantic segmentation methods. Code is available at\nhttps://github.com/bcmi/CaGNetv2-Zero-Shot-Semantic-Segmentation.",
+                    "authors": "Zhangxuan Gu, Siyuan Zhou, Li Niu, Zihan Zhao, Liqing Zhang",
+                    "published": "2020-09-25",
+                    "updated": "2022-01-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                }
+            ],
+            [
+                {
+                    "url": "http://arxiv.org/abs/2312.09128v1",
+                    "title": "Tokenize Anything via Prompting",
+                    "abstract": "We present a unified, promptable model capable of simultaneously segmenting,\nrecognizing, and captioning anything. Unlike SAM, we aim to build a versatile\nregion representation in the wild via visual prompting. To achieve this, we\ntrain a generalizable model with massive segmentation masks, e.g., SA-1B masks,\nand semantic priors from a pre-trained CLIP model with 5 billion parameters.\nSpecifically, we construct a promptable image decoder by adding a semantic\ntoken to each mask token. The semantic token is responsible for learning the\nsemantic priors in a predefined concept space. Through joint optimization of\nsegmentation on mask tokens and concept prediction on semantic tokens, our\nmodel exhibits strong regional recognition and localization capabilities. For\nexample, an additional 38M-parameter causal text decoder trained from scratch\nsets a new record with a CIDEr score of 150.7 on the Visual Genome region\ncaptioning task. We believe this model can be a versatile region-level image\ntokenizer, capable of encoding general-purpose region context for a broad range\nof perception tasks. Code and models are available at\nhttps://github.com/baaivision/tokenize-anything.",
+                    "authors": "Ting Pan, Lulu Tang, Xinlong Wang, Shiguang Shan",
+                    "published": "2023-12-14",
+                    "updated": "2023-12-14",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Original Paper",
+                    "paper_cat": "Semantic AND Segmentation AND Image",
+                    "gt": "2.1. Vision Foundation Models Vision foundation models aim to achieve strong zero and few-shot generalization capabilities across a broad range of vision tasks. Starting with CLIP [45], which simultaneously trains image and text encoders with massive imagetext pairs to align two modalities, numerous efforts have emerged to train a general-purpose vision-language representation at scale [21, 31, 54]. In addition, some works aim to build vision generalist models [23, 59, 62, 63, 80]. For example, SAM [23] introduces a large-scale dataset and trains a model for promptable segmentation. Taking user interactions as prompts, SAM demonstrates strong zero-shot performance in general segmentation tasks. Concurrent to SAM, SegGPT [63] unifies a variety of segmentation tasks into one in-context segmentation problem. After training, SegGPT showcases the capability to execute arbitrary segmentation tasks through in-context inference. Some other works seek to build a generalist model by leveraging multimodality datasets [1, 39, 55]. In this work, we aim to build a vision foundation model that serves as a versatile regionlevel image tokenizer, capable of encoding general-purpose region context for a broad range of perception tasks. 2.2. Open-Vocabulary Segmentation Unlike previous instance segmentation and semantic segmentation models [3, 5, 16, 36, 60, 61, 66] that work in a limited vocabulary, open-vocabulary segmentation (OVS) aims to classify regions that go beyond the closedvocabulary used for training [8, 9, 13, 19, 27, 32, 43, 68, 69, 73, 78, 79]. Numerous efforts focus on leveraging pretrained Vision-Language models (VLMs) like CLIP [45] and center on designing specific alignment techniques to effectively integrate VLM knowledge into existing segmentation models [8, 9, 27, 78]. For example, LSeg [27] embeds text and pixel embeddings into a common feature space, assigning label to each pixel. MaskCLIP [9] builds a two-stage model to seamlessly integrate with CLIP visual encoder. ZegFormer [8] decouples problem into a classagnostic grouping task and a region-level classification task to utilize VLM. By leveraging the caption data, some studies align visual features with texts in a weakly supervised manner [13, 19, 32, 65, 68]. For instance, GroupViT [68] is trained on image-caption pairs without pixel-level annotations, directly grouping masks based on text supervision. OVSeg [32] fine-tunes CLIP on masked images with pseudo labels generated from the nouns in image captions. CGG [65], on the other hand, combines grounding and generation losses to thoroughly explore the knowledge from image captions. Additionally, other studies [43, 73, 79] jointly learn multiple tasks within a single network or investigate text-to-image diffusion models [22, 69]. Our work 2 Figure 2. TAP accepts flexible prompts and outputs segmentation, category, and caption simultaneously. 3 aligns with CLIP-based approaches but differs from twostage models, which typically rely on an image-level CLIP to classify masks. Instead, our approach focuses on developing a single model with region-level semantic awareness. 2.3. Zero-shot Region Understanding Previous works focus on extending the open-vocabulary capabilities of VLMs to object detection tasks [14, 25, 57, 72, 76]. Recent studies [56, 70] aim to merge CLIP\u2019s proficiency in open-vocabulary classification with SAM\u2019s capability in segmentation. For instance, SAM-CLIP [56] distills knowledge from both SAM and CLIP by retraining the visual encoder with a portion of data sampled for two teachers, retaining the original strengths of both CLIP and SAM. RegionSpot [70] unifies prompting by adding an adapter trained on detection datasets, enabling SAM\u2019s mask tokens to interact with CLIP\u2019s features derived from masked image segments. Some works [28, 59, 80] attempt to construct unified models capable of recognizing objects in arbitrary regions. SEEM [80] was built upon X-Decoder [79], excelling in handling various types of prompts, including clicks, bounding boxes, scribbles, text, and referring image segments. Following SAM [23], ASM [59] created a new dataset (AS-1B) for SA-1B( [23]), constructing rich annotations of semantic tags, question-answering pairs, and detailed captions. Leveraging this dataset, they develop a new model, ASM, for panoptic visual recognition. Unlike these models relying on handcrafted multi-modal datasets, we fully leverage extensive segmentation masks from SA-1B and semantic priors from a high-performing CLIP model, aiming to develop a promptable image tokenizer that can understand semantic context for any given region.",
+                    "pre_questions": [],
+                    "main_content": "Introduction A key objective of visual perception is to efficiently localize and recognize arbitrary regions of interest. It demands a single vision model that is capable of understanding the region context and simultaneously executing perception tasks such as segmentation, recognition, and captioning. However, existing models often focus on either localizing class-agnostic masks, e.g., SAM [23], or extracting only visual semantics, e.g., CLIP [45] and its regionlevel variants. Specifically, SAM develops a segmentation foundation model that can segment anything via prompting, enabling strong generalization in pixel-wise localization tasks. On the other hand, CLIP trains a recognition foundation model via contrastive learning on web-scale image-text pairs, demonstrating powerful zero-shot abilities in recognition tasks. Accordingly, learning semantic priors from a CLIP model within SAM\u2019s architecture offers a promising pathway towards comprehensive visual perception. In this work, we aim to build a promptable model and pre-train it on a large-scale dataset, using a task that enables powerful generalization in both localization and recognition. We start by introducing a promptable tokenization task that is general enough to serve as a powerful pre-training objective, while facilitating a broad range of downstream applications. This task requires a model capable of abstract1 arXiv:2312.09128v1  [cs.CV]  14 Dec 2023 ing the general-purpose representation, e.g., mask tokens and semantic tokens, given flexible prompting that cues the region of interest. The extracted region representation can then be directly decoded into corresponding task output for general-purpose visual perception tasks. Training such a highly performant and generalizable foundation model necessitates a diverse, large-scale dataset. Nevertheless, there is currently no web-scale data source available for simultaneous segmentation and recognition. SA-1B [23] constructs 1.1B high-quality mask annotations on 11M images for training a segmentation foundation model, e.g., SAM. On the other hand, LAION-2B [48] collects 2B image-text pairs from the web, enabling the training of generalizable recognition models, e.g., CLIP. To address this challenge posed by the lack of aligned data, we introduce the SemanticSA-1B dataset (see Fig. 1c). This dataset implicitly integrates web-scale semantics from LAION-2B into SA-1B. Specifically, for each segmented region in SA-1B, we extract the distribution over a concept vocabulary predicted by a powerful CLIP model with 5B parameters, trained on massive LAION image-text pairs. With SemanticSA-1B dataset, we train a unified and generalizable model capable of simultaneously segmenting, recognizing, and captioning anything. This is achieved by merging CLIP\u2019s capabilities within SAM\u2019s architecture, leveraging the web-scale semantics and segmentation masks. We refer to this model as TAP, short for Tokenize Anything via Prompting, as illustrated in Fig. 1b. Specifically, given an image and a visual prompt, TAP tokenizes the region of interest into a mask token and a semantic token. The mask token queries for pixel-wise segmentation, similar to SAM, while the semantic token is responsible for region-level semantic prediction. Our TAP model is trained end-to-end with joint masks and semantics from the beginning. By leveraging the semantic token, we can concurrently address the open-vocabulary classification task with an MLP head, and the promptable captioning task with a lightweight text decoder using an auto-regressive process. We extensively evaluated TAP model and its components. TAP demonstrates strong zero-shot performance in instance classification, e.g., 59.0 AP on the challenging LVIS benchmark, while maintaining competitive zero-shot segmentation performance, e.g., 42.6 vs. 43.1 AP for TAP and SAM. Notably, we set a new record with a CIDEr score of 150.7 in the region caption task on Visual Genome [24], using significantly fewer parameters compared to the prior works. Our findings indicate that the tokenized region features are generalizable for both segmentation and classification tasks, and can even directly prompt causal language modeling. Above all, we believe TAP model can be a versatile regionlevel image tokenizer, capable of encoding regional context for a broad range of vision and language tasks (see Fig. 2). We introduce a promptable model that efficiently enables segmenting, recognizing, and captioning anything at once. We achieve this by predicting CLIP priors within a promptable tokenizer (Sec. 3.1) and extending the model scope to encompass generative abilities for captioning (Sec. 3.2). 3.1. Promptable Tokenization Our primary focus is to align vision and language within a promptable segmentation model, SAM, to enhance the model with region-level semantic awareness. Conventional vision-language alignment approaches rely on image-text pairs [4, 48, 51], limiting the fine-grained region understanding. In contrast to prior methods [42, 59, 80] reliant on well-collected or approximated region-text data, our approach aligns masks with language using CLIP and exhaustive segmentation data sourced from SA-1B. Since SA-1B is a class-agnostic dataset, we employ the off-the-shelf CLIP embeddings in a human-curated concept space, and align the distribution of concept vocabulary between SAM\u2019s prediction and CLIP\u2019s projection. To this end, we pre-train an all-in-one image encoder-decoder on two sub-tasks: (i) promptable segmentation and (ii) concept prediction. An overview of our method is illustrated in Fig. 3. Pre-processing. Different from prior methods [23, 80], we exclude text prompts due to their ambiguity compared to point prompts, particularly in responding to 70% of smallpart masks (area \u22641002) in SA-1B. Prior studies [14, 76] leveraging off-the-shelf CLIP alignment often extract image embeddings using box proposals from a pre-trained region proposal network [47]. In contrast, SA-1B dataset offers a high-quality mask for each object within an image. It allows us to naturally compute image embeddings based on these ground-truth masks, avoiding dataset-specific annotation biases or box prediction errors. Specifically, we employ the highly performant open-source CLIP model, 5Bparameter EVA-CLIP [54], to compute the image embeddings from the masked image crops and store them locally, thus leading to the final SemanticSA-1B dataset. Promptable segmentation. The mask decoder within SAM adopts an architecture derived from Mask2Former [6], incorporating deformable masked attention in response to input prompts for interactive segmentation. We thus consider promptable segmentation as a necessary prelude to unsealing the semantic capabilities. Following SAM, our model defaults to predicting four masks for each prompt, yet a routing strategy selects one to resolve the ambiguity. Consequently, our image decoder produces 9 output tokens: 4 semantic tokens, 4 mask tokens, and an IoU token. To improve training efficiency on the large-scale SA-1B dataset, we implement a two-stage sampling strategy with maximal 9 prompt points, as it is performed within 11 interactive stages in the original SAM. In the first stage, we sample a box or point with equal probability from the ground-truth mask. In the subsequent stage, performed across 256 GPUs, we uniformly sample 1 to 8 points from the error region between predicted and ground-truth masks. To enable sketch or mask as the prior prompt, an aspect unexplored in SAM, we introduce a non-interactive sampling method with a 50% probability in the second stage. This sampling uniformly fetches 1 to 9 points from the ground-truth mask, providing a wider prompt space. During inference, 9 points are selected from the linear space of flatted 2D coordinates of a mask or sketch to ensure determinacy. As for mask supervision, a linear combination of focal loss [34] and dice loss [41] is employed at a 20:1 ratio, following SAM [23]. Concept prediction. To enhance our model with semantic awareness, we propose predicting region concepts using 4 Prompt Encoder Image Encoder Decoder point [S] [M] mask token semantic token the rabbit is white and brown rabbit box sketch mask embedding classify segment caption Promptable Tokenization Figure 3. Overview of TAP. a) Building upon SAM\u2019s architecture, we enhance the mask decoder to a generic image decoder, adding an additional semantic token [S] to each predicted mask. b) Our model is pre-trained on SemanticSA-1B via promptable tokenization, jointly optimized for promptable segmentation and concept prediction. c) Subsequently, the pre-trained model is fine-tuned for region captioning. the semantic token. Concretely, we employ the semantic token to obtain a 1024-dimension visual embedding through a 3-layer MLP (256\u21921024\u21921024). This visual embedding is further projected to the 2560-dimension distribution logits. Subsequently, we optimize KL divergence loss between the predicted distribution and the target distribution obtained from a CLIP model. This approach effectively alleviates performance degeneration caused by pairwise joint concepts. For instance, the concept of a bulldog is a subset of the dog category, thus it should not deviate too far from related concepts like dog or cat in the representation space. More importantly, the image-text distribution provides maximum information for supervision, preventing foundation models from learning the biases of hard label. Zero-shot transfer. After pre-training, our model can conduct open-vocabulary classification for segmentation prompts. Given a visual prompt, our image decoder produces 4 masks and 9 tokens. The final mask and the associated semantic token are selected using a heuristic routing strategy. Specifically, we opt for the first mask for boundary boxes, and select the top-ranked remainder for loose points, akin to a simplified implementation of the mixtureof-experts (MoE) technique [20]. The final semantic token is utilized for zero-shot instance classification on a datasetspecific concept vocabulary (e.g., COCO and LVIS). 3.2. Promptable Captioning We draw inspiration from recent advances in large language models, where the next token prediction is used to substitute for human-crafted prediction tasks. In this section, we introduce a text generation paradigm aimed at unleashing the potential of promptable semantic tokens. Task. Many prior works fine-tune the pre-trained model with a pseudo open-vocabulary classifier on a large vocabulary dataset. However, this task is far behind the finetuning on conversational context in NLP, which encodes unbounded human knowledge from the open-world. In our ...... Causal Text Decoder [BOS] <rabbit> <the> [EOS] <the> [EOS] <rabbit> [PAD] the rabbit is white and brown semantic token <brown> ...... <is> [S] Figure 4. Promptable captioning. The semantic token from image decoder is directly used to prompt causal text generation. effort to advance visual foundation models, we develop a generative vision-language model through causal language modeling. Specifically, we adopt a causal Transformer prompted with a semantic token from the image decoder to generate the region caption. Different from the prior method [58] that roughly couples three frozen models, our model is able to challenge this task end-to-end. An overview of our text generation architecture is depicted in Fig. 4. Visual encoder. Given semantic tokens generated by the promptable tokenizer (refer to Fig. 3), we solely apply a linear projection on these semantic tokens to align their dimensions with text embeddings (see Fig. 4). This visual encoder, comprising a promptable tokenizer and a linear projector, exhibits notable efficiency compared to previous methods [64, 74] that involve parameterizing region-ofinterest features. Moreover, it perceptually encodes a segmented region for comprehensive visual understanding. Text decoder. We predict the tokenized region captions using byte-pair encoding [49] with a 32k tokens vocabulary. For text decoding, we employ an 8-layer standard Transformer with an embedding dimension of 512 to accommodate the brief description (the maximal context length is 40). This 25M-parameter lightweight text decoder refers to T5small [46] and is sufficient to perform mask-to-text translation if prompted with mighty tokens. 5 Causal modeling. We place the semantic token at the leading position of a sequence, followed by a [BOS] token, and supervise the next token prediction using cross-entropy loss. We employ rotary embedding [53] to integrate the positional encoding for multi-modal sequences. Caption inference. For caption generation, we iteratively generate up to 40 tokens with the maximum probability for each mask. To speed up attention computation, we follow a standard practice for auto-regressive that caches the key and value pairs. The final generation is selected from the multiple outputs corresponding to each prompt, employing the same routing strategy as described at the end of Sec 3.1. 4. Experiments 4.1. Experiments Setup Pre-training. We pre-train TAP models on full SA-1B, which comprises 11 M high-resolution images with around 100 regions per image, totaling 1.1B segmentation masks. Since there are no semantic annotations in SA-1B dataset, inspired by [42, 71, 76], we utilize EVA-CLIP [54] to achieve text embeddings on a merged label space from COCO [33], ADE20K [77], LVIS [15], Objects365 [50], Visual Genome [24] and OpenImagesV4 [26] datasets. This results in a concept list spanning 2560 categories to cover both things and stuff necessary for panoptic understanding. Evaluation. We assess zero-shot instance segmentation performance on COCO and LVIS. For zero-shot instance classification, we prioritize LVIS due to its broader range of 1203 generalized categories compared to COCO, which covers 80 common categories, diverging from the openworld assumption. Moving to region-level captioning task, we freeze image encoder-decoder and train a text decoder on Visual Genome (VG) [24] v1.0 train set. We report the following metrics on VG test set and RefCOCOg [40] validation set: BLEU@4, METEOR, ROUGE, and CIDEr. Modeling. We maintain SAM architecture, comprising an image encoder, a prompt encoder, and a mask decoder for segmentation, while introducing two key modifications. Firstly, we substitute the computationally intensive global attention in the image encoder with convolutional crosswindow blocks [29]. Secondly, we upgrade the mask decoder to a generic image decoder by adding one semantic token for each predicted mask. Regarding the text decoder, we incorporate a linear projector and a causal Transformer. Implementation details. In all pre-training and finetuning experiments, we utilize the AdamW [38] optimizer (\u03b21 = 0.9, \u03b22 = 0.999) with a base learning rate of 1e-3. A cosine learning rate schedule [37] is implemented, and the final learning rate is decayed to 1% of the base value. For pre-training on SemanticSA-1B, scale jitter [12] is applied with a range of [0.5, 2.0] for 90k iterations (\u223c4 epochs), with a batch size of 256 across 256 GPUs. We fine-tune VG without data augmentation for 30k iterations (\u223c25 epochs), with a batch size of 64 across 8 GPUs. Additional hyperparameters include a weight decay of 0.1, a drop path [18] rate of 0.1/0.2 for ViT-B/ViT-L, and a dropout [52] rate of 0.1/0.3 for the image/text decoder. The image encoder is initialized from MAE [17] pre-trained weights, while all other layers are from scratch. For all experiments, we adopt up to 64 sampled prompts per GPU at each sampling stage. 4.2. Main Results Zero-shot instance segmentation. We evaluate our model in zero-shot instance segmentation, a task at which the original SAM excels. Following a common practice [23, 70], we first obtain detection bounding boxes from a ViTDet-H model [29]. Subsequently, we utilize these boxes to prompt the image decoder and compare the bare segmentation performance (i.e., using the box category) on COCO and LVIS. For a fair comparison, we report results from both the original SAM and our reproduced version (denoted as our impl.). As depicted in Tab. 1, our model achieves comparable segmentation results with original SAM across different model scales. This demonstrates that additional concept prediction task dose not compromise SAM\u2019s original capability. Moreover, it also suggests that segmentation, being an elementary and geometric task, may not fully exploit the capacity of foundation models. Zero-shot instance classification. We prompt the image decoder with ground-truth (GT) boxes and evaluate the bare recognition capability (i.e., using box coordinates) on LVIS. With GT boxes as visual prompts, our model substantially surpasses RegionCLIP [76] and RegionSpot [70], which are trained on limited image regions. These promising results suggest that employing concept prediction on exhaustive image regions can effectively empower SAM with semantic awareness. As shown in Tab. 2, the highly performant EVACLIP outperforms all other methods in zero-shot evaluation, achieving an impressive rare AP. Nonetheless, deploying a standalone CLIP model to compute massive image crops is impractical for real-time vision systems. We demonstrate that large CLIP models can be integrated into a promptable image encoder-decoder with reasonable performance. Region-level captioning. We assess our model on Visual Genome [24] and RefCOCOg [40]. Initially, we utilize GT boxes to prompt the image decoder, and subsequently, we employ the resulting semantic tokens to prompt the text decoder. The evaluation results are presented in Tab. 3. Sur6 Tab 1. Zero-shot instance segmentation on COCO [33] and LVIS [15]. Proposals obtained from ViTDet-H boxes. COCO LVIS Model AP APS APM APL AP APS APM APL APR APC APF ViTDet-H [29] 51.0 32.0 54.3 68.9 46.6 35.0 58.0 66.3 35.9 46.8 51.1 SAM-B [23] 41.1 28.3 45.6 53.7 40.8 30.1 53.0 58.5 32.6 41.9 43.3 SAM-L [23] 45.5 30.2 50.1 60.4 43.8 31.9 56.7 64.2 34.3 44.7 46.9 SAM-H [23] 46.5 30.8 51.0 61.7 44.7 32.5 57.6 65.5 34.6 45.5 47.8 SAM-B (our impl.) 45.1 28.1 50.1 61.4 42.1 29.3 54.9 64.2 33.2 43.2 44.7 SAM-L (our impl.) 45.8 29.0 50.7 62.2 43.1 30.2 56.0 65.3 33.4 44.2 46.1 TAP-B 44.9 28.0 49.9 60.9 41.7 28.9 54.5 63.7 33.1 42.7 44.2 TAP-L 45.6 29.0 50.6 61.4 42.6 29.8 55.5 64.8 33.3 43.6 45.5 Tab 2. Zero-shot instance classification on LVIS [15]. All entries are evaluated using GT boxes for a fair comparison. Model Training data AP APR APC APF Supervised methods (baseline): ViTDet-B [29] LVIS 61.9 40.8 58.5 74.9 ViTDet-L [29] LVIS 68.8 51.5 65.6 79.9 ViTDet-H [29] LVIS 69.3 52.8 66.3 79.8 Image-level CLIP (evaluation with image crops): CLIP-L [45] WIT-400M 48.8 52.8 50.0 45.6 EVA-CLIP-E [54] LAION-2B 64.3 72.4 65.3 59.7 Region-level CLIP: RegionCLIP-R50x4 [76] CC-3M 50.7 50.1 50.1 51.7 RegionSpot-BL [70] O365,OI,V3D 56.6 50.6 50.2 68.8 TAP-B SemanticSA-1B 56.4 55.6 55.6 57.7 TAP-L SemanticSA-1B 59.0 60.5 58.7 58.7 prisingly, our model achieves a new record with a 150.7 CIDEr score on Visual Genome, even with a frozen image encoder-decoder due to the absence of masks and labels. It is noteworthy that the concurrent work ASM [59] is trained on a multi-modal dataset, including a vast repository of region-text pairs. Semantic knowledge of our model is learnt from a CLIP model. Another concurrent work, SCA [67], additionally trains a 12-layer image decoder to learn 14 tokens for captioning task. These results suggest that our semantic token effectively encodes sufficient region-level information for captioning, supporting our earlier claim that TAP can function as a location-aware image tokenizer. 4.3. Ablation Study Semantic prediction. Ablation studies on pretext tasks for pre-training are presented in Tabs. 4 and 5, where \u2018Mask\u2019, \u2018Feature\u2019, and \u2018Concept\u2019 represent pre-training with segmentation, feature prediction, and concept prediction, respectively. As observed in Tab. 5, caption metrics are remarkably low when pre-trained with \u2018Mask\u2019 alone (Model A). When combined with semantic prediction (Model B/C), either feature or concept prediction, the caption performance sees a significant improvement. Despite showing semantic awareness, feature prediction is inferior to concept prediction in both classification and region-level captioning tasks. These findings indicate that an orthogonal space (e.g. concept space) is crucial for acquiring CLIP priors. We conjecture that concept prediction has efficiently facilitated the model in learning sufficient negative retrievals from CLIP. Semantic tokenization. To assess the effectiveness of semantic tokens, we conduct four experiments. Firstly, we pre-train our model using the approach listed in \u2018Pre-train\u2019 column. Subsequently, we fine-tune the text decoder using the items outlined in \u2018TextPrompt\u2019, generated from the frozen pre-trained model. Model A serves as our baseline, pre-trained with only \u2018Mask\u2019. Here, mask tokens are directly used for region-level captioning task, akin to using the original SAM\u2019s output to train the text decoder. Model D is our default model, jointly optimized with promptable segmentation and concept prediction. Semantic tokens are used to prompt the text decoder. As demonstrated in Tab. 5, semantic tokens consistently outperform mask tokens in captioning task, while achieving comparable AP in segmentation task. Eventually, semantic tokenization via concept prediction proves to be the most effective. This suggests that semantic tokenization significantly unlocks the potential of foundation model, facilitating more perception tasks. Scaling pre-training. We ablate the pre-training configurations from model, data, and training schedule. Default settings, highlighted in gray, involve pre-training with a ViT-L image encoder, 50% of the full SA-1B dataset, and 90k iterations. As observed in Tab. 6a, a longer pre-training sched7 Figure 5. Visualization of crowd understanding. Best viewed in color with zoom. 8 Figure 6. Visualization of understanding open-world knowledge. 9 Tab 3. Region captioning on Visual Genome [24] and RefCOCOg [40]. Proposals obtained from GT boxes. Visual Genome RefCOCOg Method VisualEncoder TextDecoder TextPrompt METEOR CIDEr METEOR CIDEr GRiT [64] ViT-B Small-43M BoxFeature 17.1 142.0 15.2 71.6 GPT4ROI [74] CLIP-H Vicuna-7B [75] BoxFeature 17.4 145.2 GPT4ROI [74] CLIP-H Vicuna-13B [75] BoxFeature 17.6 146.8 ASM [59] ViT-g Husky-7B [35] BoxFeature 18.0 145.1 20.8 103.0 SCA [67] SAM-H GPT2-774M [44] (8 Caption + 6 Task) Tokens 17.4 148.8 15.3 70.5 SCA [67] SAM-H Llama-3B [11] (8 Caption + 6 Task) Tokens 17.4 149.8 15.6 74.0 TAP ViT-B Small-38M 1 SemanticToken 17.4 149.2 15.6 77.1 TAP ViT-L Small-38M 1 SemanticToken 17.5 150.7 15.8 79.5 Tab 4. Ablation study on semantic prediction tasks for zero-shot classification. Default tasks are marked in gray . COCO LVIS VisualEncoder Pre-train AP APS APM APL AP APS APM APL APR APC APF ViT-L Mask,Feature 61.9 44.4 69.7 75.2 39.0 27.5 50.4 63.5 35.4 37.3 42.5 ViT-L Mask,Concept 76.8 (+14.9) 59.8 83.4 90.0 59.0 (+20.0) 44.3 71.8 82.7 60.5 58.7 58.7 ule with more data does not lead to a significant improvement in classification performance. We hypothesize that incorporating diverse data sources or applying stronger data augmentation techniques may help TAP models achieve a performance comparable to EVA-CLIP-E. Scaling language model. We scale up text decoder along the depth and embedding dimension to ablate the caption bottleneck. As illustrated in Tab. 6b, there is no substantial improvement with an increase in decoder depth or embedding dimension. This suggests that employing larger language models for region captioning task may not be necessary unless the training data could be further scaled up. 4.4. Qualitative Results We qualitatively evaluate TAP using point-based prompts. By simply drawing a click, box or scribble on a target region, the TAP model can concurrently generate the segmentation mask, class name, as well as text description. Panoptic understanding. Fig. 2 visualizes examples that typically require panoptic segmentation. The results showcase the challenge of retrieving stuff categories without access to human annotations. However, the generated region-level captions are precise in most scenarios, wherever clicked on stuff or things. This indicates that text decoder effectively activates the rich representation in the semantic tokens comparing to the classification head. Crowd understanding. Fig. 5 visualizes the result of challenging crowd regions. Our model accurately identifies and segments various elements within crowded or bustling environments. The segmentation masks precisely outline the distinct regions occupied by people, animals, food, as well as various uncommon commodities and stationeries. The associated class labels provide detailed categorization, distinguishing between different entities such as bird and dog. Furthermore, the accompanying caption provides an overall summary, including the presence of a crowd and the quantity of entities coexisting within the environment. Open-world knowledge. Fig. 6 showcases the instances that pose challenges in open-world scenarios. Due to the subjective nature of vocabulary design, curated concepts such as \u2018pepsi\u2019, \u2018cocacola\u2019, \u2018dragon\u2019, \u2018spider-man\u2019 and \u2018whisky\u2019 could hardly be selected via discriminative retrievals (i.e., classification). However, our model demonstrates distinction in these concept-related instances, demonstrating its capacity to handle open-world knowledge. This adaptability is particularly notable, as it enables the model to contribute effectively to generative tasks. 5. Conclusion In this study, we propose TAP, a promptable model trained on SemanticSA-1B, capable of concurrently segmenting, recognizing, and captioning objects within arbitrary regions. Our key findings include: a) Visual prompting can 10 Tab 5. Ablation study on pre-training tasks and text prompts. Default settings are marked in gray . VG Caption Segmentation Model Pre-train TextPrompt BLEU@4 METEOR ROUGE CIDEr APCOCO APLVIS Model A Mask MaskToken 8.6 13.1 28.7 103.4 45.8 43.1 Model B Mask,Concept MaskToken 11.0 16.5 33.9 138.3 45.6 42.6 Model C Mask,Feature SemanticToken 11.3 16.8 34.5 142.0 45.7 42.8 Model D Mask,Concept SemanticToken 11.9 17.4 35.6 149.5 45.6 42.6 Tab 6. Ablation on pre-training and fine-tuning configurations. Default configurations are marked in gray . (a) Model, data and schedule for pre-training. Pre-train Classification Model Proportion of SA-1B Schedule APCOCO APLVIS TAP-B 50% 90k 74.5 56.4 TAP-B 100% 180k 74.8 56.5 TAP-L 50% 90k 76.8 59.0 TAP-L 100% 180k 77.0 59.1 (b) Depth and dimension for language fine-tuning. TextDecoder VG Caption Model Params Depth Dim BLEU@4 METEOR ROUGE CIDEr TAP-L 20M 6 512 11.8 17.4 35.5 148.8 TAP-L 25M 8 512 11.9 17.4 35.6 149.5 TAP-L 38M 12 512 12.0 17.5 35.7 150.7 TAP-L 43M 6 768 11.9 17.4 35.6 149.8 facilitate a broader range of tasks beyond mere segmentation. b) SAM can be augmented with regional semantic awareness using an image-level CLIP, without compromising mask AP. c) An orthogonal space, such as the vocabulary concept space, is essential for effective learning of CLIP priors. d) Moreover, with a visual prompt, TAP acts as a versatile location-aware image tokenizer, where the tokenized region features can be directly used to prompt causal language modeling. We hope our work could inspire community to develop more compact and significant visionlanguage fundamental models. Limitations. Despite its advancements, TAP has two main constraints. It is trained using a human-curated label space, which still falls short of open-world assumption. This constraint also leads to an unstable ranking of similar concepts during inference (Fig. 8 left). Additionally, the text decoder, fine-tuned on a constrained set of region caption data, may limit the model\u2019s scalability and breadth of vision-language understanding. For example, the object counting cannot be well solved with obscure annotation on quantity (Fig. 8 right). Expanding the diversity of captions is anticipated to instruct model for complex understandings. Acknowledgement This project is supported by the National Key R&D Program of China (2022ZD0116302). We would like to thank Hanxiao Qu, Yan Tian and Xigang Cao for their help on Cambricon MLU resources for pre-training, as well as other colleagues at BAAI for their support to this project."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2111.12698v2",
+                    "title": "Open-Vocabulary Instance Segmentation via Robust Cross-Modal Pseudo-Labeling",
+                    "abstract": "Open-vocabulary instance segmentation aims at segmenting novel classes\nwithout mask annotations. It is an important step toward reducing laborious\nhuman supervision. Most existing works first pretrain a model on captioned\nimages covering many novel classes and then finetune it on limited base classes\nwith mask annotations. However, the high-level textual information learned from\ncaption pretraining alone cannot effectively encode the details required for\npixel-wise segmentation. To address this, we propose a cross-modal\npseudo-labeling framework, which generates training pseudo masks by aligning\nword semantics in captions with visual features of object masks in images.\nThus, our framework is capable of labeling novel classes in captions via their\nword semantics to self-train a student model. To account for noises in pseudo\nmasks, we design a robust student model that selectively distills mask\nknowledge by estimating the mask noise levels, hence mitigating the adverse\nimpact of noisy pseudo masks. By extensive experiments, we show the\neffectiveness of our framework, where we significantly improve mAP score by\n4.5% on MS-COCO and 5.1% on the large-scale Open Images & Conceptual Captions\ndatasets compared to the state-of-the-art.",
+                    "authors": "Dat Huynh, Jason Kuen, Zhe Lin, Jiuxiang Gu, Ehsan Elhamifar",
+                    "published": "2021-11-24",
+                    "updated": "2022-04-19",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2304.03284v1",
+                    "title": "SegGPT: Segmenting Everything In Context",
+                    "abstract": "We present SegGPT, a generalist model for segmenting everything in context.\nWe unify various segmentation tasks into a generalist in-context learning\nframework that accommodates different kinds of segmentation data by\ntransforming them into the same format of images. The training of SegGPT is\nformulated as an in-context coloring problem with random color mapping for each\ndata sample. The objective is to accomplish diverse tasks according to the\ncontext, rather than relying on specific colors. After training, SegGPT can\nperform arbitrary segmentation tasks in images or videos via in-context\ninference, such as object instance, stuff, part, contour, and text. SegGPT is\nevaluated on a broad range of tasks, including few-shot semantic segmentation,\nvideo object segmentation, semantic segmentation, and panoptic segmentation.\nOur results show strong capabilities in segmenting in-domain and out-of-domain\ntargets, either qualitatively or quantitatively.",
+                    "authors": "Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, Tiejun Huang",
+                    "published": "2023-04-06",
+                    "updated": "2023-04-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2304.02643v1",
+                    "title": "Segment Anything",
+                    "abstract": "We introduce the Segment Anything (SA) project: a new task, model, and\ndataset for image segmentation. Using our efficient model in a data collection\nloop, we built the largest segmentation dataset to date (by far), with over 1\nbillion masks on 11M licensed and privacy respecting images. The model is\ndesigned and trained to be promptable, so it can transfer zero-shot to new\nimage distributions and tasks. We evaluate its capabilities on numerous tasks\nand find that its zero-shot performance is impressive -- often competitive with\nor even superior to prior fully supervised results. We are releasing the\nSegment Anything Model (SAM) and corresponding dataset (SA-1B) of 1B masks and\n11M images at https://segment-anything.com to foster research into foundation\nmodels for computer vision.",
+                    "authors": "Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Doll\u00e1r, Ross Girshick",
+                    "published": "2023-04-05",
+                    "updated": "2023-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1912.04488v3",
+                    "title": "SOLO: Segmenting Objects by Locations",
+                    "abstract": "We present a new, embarrassingly simple approach to instance segmentation in\nimages. Compared to many other dense prediction tasks, e.g., semantic\nsegmentation, it is the arbitrary number of instances that have made instance\nsegmentation much more challenging. In order to predict a mask for each\ninstance, mainstream approaches either follow the 'detect-thensegment' strategy\nas used by Mask R-CNN, or predict category masks first then use clustering\ntechniques to group pixels into individual instances. We view the task of\ninstance segmentation from a completely new perspective by introducing the\nnotion of \"instance categories\", which assigns categories to each pixel within\nan instance according to the instance's location and size, thus nicely\nconverting instance mask segmentation into a classification-solvable problem.\nNow instance segmentation is decomposed into two classification tasks. We\ndemonstrate a much simpler and flexible instance segmentation framework with\nstrong performance, achieving on par accuracy with Mask R-CNN and outperforming\nrecent singleshot instance segmenters in accuracy. We hope that this very\nsimple and strong framework can serve as a baseline for many instance-level\nrecognition tasks besides instance segmentation.",
+                    "authors": "Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, Lei Li",
+                    "published": "2019-12-10",
+                    "updated": "2020-07-19",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.10221v1",
+                    "title": "Unified Perceptual Parsing for Scene Understanding",
+                    "abstract": "Humans recognize the visual world at multiple levels: we effortlessly\ncategorize scenes and detect objects inside, while also identifying the\ntextures and surfaces of the objects along with their different compositional\nparts. In this paper, we study a new task called Unified Perceptual Parsing,\nwhich requires the machine vision systems to recognize as many visual concepts\nas possible from a given image. A multi-task framework called UPerNet and a\ntraining strategy are developed to learn from heterogeneous image annotations.\nWe benchmark our framework on Unified Perceptual Parsing and show that it is\nable to effectively segment a wide range of concepts from images. The trained\nnetworks are further applied to discover visual knowledge in natural scenes.\nModels are available at \\url{https://github.com/CSAILVision/unifiedparsing}.",
+                    "authors": "Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun",
+                    "published": "2018-07-26",
+                    "updated": "2018-07-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.06278v2",
+                    "title": "Per-Pixel Classification is Not All You Need for Semantic Segmentation",
+                    "abstract": "Modern approaches typically formulate semantic segmentation as a per-pixel\nclassification task, while instance-level segmentation is handled with an\nalternative mask classification. Our key insight: mask classification is\nsufficiently general to solve both semantic- and instance-level segmentation\ntasks in a unified manner using the exact same model, loss, and training\nprocedure. Following this observation, we propose MaskFormer, a simple mask\nclassification model which predicts a set of binary masks, each associated with\na single global class label prediction. Overall, the proposed mask\nclassification-based method simplifies the landscape of effective approaches to\nsemantic and panoptic segmentation tasks and shows excellent empirical results.\nIn particular, we observe that MaskFormer outperforms per-pixel classification\nbaselines when the number of classes is large. Our mask classification-based\nmethod outperforms both current state-of-the-art semantic (55.6 mIoU on ADE20K)\nand panoptic segmentation (52.7 PQ on COCO) models.",
+                    "authors": "Bowen Cheng, Alexander G. Schwing, Alexander Kirillov",
+                    "published": "2021-07-13",
+                    "updated": "2021-10-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.04803v4",
+                    "title": "Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models",
+                    "abstract": "We present ODISE: Open-vocabulary DIffusion-based panoptic SEgmentation,\nwhich unifies pre-trained text-image diffusion and discriminative models to\nperform open-vocabulary panoptic segmentation. Text-to-image diffusion models\nhave the remarkable ability to generate high-quality images with diverse\nopen-vocabulary language descriptions. This demonstrates that their internal\nrepresentation space is highly correlated with open concepts in the real world.\nText-image discriminative models like CLIP, on the other hand, are good at\nclassifying images into open-vocabulary labels. We leverage the frozen internal\nrepresentations of both these models to perform panoptic segmentation of any\ncategory in the wild. Our approach outperforms the previous state of the art by\nsignificant margins on both open-vocabulary panoptic and semantic segmentation\ntasks. In particular, with COCO training only, our method achieves 23.4 PQ and\n30.0 mIoU on the ADE20K dataset, with 8.3 PQ and 7.9 mIoU absolute improvement\nover the previous state of the art. We open-source our code and models at\nhttps://github.com/NVlabs/ODISE .",
+                    "authors": "Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, Shalini De Mello",
+                    "published": "2023-03-08",
+                    "updated": "2023-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.09316v1",
+                    "title": "Diffusion Models for Zero-Shot Open-Vocabulary Segmentation",
+                    "abstract": "The variety of objects in the real world is nearly unlimited and is thus\nimpossible to capture using models trained on a fixed set of categories. As a\nresult, in recent years, open-vocabulary methods have attracted the interest of\nthe community. This paper proposes a new method for zero-shot open-vocabulary\nsegmentation. Prior work largely relies on contrastive training using\nimage-text pairs, leveraging grouping mechanisms to learn image features that\nare both aligned with language and well-localised. This however can introduce\nambiguity as the visual appearance of images with similar captions often\nvaries. Instead, we leverage the generative properties of large-scale\ntext-to-image diffusion models to sample a set of support images for a given\ntextual category. This provides a distribution of appearances for a given text\ncircumventing the ambiguity problem. We further propose a mechanism that\nconsiders the contextual background of the sampled images to better localise\nobjects and segment the background directly. We show that our method can be\nused to ground several existing pre-trained self-supervised feature extractors\nin natural language and provide explainable predictions by mapping back to\nregions in the support set. Our proposal is training-free, relying on\npre-trained components only, yet, shows strong performance on a range of\nopen-vocabulary segmentation benchmarks, obtaining a lead of more than 10% on\nthe Pascal VOC benchmark.",
+                    "authors": "Laurynas Karazija, Iro Laina, Andrea Vedaldi, Christian Rupprecht",
+                    "published": "2023-06-15",
+                    "updated": "2023-06-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2003.10152v3",
+                    "title": "SOLOv2: Dynamic and Fast Instance Segmentation",
+                    "abstract": "In this work, we aim at building a simple, direct, and fast instance\nsegmentation framework with strong performance. We follow the principle of the\nSOLO method of Wang et al. \"SOLO: segmenting objects by locations\".\nImportantly, we take one step further by dynamically learning the mask head of\nthe object segmenter such that the mask head is conditioned on the location.\nSpecifically, the mask branch is decoupled into a mask kernel branch and mask\nfeature branch, which are responsible for learning the convolution kernel and\nthe convolved features respectively. Moreover, we propose Matrix NMS (non\nmaximum suppression) to significantly reduce the inference time overhead due to\nNMS of masks. Our Matrix NMS performs NMS with parallel matrix operations in\none shot, and yields better results. We demonstrate a simple direct instance\nsegmentation system, outperforming a few state-of-the-art methods in both speed\nand accuracy. A light-weight version of SOLOv2 executes at 31.3 FPS and yields\n37.1% AP. Moreover, our state-of-the-art results in object detection (from our\nmask byproduct) and panoptic segmentation show the potential to serve as a new\nstrong baseline for many instance-level recognition tasks besides instance\nsegmentation. Code is available at: https://git.io/AdelaiDet",
+                    "authors": "Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, Chunhua Shen",
+                    "published": "2020-03-23",
+                    "updated": "2020-10-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.04150v3",
+                    "title": "Open-Vocabulary Semantic Segmentation with Mask-adapted CLIP",
+                    "abstract": "Open-vocabulary semantic segmentation aims to segment an image into semantic\nregions according to text descriptions, which may not have been seen during\ntraining. Recent two-stage methods first generate class-agnostic mask proposals\nand then leverage pre-trained vision-language models, e.g., CLIP, to classify\nmasked regions. We identify the performance bottleneck of this paradigm to be\nthe pre-trained CLIP model, since it does not perform well on masked images. To\naddress this, we propose to finetune CLIP on a collection of masked image\nregions and their corresponding text descriptions. We collect training data by\nmining an existing image-caption dataset (e.g., COCO Captions), using CLIP to\nmatch masked image regions to nouns in the image captions. Compared with the\nmore precise and manually annotated segmentation labels with fixed classes\n(e.g., COCO-Stuff), we find our noisy but diverse dataset can better retain\nCLIP's generalization ability. Along with finetuning the entire model, we\nutilize the \"blank\" areas in masked images using a method we dub mask prompt\ntuning. Experiments demonstrate mask prompt tuning brings significant\nimprovement without modifying any weights of CLIP, and it can further improve a\nfully finetuned model. In particular, when trained on COCO and evaluated on\nADE20K-150, our best model achieves 29.6% mIoU, which is +8.5% higher than the\nprevious state-of-the-art. For the first time, open-vocabulary generalist\nmodels match the performance of supervised specialist models in 2017 without\ndataset-specific adaptations.",
+                    "authors": "Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, Diana Marculescu",
+                    "published": "2022-10-09",
+                    "updated": "2023-04-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2103.00020v1",
+                    "title": "Learning Transferable Visual Models From Natural Language Supervision",
+                    "abstract": "State-of-the-art computer vision systems are trained to predict a fixed set\nof predetermined object categories. This restricted form of supervision limits\ntheir generality and usability since additional labeled data is needed to\nspecify any other visual concept. Learning directly from raw text about images\nis a promising alternative which leverages a much broader source of\nsupervision. We demonstrate that the simple pre-training task of predicting\nwhich caption goes with which image is an efficient and scalable way to learn\nSOTA image representations from scratch on a dataset of 400 million (image,\ntext) pairs collected from the internet. After pre-training, natural language\nis used to reference learned visual concepts (or describe new ones) enabling\nzero-shot transfer of the model to downstream tasks. We study the performance\nof this approach by benchmarking on over 30 different existing computer vision\ndatasets, spanning tasks such as OCR, action recognition in videos,\ngeo-localization, and many types of fine-grained object classification. The\nmodel transfers non-trivially to most tasks and is often competitive with a\nfully supervised baseline without the need for any dataset specific training.\nFor instance, we match the accuracy of the original ResNet-50 on ImageNet\nzero-shot without needing to use any of the 1.28 million training examples it\nwas trained on. We release our code and pre-trained model weights at\nhttps://github.com/OpenAI/CLIP.",
+                    "authors": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever",
+                    "published": "2021-02-26",
+                    "updated": "2021-02-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2202.11094v5",
+                    "title": "GroupViT: Semantic Segmentation Emerges from Text Supervision",
+                    "abstract": "Grouping and recognition are important components of visual scene\nunderstanding, e.g., for object detection and semantic segmentation. With\nend-to-end deep learning systems, grouping of image regions usually happens\nimplicitly via top-down supervision from pixel-level recognition labels.\nInstead, in this paper, we propose to bring back the grouping mechanism into\ndeep networks, which allows semantic segments to emerge automatically with only\ntext supervision. We propose a hierarchical Grouping Vision Transformer\n(GroupViT), which goes beyond the regular grid structure representation and\nlearns to group image regions into progressively larger arbitrary-shaped\nsegments. We train GroupViT jointly with a text encoder on a large-scale\nimage-text dataset via contrastive losses. With only text supervision and\nwithout any pixel-level annotations, GroupViT learns to group together semantic\nregions and successfully transfers to the task of semantic segmentation in a\nzero-shot manner, i.e., without any further fine-tuning. It achieves a\nzero-shot accuracy of 52.3% mIoU on the PASCAL VOC 2012 and 22.4% mIoU on\nPASCAL Context datasets, and performs competitively to state-of-the-art\ntransfer-learning methods requiring greater levels of supervision. We\nopen-source our code at https://github.com/NVlabs/GroupViT .",
+                    "authors": "Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang",
+                    "published": "2022-02-22",
+                    "updated": "2022-07-18",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.01907v1",
+                    "title": "The All-Seeing Project: Towards Panoptic Visual Recognition and Understanding of the Open World",
+                    "abstract": "We present the All-Seeing (AS) project: a large-scale data and model for\nrecognizing and understanding everything in the open world. Using a scalable\ndata engine that incorporates human feedback and efficient models in the loop,\nwe create a new dataset (AS-1B) with over 1 billion regions annotated with\nsemantic tags, question-answering pairs, and detailed captions. It covers a\nwide range of 3.5 million common and rare concepts in the real world, and has\n132.2 billion tokens that describe the concepts and their attributes.\nLeveraging this new dataset, we develop the All-Seeing model (ASM), a unified\nframework for panoptic visual recognition and understanding. The model is\ntrained with open-ended language prompts and locations, which allows it to\ngeneralize to various vision and language tasks with remarkable zero-shot\nperformance, including region-text retrieval, region recognition, captioning,\nand question-answering. We hope that this project can serve as a foundation for\nvision-language artificial general intelligence research. Models and the\ndataset shall be released at https://github.com/OpenGVLab/All-Seeing, and demo\ncan be seen at https://huggingface.co/spaces/OpenGVLab/all-seeing.",
+                    "authors": "Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, Yushi Chen, Tong Lu, Jifeng Dai, Yu Qiao",
+                    "published": "2023-08-03",
+                    "updated": "2023-08-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2209.15639v2",
+                    "title": "F-VLM: Open-Vocabulary Object Detection upon Frozen Vision and Language Models",
+                    "abstract": "We present F-VLM, a simple open-vocabulary object detection method built upon\nFrozen Vision and Language Models. F-VLM simplifies the current multi-stage\ntraining pipeline by eliminating the need for knowledge distillation or\ndetection-tailored pretraining. Surprisingly, we observe that a frozen VLM: 1)\nretains the locality-sensitive features necessary for detection, and 2) is a\nstrong region classifier. We finetune only the detector head and combine the\ndetector and VLM outputs for each region at inference time. F-VLM shows\ncompelling scaling behavior and achieves +6.5 mask AP improvement over the\nprevious state of the art on novel categories of LVIS open-vocabulary detection\nbenchmark. In addition, we demonstrate very competitive results on COCO\nopen-vocabulary detection benchmark and cross-dataset transfer detection, in\naddition to significant training speed-up and compute savings. Code will be\nreleased at the https://sites.google.com/view/f-vlm/home",
+                    "authors": "Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, Anelia Angelova",
+                    "published": "2022-09-30",
+                    "updated": "2023-02-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2201.03546v2",
+                    "title": "Language-driven Semantic Segmentation",
+                    "abstract": "We present LSeg, a novel model for language-driven semantic image\nsegmentation. LSeg uses a text encoder to compute embeddings of descriptive\ninput labels (e.g., \"grass\" or \"building\") together with a transformer-based\nimage encoder that computes dense per-pixel embeddings of the input image. The\nimage encoder is trained with a contrastive objective to align pixel embeddings\nto the text embedding of the corresponding semantic class. The text embeddings\nprovide a flexible label representation in which semantically similar labels\nmap to similar regions in the embedding space (e.g., \"cat\" and \"furry\"). This\nallows LSeg to generalize to previously unseen categories at test time, without\nretraining or even requiring a single additional training sample. We\ndemonstrate that our approach achieves highly competitive zero-shot performance\ncompared to existing zero- and few-shot semantic segmentation methods, and even\nmatches the accuracy of traditional segmentation algorithms when a fixed label\nset is provided. Code and demo are available at\nhttps://github.com/isl-org/lang-seg.",
+                    "authors": "Boyi Li, Kilian Q. Weinberger, Serge Belongie, Vladlen Koltun, Ren\u00e9 Ranftl",
+                    "published": "2022-01-10",
+                    "updated": "2022-04-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.CL",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2212.02499v2",
+                    "title": "Images Speak in Images: A Generalist Painter for In-Context Visual Learning",
+                    "abstract": "In-context learning, as a new paradigm in NLP, allows the model to rapidly\nadapt to various tasks with only a handful of prompts and examples. But in\ncomputer vision, the difficulties for in-context learning lie in that tasks\nvary significantly in the output representations, thus it is unclear how to\ndefine the general-purpose task prompts that the vision model can understand\nand transfer to out-of-domain tasks. In this work, we present Painter, a\ngeneralist model which addresses these obstacles with an \"image\"-centric\nsolution, that is, to redefine the output of core vision tasks as images, and\nspecify task prompts as also images. With this idea, our training process is\nextremely simple, which performs standard masked image modeling on the stitch\nof input and output image pairs. This makes the model capable of performing\ntasks conditioned on visible image patches. Thus, during inference, we can\nadopt a pair of input and output images from the same task as the input\ncondition, to indicate which task to perform. Without bells and whistles, our\ngeneralist Painter can achieve competitive performance compared to\nwell-established task-specific models, on seven representative vision tasks\nranging from high-level visual understanding to low-level image processing. In\naddition, Painter significantly outperforms recent generalist models on several\nchallenging tasks.",
+                    "authors": "Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, Tiejun Huang",
+                    "published": "2022-12-05",
+                    "updated": "2023-03-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2112.07910v2",
+                    "title": "Decoupling Zero-Shot Semantic Segmentation",
+                    "abstract": "Zero-shot semantic segmentation (ZS3) aims to segment the novel categories\nthat have not been seen in the training. Existing works formulate ZS3 as a\npixel-level zeroshot classification problem, and transfer semantic knowledge\nfrom seen classes to unseen ones with the help of language models pre-trained\nonly with texts. While simple, the pixel-level ZS3 formulation shows the\nlimited capability to integrate vision-language models that are often\npre-trained with image-text pairs and currently demonstrate great potential for\nvision tasks. Inspired by the observation that humans often perform\nsegment-level semantic labeling, we propose to decouple the ZS3 into two\nsub-tasks: 1) a classagnostic grouping task to group the pixels into segments.\n2) a zero-shot classification task on segments. The former task does not\ninvolve category information and can be directly transferred to group pixels\nfor unseen classes. The latter task performs at segment-level and provides a\nnatural way to leverage large-scale vision-language models pre-trained with\nimage-text pairs (e.g. CLIP) for ZS3. Based on the decoupling formulation, we\npropose a simple and effective zero-shot semantic segmentation model, called\nZegFormer, which outperforms the previous methods on ZS3 standard benchmarks by\nlarge margins, e.g., 22 points on the PASCAL VOC and 3 points on the COCO-Stuff\nin terms of mIoU for unseen classes. Code will be released at\nhttps://github.com/dingjiansw101/ZegFormer.",
+                    "authors": "Jian Ding, Nan Xue, Gui-Song Xia, Dengxin Dai",
+                    "published": "2021-12-15",
+                    "updated": "2022-04-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.13921v3",
+                    "title": "Open-vocabulary Object Detection via Vision and Language Knowledge Distillation",
+                    "abstract": "We aim at advancing open-vocabulary object detection, which detects objects\ndescribed by arbitrary text inputs. The fundamental challenge is the\navailability of training data. It is costly to further scale up the number of\nclasses contained in existing object detection datasets. To overcome this\nchallenge, we propose ViLD, a training method via Vision and Language knowledge\nDistillation. Our method distills the knowledge from a pretrained\nopen-vocabulary image classification model (teacher) into a two-stage detector\n(student). Specifically, we use the teacher model to encode category texts and\nimage regions of object proposals. Then we train a student detector, whose\nregion embeddings of detected boxes are aligned with the text and image\nembeddings inferred by the teacher. We benchmark on LVIS by holding out all\nrare categories as novel categories that are not seen during training. ViLD\nobtains 16.1 mask AP$_r$ with a ResNet-50 backbone, even outperforming the\nsupervised counterpart by 3.8. When trained with a stronger teacher model\nALIGN, ViLD achieves 26.3 AP$_r$. The model can directly transfer to other\ndatasets without finetuning, achieving 72.2 AP$_{50}$ on PASCAL VOC, 36.6 AP on\nCOCO and 11.8 AP on Objects365. On COCO, ViLD outperforms the previous\nstate-of-the-art by 4.8 on novel AP and 11.4 on overall AP. Code and demo are\nopen-sourced at\nhttps://github.com/tensorflow/tpu/tree/master/models/official/detection/projects/vild.",
+                    "authors": "Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, Yin Cui",
+                    "published": "2021-04-28",
+                    "updated": "2022-05-12",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2112.09106v1",
+                    "title": "RegionCLIP: Region-based Language-Image Pretraining",
+                    "abstract": "Contrastive language-image pretraining (CLIP) using image-text pairs has\nachieved impressive results on image classification in both zero-shot and\ntransfer learning settings. However, we show that directly applying such models\nto recognize image regions for object detection leads to poor performance due\nto a domain shift: CLIP was trained to match an image as a whole to a text\ndescription, without capturing the fine-grained alignment between image regions\nand text spans. To mitigate this issue, we propose a new method called\nRegionCLIP that significantly extends CLIP to learn region-level visual\nrepresentations, thus enabling fine-grained alignment between image regions and\ntextual concepts. Our method leverages a CLIP model to match image regions with\ntemplate captions and then pretrains our model to align these region-text pairs\nin the feature space. When transferring our pretrained model to the\nopen-vocabulary object detection tasks, our method significantly outperforms\nthe state of the art by 3.8 AP50 and 2.2 AP for novel categories on COCO and\nLVIS datasets, respectively. Moreoever, the learned region representations\nsupport zero-shot inference for object detection, showing promising results on\nboth COCO and LVIS datasets. Our code is available at\nhttps://github.com/microsoft/RegionCLIP.",
+                    "authors": "Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, Jianfeng Gao",
+                    "published": "2021-12-16",
+                    "updated": "2021-12-16",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2011.10678v2",
+                    "title": "Open-Vocabulary Object Detection Using Captions",
+                    "abstract": "Despite the remarkable accuracy of deep neural networks in object detection,\nthey are costly to train and scale due to supervision requirements.\nParticularly, learning more object categories typically requires proportionally\nmore bounding box annotations. Weakly supervised and zero-shot learning\ntechniques have been explored to scale object detectors to more categories with\nless supervision, but they have not been as successful and widely adopted as\nsupervised models. In this paper, we put forth a novel formulation of the\nobject detection problem, namely open-vocabulary object detection, which is\nmore general, more practical, and more effective than weakly supervised and\nzero-shot approaches. We propose a new method to train object detectors using\nbounding box annotations for a limited set of object categories, as well as\nimage-caption pairs that cover a larger variety of objects at a significantly\nlower cost. We show that the proposed method can detect and localize objects\nfor which no bounding box annotation is provided during training, at a\nsignificantly higher accuracy than zero-shot approaches. Meanwhile, objects\nwith bounding box annotation can be detected almost as accurately as supervised\nmethods, which is significantly better than weakly supervised baselines.\nAccordingly, we establish a new state of the art for scalable object detection.",
+                    "authors": "Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, Shih-Fu Chang",
+                    "published": "2020-11-20",
+                    "updated": "2021-03-14",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2102.05918v2",
+                    "title": "Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision",
+                    "abstract": "Pre-trained representations are becoming crucial for many NLP and perception\ntasks. While representation learning in NLP has transitioned to training on raw\ntext without human annotations, visual and vision-language representations\nstill rely heavily on curated training datasets that are expensive or require\nexpert knowledge. For vision applications, representations are mostly learned\nusing datasets with explicit class labels such as ImageNet or OpenImages. For\nvision-language, popular datasets like Conceptual Captions, MSCOCO, or CLIP all\ninvolve a non-trivial data collection (and cleaning) process. This costly\ncuration process limits the size of datasets and hence hinders the scaling of\ntrained models. In this paper, we leverage a noisy dataset of over one billion\nimage alt-text pairs, obtained without expensive filtering or post-processing\nsteps in the Conceptual Captions dataset. A simple dual-encoder architecture\nlearns to align visual and language representations of the image and text pairs\nusing a contrastive loss. We show that the scale of our corpus can make up for\nits noise and leads to state-of-the-art representations even with such a simple\nlearning scheme. Our visual representation achieves strong performance when\ntransferred to classification tasks such as ImageNet and VTAB. The aligned\nvisual and language representations enables zero-shot image classification and\nalso set new state-of-the-art results on Flickr30K and MSCOCO image-text\nretrieval benchmarks, even when compared with more sophisticated\ncross-attention models. The representations also enable cross-modality search\nwith complex text and text + image queries.",
+                    "authors": "Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig",
+                    "published": "2021-02-11",
+                    "updated": "2021-06-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.CL",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2307.05222v2",
+                    "title": "Emu: Generative Pretraining in Multimodality",
+                    "abstract": "We present Emu, a Transformer-based multimodal foundation model, which can\nseamlessly generate images and texts in multimodal context. This omnivore model\ncan take in any single-modality or multimodal data input indiscriminately\n(e.g., interleaved image, text and video) through a one-model-for-all\nautoregressive training process. First, visual signals are encoded into\nembeddings, and together with text tokens form an interleaved input sequence.\nEmu is then end-to-end trained with a unified objective of classifying the next\ntext token or regressing the next visual embedding in the multimodal sequence.\nThis versatile multimodality empowers the exploration of diverse pretraining\ndata sources at scale, such as videos with interleaved frames and text,\nwebpages with interleaved images and text, as well as web-scale image-text\npairs and video-text pairs. Emu can serve as a generalist multimodal interface\nfor both image-to-text and text-to-image tasks, and supports in-context image\nand text generation. Across a broad range of zero-shot/few-shot tasks including\nimage captioning, visual question answering, video question answering and\ntext-to-image generation, Emu demonstrates superb performance compared to\nstate-of-the-art large multimodal models. Extended capabilities such as\nmultimodal assistants via instruction tuning are also demonstrated with\nimpressive performance.",
+                    "authors": "Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, Xinlong Wang",
+                    "published": "2023-07-11",
+                    "updated": "2024-05-08",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1411.4038v2",
+                    "title": "Fully Convolutional Networks for Semantic Segmentation",
+                    "abstract": "Convolutional networks are powerful visual models that yield hierarchies of\nfeatures. We show that convolutional networks by themselves, trained\nend-to-end, pixels-to-pixels, exceed the state-of-the-art in semantic\nsegmentation. Our key insight is to build \"fully convolutional\" networks that\ntake input of arbitrary size and produce correspondingly-sized output with\nefficient inference and learning. We define and detail the space of fully\nconvolutional networks, explain their application to spatially dense prediction\ntasks, and draw connections to prior models. We adapt contemporary\nclassification networks (AlexNet, the VGG net, and GoogLeNet) into fully\nconvolutional networks and transfer their learned representations by\nfine-tuning to the segmentation task. We then define a novel architecture that\ncombines semantic information from a deep, coarse layer with appearance\ninformation from a shallow, fine layer to produce accurate and detailed\nsegmentations. Our fully convolutional network achieves state-of-the-art\nsegmentation of PASCAL VOC (20% relative improvement to 62.2% mean IU on 2012),\nNYUDv2, and SIFT Flow, while inference takes one third of a second for a\ntypical image.",
+                    "authors": "Jonathan Long, Evan Shelhamer, Trevor Darrell",
+                    "published": "2014-11-14",
+                    "updated": "2015-03-08",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.17225v1",
+                    "title": "FreeSeg: Unified, Universal and Open-Vocabulary Image Segmentation",
+                    "abstract": "Recently, open-vocabulary learning has emerged to accomplish segmentation for\narbitrary categories of text-based descriptions, which popularizes the\nsegmentation system to more general-purpose application scenarios. However,\nexisting methods devote to designing specialized architectures or parameters\nfor specific segmentation tasks. These customized design paradigms lead to\nfragmentation between various segmentation tasks, thus hindering the uniformity\nof segmentation models. Hence in this paper, we propose FreeSeg, a generic\nframework to accomplish Unified, Universal and Open-Vocabulary Image\nSegmentation. FreeSeg optimizes an all-in-one network via one-shot training and\nemploys the same architecture and parameters to handle diverse segmentation\ntasks seamlessly in the inference procedure. Additionally, adaptive prompt\nlearning facilitates the unified model to capture task-aware and\ncategory-sensitive concepts, improving model robustness in multi-task and\nvaried scenarios. Extensive experimental results demonstrate that FreeSeg\nestablishes new state-of-the-art results in performance and generalization on\nthree segmentation tasks, which outperforms the best task-specific\narchitectures by a large margin: 5.5% mIoU on semantic segmentation, 17.6% mAP\non instance segmentation, 20.1% PQ on panoptic segmentation for the unseen\nclass on COCO.",
+                    "authors": "Jie Qin, Jie Wu, Pengxiang Yan, Ming Li, Ren Yuxi, Xuefeng Xiao, Yitong Wang, Rui Wang, Shilei Wen, Xin Pan, Xingang Wang",
+                    "published": "2023-03-30",
+                    "updated": "2023-03-30",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2304.06718v4",
+                    "title": "Segment Everything Everywhere All at Once",
+                    "abstract": "In this work, we present SEEM, a promptable and interactive model for\nsegmenting everything everywhere all at once in an image, as shown in Fig.1. In\nSEEM, we propose a novel decoding mechanism that enables diverse prompting for\nall types of segmentation tasks, aiming at a universal segmentation interface\nthat behaves like large language models (LLMs). More specifically, SEEM is\ndesigned with four desiderata: i) Versatility. We introduce a new visual prompt\nto unify different spatial queries including points, boxes, scribbles and\nmasks, which can further generalize to a different referring image; ii)\nCompositionality. We learn a joint visual-semantic space between text and\nvisual prompts, which facilitates the dynamic composition of two prompt types\nrequired for various segmentation tasks; iii) Interactivity. We further\nincorporate learnable memory prompts into the decoder to retain segmentation\nhistory through mask-guided cross-attention from decoder to image features; and\niv) Semantic-awareness. We use a text encoder to encode text queries and mask\nlabels into the same semantic space for open-vocabulary segmentation. We\nconduct a comprehensive empirical study to validate the effectiveness of SEEM\nacross diverse segmentation tasks. Notably, our single SEEM model achieves\ncompetitive performance across interactive segmentation, generic segmentation,\nreferring segmentation, and video object segmentation on 9 datasets with\nminimum 1/100 supervision. Furthermore, SEEM showcases a remarkable capacity\nfor generalization to novel prompts or their combinations, rendering it a\nreadily universal image segmentation interface.",
+                    "authors": "Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, Yong Jae Lee",
+                    "published": "2023-04-13",
+                    "updated": "2023-07-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.05892v1",
+                    "title": "Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection",
+                    "abstract": "Open-vocabulary object detection aims to provide object detectors trained on\na fixed set of object categories with the generalizability to detect objects\ndescribed by arbitrary text queries. Previous methods adopt knowledge\ndistillation to extract knowledge from Pretrained Vision-and-Language Models\n(PVLMs) and transfer it to detectors. However, due to the non-adaptive proposal\ncropping and single-level feature mimicking processes, they suffer from\ninformation destruction during knowledge extraction and inefficient knowledge\ntransfer. To remedy these limitations, we propose an Object-Aware Distillation\nPyramid (OADP) framework, including an Object-Aware Knowledge Extraction (OAKE)\nmodule and a Distillation Pyramid (DP) mechanism. When extracting object\nknowledge from PVLMs, the former adaptively transforms object proposals and\nadopts object-aware mask attention to obtain precise and complete knowledge of\nobjects. The latter introduces global and block distillation for more\ncomprehensive knowledge transfer to compensate for the missing relation\ninformation in object distillation. Extensive experiments show that our method\nachieves significant improvement compared to current methods. Especially on the\nMS-COCO dataset, our OADP framework reaches $35.6$ mAP$^{\\text{N}}_{50}$,\nsurpassing the current state-of-the-art method by $3.3$ mAP$^{\\text{N}}_{50}$.\nCode is released at https://github.com/LutingWang/OADP.",
+                    "authors": "Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, Si Liu",
+                    "published": "2023-03-10",
+                    "updated": "2023-03-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2307.04767v1",
+                    "title": "Semantic-SAM: Segment and Recognize Anything at Any Granularity",
+                    "abstract": "In this paper, we introduce Semantic-SAM, a universal image segmentation\nmodel to enable segment and recognize anything at any desired granularity. Our\nmodel offers two key advantages: semantic-awareness and granularity-abundance.\nTo achieve semantic-awareness, we consolidate multiple datasets across three\ngranularities and introduce decoupled classification for objects and parts.\nThis allows our model to capture rich semantic information. For the\nmulti-granularity capability, we propose a multi-choice learning scheme during\ntraining, enabling each click to generate masks at multiple levels that\ncorrespond to multiple ground-truth masks. Notably, this work represents the\nfirst attempt to jointly train a model on SA-1B, generic, and part segmentation\ndatasets. Experimental results and visualizations demonstrate that our model\nsuccessfully achieves semantic-awareness and granularity-abundance.\nFurthermore, combining SA-1B training with other segmentation tasks, such as\npanoptic and part segmentation, leads to performance improvements. We will\nprovide code and a demo for further exploration and evaluation.",
+                    "authors": "Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, Jianfeng Gao",
+                    "published": "2023-07-10",
+                    "updated": "2023-07-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2301.00805v2",
+                    "title": "Betrayed by Captions: Joint Caption Grounding and Generation for Open Vocabulary Instance Segmentation",
+                    "abstract": "In this work, we focus on open vocabulary instance segmentation to expand a\nsegmentation model to classify and segment instance-level novel categories.\nPrevious approaches have relied on massive caption datasets and complex\npipelines to establish one-to-one mappings between image regions and words in\ncaptions. However, such methods build noisy supervision by matching non-visible\nwords to image regions, such as adjectives and verbs. Meanwhile, context words\nare also important for inferring the existence of novel objects as they show\nhigh inter-correlations with novel categories. To overcome these limitations,\nwe devise a joint \\textbf{Caption Grounding and Generation (CGG)} framework,\nwhich incorporates a novel grounding loss that only focuses on matching object\nnouns to improve learning efficiency. We also introduce a caption generation\nhead that enables additional supervision and contextual modeling as a\ncomplementation to the grounding loss. Our analysis and results demonstrate\nthat grounding and generation components complement each other, significantly\nenhancing the segmentation performance for novel classes. Experiments on the\nCOCO dataset with two settings: Open Vocabulary Instance Segmentation (OVIS)\nand Open Set Panoptic Segmentation (OSPS) demonstrate the superiority of the\nCGG. Specifically, CGG achieves a substantial improvement of 6.8% mAP for novel\nclasses without extra data on the OVIS task and 15% PQ improvements for novel\nclasses on the OSPS benchmark.",
+                    "authors": "Jianzong Wu, Xiangtai Li, Henghui Ding, Xia Li, Guangliang Cheng, Yunhai Tong, Chen Change Loy",
+                    "published": "2023-01-02",
+                    "updated": "2023-07-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.08131v3",
+                    "title": "A Simple Framework for Open-Vocabulary Segmentation and Detection",
+                    "abstract": "We present OpenSeeD, a simple Open-vocabulary Segmentation and Detection\nframework that jointly learns from different segmentation and detection\ndatasets. To bridge the gap of vocabulary and annotation granularity, we first\nintroduce a pre-trained text encoder to encode all the visual concepts in two\ntasks and learn a common semantic space for them. This gives us reasonably good\nresults compared with the counterparts trained on segmentation task only. To\nfurther reconcile them, we locate two discrepancies: $i$) task discrepancy --\nsegmentation requires extracting masks for both foreground objects and\nbackground stuff, while detection merely cares about the former; $ii$) data\ndiscrepancy -- box and mask annotations are with different spatial granularity,\nand thus not directly interchangeable. To address these issues, we propose a\ndecoupled decoding to reduce the interference between foreground/background and\na conditioned mask decoding to assist in generating masks for given boxes. To\nthis end, we develop a simple encoder-decoder model encompassing all three\ntechniques and train it jointly on COCO and Objects365. After pre-training, our\nmodel exhibits competitive or stronger zero-shot transferability for both\nsegmentation and detection. Specifically, OpenSeeD beats the state-of-the-art\nmethod for open-vocabulary instance and panoptic segmentation across 5\ndatasets, and outperforms previous work for open-vocabulary detection on LVIS\nand ODinW under similar settings. When transferred to specific tasks, our model\nachieves new SoTA for panoptic segmentation on COCO and ADE20K, and instance\nsegmentation on ADE20K and Cityscapes.\n  Finally, we note that OpenSeeD is the first to explore the potential of joint\ntraining on segmentation and detection, and hope it can be received as a strong\nbaseline for developing a single model for both tasks in open world.",
+                    "authors": "Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianfeng Gao, Jianwei Yang, Lei Zhang",
+                    "published": "2023-03-14",
+                    "updated": "2023-03-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2112.01071v2",
+                    "title": "Extract Free Dense Labels from CLIP",
+                    "abstract": "Contrastive Language-Image Pre-training (CLIP) has made a remarkable\nbreakthrough in open-vocabulary zero-shot image recognition. Many recent\nstudies leverage the pre-trained CLIP models for image-level classification and\nmanipulation. In this paper, we wish examine the intrinsic potential of CLIP\nfor pixel-level dense prediction, specifically in semantic segmentation. To\nthis end, with minimal modification, we show that MaskCLIP yields compelling\nsegmentation results on open concepts across various datasets in the absence of\nannotations and fine-tuning. By adding pseudo labeling and self-training,\nMaskCLIP+ surpasses SOTA transductive zero-shot semantic segmentation methods\nby large margins, e.g., mIoUs of unseen classes on PASCAL VOC/PASCAL\nContext/COCO Stuff are improved from 35.6/20.7/30.3 to 86.1/66.7/54.7. We also\ntest the robustness of MaskCLIP under input corruption and evaluate its\ncapability in discriminating fine-grained objects and novel concepts. Our\nfinding suggests that MaskCLIP can serve as a new reliable source of\nsupervision for dense prediction tasks to achieve annotation-free segmentation.\nSource code is available at https://github.com/chongzhou96/MaskCLIP.",
+                    "authors": "Chong Zhou, Chen Change Loy, Bo Dai",
+                    "published": "2021-12-02",
+                    "updated": "2022-07-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.CL"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2206.08916v2",
+                    "title": "Unified-IO: A Unified Model for Vision, Language, and Multi-Modal Tasks",
+                    "abstract": "We propose Unified-IO, a model that performs a large variety of AI tasks\nspanning classical computer vision tasks, including pose estimation, object\ndetection, depth estimation and image generation, vision-and-language tasks\nsuch as region captioning and referring expression, to natural language\nprocessing tasks such as question answering and paraphrasing. Developing a\nsingle unified model for such a large variety of tasks poses unique challenges\ndue to the heterogeneous inputs and outputs pertaining to each task, including\nRGB images, per-pixel maps, binary masks, bounding boxes, and language. We\nachieve this unification by homogenizing every supported input and output into\na sequence of discrete vocabulary tokens. This common representation across all\ntasks allows us to train a single transformer-based architecture, jointly on\nover 90 diverse datasets in the vision and language fields. Unified-IO is the\nfirst model capable of performing all 7 tasks on the GRIT benchmark and\nproduces strong results across 16 diverse benchmarks like NYUv2-Depth,\nImageNet, VQA2.0, OK-VQA, Swig, VizWizGround, BoolQ, and SciTail, with no\ntask-specific fine-tuning. Code and demos for Unified-IO are available at:\nhttps://unified-io.allenai.org.",
+                    "authors": "Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, Aniruddha Kembhavi",
+                    "published": "2022-06-17",
+                    "updated": "2022-10-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2212.11270v1",
+                    "title": "Generalized Decoding for Pixel, Image, and Language",
+                    "abstract": "We present X-Decoder, a generalized decoding model that can predict\npixel-level segmentation and language tokens seamlessly. X-Decodert takes as\ninput two types of queries: (i) generic non-semantic queries and (ii) semantic\nqueries induced from text inputs, to decode different pixel-level and\ntoken-level outputs in the same semantic space. With such a novel design,\nX-Decoder is the first work that provides a unified way to support all types of\nimage segmentation and a variety of vision-language (VL) tasks. Further, our\ndesign enables seamless interactions across tasks at different granularities\nand brings mutual benefits by learning a common and rich pixel-level\nvisual-semantic understanding space, without any pseudo-labeling. After\npretraining on a mixed set of a limited amount of segmentation data and\nmillions of image-text pairs, X-Decoder exhibits strong transferability to a\nwide range of downstream tasks in both zero-shot and finetuning settings.\nNotably, it achieves (1) state-of-the-art results on open-vocabulary\nsegmentation and referring segmentation on eight datasets; (2) better or\ncompetitive finetuned performance to other generalist and specialist models on\nsegmentation and VL tasks; and (3) flexibility for efficient finetuning and\nnovel task composition (e.g., referring captioning and image editing). Code,\ndemo, video, and visualization are available at https://x-decoder-vl.github.io.",
+                    "authors": "Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, Nanyun Peng, Lijuan Wang, Yong Jae Lee, Jianfeng Gao",
+                    "published": "2022-12-21",
+                    "updated": "2022-12-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.CL"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2311.01373v1",
+                    "title": "Recognize Any Regions",
+                    "abstract": "Understanding the semantics of individual regions or patches within\nunconstrained images, such as in open-world object detection, represents a\ncritical yet challenging task in computer vision. Building on the success of\npowerful image-level vision-language (ViL) foundation models like CLIP, recent\nefforts have sought to harness their capabilities by either training a\ncontrastive model from scratch with an extensive collection of region-label\npairs or aligning the outputs of a detection model with image-level\nrepresentations of region proposals. Despite notable progress, these approaches\nare plagued by computationally intensive training requirements, susceptibility\nto data noise, and deficiency in contextual information. To address these\nlimitations, we explore the synergistic potential of off-the-shelf foundation\nmodels, leveraging their respective strengths in localization and semantics. We\nintroduce a novel, generic, and efficient region recognition architecture,\nnamed RegionSpot, designed to integrate position-aware localization knowledge\nfrom a localization foundation model (e.g., SAM) with semantic information\nextracted from a ViL model (e.g., CLIP). To fully exploit pretrained knowledge\nwhile minimizing training overhead, we keep both foundation models frozen,\nfocusing optimization efforts solely on a lightweight attention-based knowledge\nintegration module. Through extensive experiments in the context of open-world\nobject recognition, our RegionSpot demonstrates significant performance\nimprovements over prior alternatives, while also providing substantial\ncomputational savings. For instance, training our model with 3 million data in\na single day using 8 V100 GPUs. Our model outperforms GLIP by 6.5 % in mean\naverage precision (mAP), with an even larger margin by 14.8 % for more\nchallenging and rare categories.",
+                    "authors": "Haosen Yang, Chuofan Ma, Bin Wen, Yi Jiang, Zehuan Yuan, Xiatian Zhu",
+                    "published": "2023-11-02",
+                    "updated": "2023-11-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2212.00794v2",
+                    "title": "Scaling Language-Image Pre-training via Masking",
+                    "abstract": "We present Fast Language-Image Pre-training (FLIP), a simple and more\nefficient method for training CLIP. Our method randomly masks out and removes a\nlarge portion of image patches during training. Masking allows us to learn from\nmore image-text pairs given the same wall-clock time and contrast more samples\nper iteration with similar memory footprint. It leads to a favorable trade-off\nbetween accuracy and training time. In our experiments on 400 million\nimage-text pairs, FLIP improves both accuracy and speed over the no-masking\nbaseline. On a large diversity of downstream tasks, FLIP dominantly outperforms\nthe CLIP counterparts trained on the same data. Facilitated by the speedup, we\nexplore the scaling behavior of increasing the model size, data size, or\ntraining length, and report encouraging results and comparisons. We hope that\nour work will foster future research on scaling vision-language learning.",
+                    "authors": "Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, Kaiming He",
+                    "published": "2022-12-01",
+                    "updated": "2023-03-30",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2204.14198v2",
+                    "title": "Flamingo: a Visual Language Model for Few-Shot Learning",
+                    "abstract": "Building models that can be rapidly adapted to novel tasks using only a\nhandful of annotated examples is an open challenge for multimodal machine\nlearning research. We introduce Flamingo, a family of Visual Language Models\n(VLM) with this ability. We propose key architectural innovations to: (i)\nbridge powerful pretrained vision-only and language-only models, (ii) handle\nsequences of arbitrarily interleaved visual and textual data, and (iii)\nseamlessly ingest images or videos as inputs. Thanks to their flexibility,\nFlamingo models can be trained on large-scale multimodal web corpora containing\narbitrarily interleaved text and images, which is key to endow them with\nin-context few-shot learning capabilities. We perform a thorough evaluation of\nour models, exploring and measuring their ability to rapidly adapt to a variety\nof image and video tasks. These include open-ended tasks such as visual\nquestion-answering, where the model is prompted with a question which it has to\nanswer; captioning tasks, which evaluate the ability to describe a scene or an\nevent; and close-ended tasks such as multiple-choice visual question-answering.\nFor tasks lying anywhere on this spectrum, a single Flamingo model can achieve\na new state of the art with few-shot learning, simply by prompting the model\nwith task-specific examples. On numerous benchmarks, Flamingo outperforms\nmodels fine-tuned on thousands of times more task-specific data.",
+                    "authors": "Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, Karen Simonyan",
+                    "published": "2022-04-29",
+                    "updated": "2022-11-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2112.12143v2",
+                    "title": "Scaling Open-Vocabulary Image Segmentation with Image-Level Labels",
+                    "abstract": "We design an open-vocabulary image segmentation model to organize an image\ninto meaningful regions indicated by arbitrary texts. Recent works (CLIP and\nALIGN), despite attaining impressive open-vocabulary classification accuracy\nwith image-level caption labels, are unable to segment visual concepts with\npixels. We argue that these models miss an important step of visual grouping,\nwhich organizes pixels into groups before learning visual-semantic alignments.\nWe propose OpenSeg to address the above issue while still making use of\nscalable image-level supervision of captions. First, it learns to propose\nsegmentation masks for possible organizations. Then it learns visual-semantic\nalignments by aligning each word in a caption to one or a few predicted masks.\nWe find the mask representations are the key to support learning image\nsegmentation from captions, making it possible to scale up the dataset and\nvocabulary sizes. OpenSeg significantly outperforms the recent open-vocabulary\nmethod of LSeg by +19.9 mIoU on PASCAL dataset, thanks to its scalability.",
+                    "authors": "Golnaz Ghiasi, Xiuye Gu, Yin Cui, Tsung-Yi Lin",
+                    "published": "2021-12-22",
+                    "updated": "2022-07-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.15389v1",
+                    "title": "EVA-CLIP: Improved Training Techniques for CLIP at Scale",
+                    "abstract": "Contrastive language-image pre-training, CLIP for short, has gained\nincreasing attention for its potential in various scenarios. In this paper, we\npropose EVA-CLIP, a series of models that significantly improve the efficiency\nand effectiveness of CLIP training. Our approach incorporates new techniques\nfor representation learning, optimization, and augmentation, enabling EVA-CLIP\nto achieve superior performance compared to previous CLIP models with the same\nnumber of parameters but significantly smaller training costs. Notably, our\nlargest 5.0B-parameter EVA-02-CLIP-E/14+ with only 9 billion seen samples\nachieves 82.0 zero-shot top-1 accuracy on ImageNet-1K val. A smaller\nEVA-02-CLIP-L/14+ with only 430 million parameters and 6 billion seen samples\nachieves 80.4 zero-shot top-1 accuracy on ImageNet-1K val. To facilitate open\naccess and open research, we release the complete suite of EVA-CLIP to the\ncommunity at https://github.com/baaivision/EVA/tree/master/EVA-CLIP.",
+                    "authors": "Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, Yue Cao",
+                    "published": "2023-03-27",
+                    "updated": "2023-03-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.02689v2",
+                    "title": "YOLACT: Real-time Instance Segmentation",
+                    "abstract": "We present a simple, fully-convolutional model for real-time instance\nsegmentation that achieves 29.8 mAP on MS COCO at 33.5 fps evaluated on a\nsingle Titan Xp, which is significantly faster than any previous competitive\napproach. Moreover, we obtain this result after training on only one GPU. We\naccomplish this by breaking instance segmentation into two parallel subtasks:\n(1) generating a set of prototype masks and (2) predicting per-instance mask\ncoefficients. Then we produce instance masks by linearly combining the\nprototypes with the mask coefficients. We find that because this process\ndoesn't depend on repooling, this approach produces very high-quality masks and\nexhibits temporal stability for free. Furthermore, we analyze the emergent\nbehavior of our prototypes and show they learn to localize instances on their\nown in a translation variant manner, despite being fully-convolutional.\nFinally, we also propose Fast NMS, a drop-in 12 ms faster replacement for\nstandard NMS that only has a marginal performance penalty.",
+                    "authors": "Daniel Bolya, Chong Zhou, Fanyi Xiao, Yong Jae Lee",
+                    "published": "2019-04-04",
+                    "updated": "2019-10-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.15308v2",
+                    "title": "SAM-CLIP: Merging Vision Foundation Models towards Semantic and Spatial Understanding",
+                    "abstract": "The landscape of publicly available vision foundation models (VFMs), such as\nCLIP and Segment Anything Model (SAM), is expanding rapidly. VFMs are endowed\nwith distinct capabilities stemming from their pre-training objectives. For\ninstance, CLIP excels in semantic understanding, while SAM specializes in\nspatial understanding for segmentation. In this work, we introduce a simple\nrecipe to efficiently merge VFMs into a unified model that absorbs their\nexpertise. Our method integrates techniques of multi-task learning, continual\nlearning, and distillation. Further, it demands significantly less\ncomputational cost compared to traditional multi-task training from scratch,\nand it only needs a small fraction of the pre-training datasets that were\ninitially used to train individual models. By applying our method to SAM and\nCLIP, we obtain SAM-CLIP: a unified model that combines the capabilities of SAM\nand CLIP into a single vision transformer. Compared with deploying SAM and CLIP\nindependently, our merged model, SAM-CLIP, reduces storage and compute costs\nfor inference, making it well-suited for edge device applications. We show that\nSAM-CLIP not only retains the foundational strengths of SAM and CLIP, but also\nintroduces synergistic functionalities, notably in zero-shot semantic\nsegmentation, where SAM-CLIP establishes new state-of-the-art results on 5\nbenchmarks. It outperforms previous models that are specifically designed for\nthis task by a large margin, including +6.8% and +5.9% mean IoU improvement on\nPascal-VOC and COCO-Stuff datasets, respectively.",
+                    "authors": "Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, Hadi Pouransari",
+                    "published": "2023-10-23",
+                    "updated": "2023-11-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.03273v1",
+                    "title": "Semantic Segmentation using Vision Transformers: A survey",
+                    "abstract": "Semantic segmentation has a broad range of applications in a variety of\ndomains including land coverage analysis, autonomous driving, and medical image\nanalysis. Convolutional neural networks (CNN) and Vision Transformers (ViTs)\nprovide the architecture models for semantic segmentation. Even though ViTs\nhave proven success in image classification, they cannot be directly applied to\ndense prediction tasks such as image segmentation and object detection since\nViT is not a general purpose backbone due to its patch partitioning scheme. In\nthis survey, we discuss some of the different ViT architectures that can be\nused for semantic segmentation and how their evolution managed the above-stated\nchallenge. The rise of ViT and its performance with a high success rate\nmotivated the community to slowly replace the traditional convolutional neural\nnetworks in various computer vision tasks. This survey aims to review and\ncompare the performances of ViT architectures designed for semantic\nsegmentation using benchmarking datasets. This will be worthwhile for the\ncommunity to yield knowledge regarding the implementations carried out in\nsemantic segmentation and to discover more efficient methodologies using ViTs.",
+                    "authors": "Hans Thisanke, Chamli Deshan, Kavindu Chamith, Sachith Seneviratne, Rajith Vidanaarachchi, Damayanthi Herath",
+                    "published": "2023-05-05",
+                    "updated": "2023-05-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.03983v1",
+                    "title": "Weakly Supervised Semantic Segmentation of Satellite Images",
+                    "abstract": "When one wants to train a neural network to perform semantic segmentation,\ncreating pixel-level annotations for each of the images in the database is a\ntedious task. If he works with aerial or satellite images, which are usually\nvery large, it is even worse. With that in mind, we investigate how to use\nimage-level annotations in order to perform semantic segmentation. Image-level\nannotations are much less expensive to acquire than pixel-level annotations,\nbut we lose a lot of information for the training of the model. From the\nannotations of the images, the model must find by itself how to classify the\ndifferent regions of the image. In this work, we use the method proposed by Anh\nand Kwak [1] to produce pixel-level annotation from image level annotation. We\ncompare the overall quality of our generated dataset with the original dataset.\nIn addition, we propose an adaptation of the AffinityNet that allows us to\ndirectly perform a semantic segmentation. Our results show that the generated\nlabels lead to the same performances for the training of several segmentation\nnetworks. Also, the quality of semantic segmentation performed directly by the\nAffinityNet and the Random Walk is close to the one of the best\nfully-supervised approaches.",
+                    "authors": "Adrien Nivaggioli, Hicham Randrianarivo",
+                    "published": "2019-04-08",
+                    "updated": "2019-04-08",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1502.04983v1",
+                    "title": "Context Tricks for Cheap Semantic Segmentation",
+                    "abstract": "Accurate semantic labeling of image pixels is difficult because intra-class\nvariability is often greater than inter-class variability. In turn, fast\nsemantic segmentation is hard because accurate models are usually too\ncomplicated to also run quickly at test-time. Our experience with building and\nrunning semantic segmentation systems has also shown a reasonably obvious\nbottleneck on model complexity, imposed by small training datasets. We\ntherefore propose two simple complementary strategies that leverage context to\ngive better semantic segmentation, while scaling up or down to train on\ndifferent-sized datasets.\n  As easy modifications for existing semantic segmentation algorithms, we\nintroduce Decorrelated Semantic Texton Forests, and the Context Sensitive Image\nLevel Prior. The proposed modifications are tested using a Semantic Texton\nForest (STF) system, and the modifications are validated on two standard\nbenchmark datasets, MSRC-21 and PascalVOC-2010. In Python based comparisons,\nour system is insignificantly slower than STF at test-time, yet produces\nsuperior semantic segmentations overall, with just push-button training.",
+                    "authors": "Thanapong Intharah, Gabriel J. Brostow",
+                    "published": "2015-02-17",
+                    "updated": "2015-02-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.13697v1",
+                    "title": "Generalizable Semantic Vision Query Generation for Zero-shot Panoptic and Semantic Segmentation",
+                    "abstract": "Zero-shot Panoptic Segmentation (ZPS) aims to recognize foreground instances\nand background stuff without images containing unseen categories in training.\nDue to the visual data sparsity and the difficulty of generalizing from seen to\nunseen categories, this task remains challenging. To better generalize to\nunseen classes, we propose Conditional tOken aligNment and Cycle trAnsiTion\n(CONCAT), to produce generalizable semantic vision queries. First, a feature\nextractor is trained by CON to link the vision and semantics for providing\ntarget queries. Formally, CON is proposed to align the semantic queries with\nthe CLIP visual CLS token extracted from complete and masked images. To address\nthe lack of unseen categories, a generator is required. However, one of the\ngaps in synthesizing pseudo vision queries, ie, vision queries for unseen\ncategories, is describing fine-grained visual details through semantic\nembeddings. Therefore, we approach CAT to train the generator in\nsemantic-vision and vision-semantic manners. In semantic-vision, visual query\ncontrast is proposed to model the high granularity of vision by pulling the\npseudo vision queries with the corresponding targets containing segments while\npushing those without segments away. To ensure the generated queries retain\nsemantic information, in vision-semantic, the pseudo vision queries are mapped\nback to semantic and supervised by real semantic embeddings. Experiments on ZPS\nachieve a 5.2% hPQ increase surpassing SOTA. We also examine inductive ZPS and\nopen-vocabulary semantic segmentation and obtain comparative results while\nbeing 2 times faster in testing.",
+                    "authors": "Jialei Chen, Daisuke Deguchi, Chenkai Zhang, Hiroshi Murase",
+                    "published": "2024-02-21",
+                    "updated": "2024-02-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.05490v1",
+                    "title": "Learning Semantic Segmentation with Query Points Supervision on Aerial Images",
+                    "abstract": "Semantic segmentation is crucial in remote sensing, where high-resolution\nsatellite images are segmented into meaningful regions. Recent advancements in\ndeep learning have significantly improved satellite image segmentation.\nHowever, most of these methods are typically trained in fully supervised\nsettings that require high-quality pixel-level annotations, which are expensive\nand time-consuming to obtain. In this work, we present a weakly supervised\nlearning algorithm to train semantic segmentation algorithms that only rely on\nquery point annotations instead of full mask labels. Our proposed approach\nperforms accurate semantic segmentation and improves efficiency by\nsignificantly reducing the cost and time required for manual annotation.\nSpecifically, we generate superpixels and extend the query point labels into\nthose superpixels that group similar meaningful semantics. Then, we train\nsemantic segmentation models, supervised with images partially labeled with the\nsuperpixels pseudo-labels. We benchmark our weakly supervised training approach\non an aerial image dataset and different semantic segmentation architectures,\nshowing that we can reach competitive performance compared to fully supervised\ntraining while reducing the annotation effort.",
+                    "authors": "Santiago Rivier, Carlos Hinojosa, Silvio Giancola, Bernard Ghanem",
+                    "published": "2023-09-11",
+                    "updated": "2023-09-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1809.10198v1",
+                    "title": "Recent progress in semantic image segmentation",
+                    "abstract": "Semantic image segmentation, which becomes one of the key applications in\nimage processing and computer vision domain, has been used in multiple domains\nsuch as medical area and intelligent transportation. Lots of benchmark datasets\nare released for researchers to verify their algorithms. Semantic segmentation\nhas been studied for many years. Since the emergence of Deep Neural Network\n(DNN), segmentation has made a tremendous progress. In this paper, we divide\nsemantic image segmentation methods into two categories: traditional and recent\nDNN method. Firstly, we briefly summarize the traditional method as well as\ndatasets released for segmentation, then we comprehensively investigate recent\nmethods based on DNN which are described in the eight aspects: fully\nconvolutional network, upsample ways, FCN joint with CRF methods, dilated\nconvolution approaches, progresses in backbone network, pyramid methods,\nMulti-level feature and multi-stage method, supervised, weakly-supervised and\nunsupervised methods. Finally, a conclusion in this area is drawn.",
+                    "authors": "Xiaolong Liu, Zhidong Deng, Yuhan Yang",
+                    "published": "2018-09-20",
+                    "updated": "2018-09-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.13978v3",
+                    "title": "Personalized Image Semantic Segmentation",
+                    "abstract": "Semantic segmentation models trained on public datasets have achieved great\nsuccess in recent years. However, these models didn't consider the\npersonalization issue of segmentation though it is important in practice. In\nthis paper, we address the problem of personalized image segmentation. The\nobjective is to generate more accurate segmentation results on unlabeled\npersonalized images by investigating the data's personalized traits. To open up\nfuture research in this area, we collect a large dataset containing various\nusers' personalized images called PIS (Personalized Image Semantic\nSegmentation). We also survey some recent researches related to this problem\nand report their performance on our dataset. Furthermore, by observing the\ncorrelation among a user's personalized images, we propose a baseline method\nthat incorporates the inter-image context when segmenting certain images.\nExtensive experiments show that our method outperforms the existing methods on\nthe proposed dataset. The code and the PIS dataset will be made publicly\navailable.",
+                    "authors": "Yu Zhang, Chang-Bin Zhang, Peng-Tao Jiang, Ming-Ming Cheng, Feng Mao",
+                    "published": "2021-07-24",
+                    "updated": "2021-09-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.07898v5",
+                    "title": "ISLE: A Framework for Image Level Semantic Segmentation Ensemble",
+                    "abstract": "One key bottleneck of employing state-of-the-art semantic segmentation\nnetworks in the real world is the availability of training labels. Conventional\nsemantic segmentation networks require massive pixel-wise annotated labels to\nreach state-of-the-art prediction quality. Hence, several works focus on\nsemantic segmentation networks trained with only image-level annotations.\nHowever, when scrutinizing the results of state-of-the-art in more detail, we\nnotice that they are remarkably close to each other on average prediction\nquality, different approaches perform better in different classes while\nproviding low quality in others. To address this problem, we propose a novel\nframework, ISLE, which employs an ensemble of the \"pseudo-labels\" for a given\nset of different semantic segmentation techniques on a class-wise level.\nPseudo-labels are the pixel-wise predictions of the image-level semantic\nsegmentation frameworks used to train the final segmentation model. Our\npseudo-labels seamlessly combine the strong points of multiple segmentation\ntechniques approaches to reach superior prediction quality. We reach up to 2.4%\nimprovement over ISLE's individual components. An exhaustive analysis was\nperformed to demonstrate ISLE's effectiveness over state-of-the-art frameworks\nfor image-level semantic segmentation.",
+                    "authors": "Erik Ostrowski, Muhammad Shafique",
+                    "published": "2023-03-14",
+                    "updated": "2023-09-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.00487v1",
+                    "title": "Linear Semantics in Generative Adversarial Networks",
+                    "abstract": "Generative Adversarial Networks (GANs) are able to generate high-quality\nimages, but it remains difficult to explicitly specify the semantics of\nsynthesized images. In this work, we aim to better understand the semantic\nrepresentation of GANs, and thereby enable semantic control in GAN's generation\nprocess. Interestingly, we find that a well-trained GAN encodes image semantics\nin its internal feature maps in a surprisingly simple way: a linear\ntransformation of feature maps suffices to extract the generated image\nsemantics. To verify this simplicity, we conduct extensive experiments on\nvarious GANs and datasets; and thanks to this simplicity, we are able to learn\na semantic segmentation model for a trained GAN from a small number (e.g., 8)\nof labeled images. Last but not least, leveraging our findings, we propose two\nfew-shot image editing approaches, namely Semantic-Conditional Sampling and\nSemantic Image Editing. Given a trained GAN and as few as eight semantic\nannotations, the user is able to generate diverse images subject to a\nuser-provided semantic layout, and control the synthesized image semantics. We\nhave made the code publicly available.",
+                    "authors": "Jianjin Xu, Changxi Zheng",
+                    "published": "2021-04-01",
+                    "updated": "2021-04-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.04618v1",
+                    "title": "Multi-Scale Semantic Segmentation with Modified MBConv Blocks",
+                    "abstract": "Recently, MBConv blocks, initially designed for efficiency in\nresource-limited settings and later adapted for cutting-edge image\nclassification performances, have demonstrated significant potential in image\nclassification tasks. Despite their success, their application in semantic\nsegmentation has remained relatively unexplored. This paper introduces a novel\nadaptation of MBConv blocks specifically tailored for semantic segmentation.\nOur modification stems from the insight that semantic segmentation requires the\nextraction of more detailed spatial information than image classification. We\nargue that to effectively perform multi-scale semantic segmentation, each\nbranch of a U-Net architecture, regardless of its resolution, should possess\nequivalent segmentation capabilities. By implementing these changes, our\napproach achieves impressive mean Intersection over Union (IoU) scores of 84.5%\nand 84.0% on the Cityscapes test and validation datasets, respectively,\ndemonstrating the efficacy of our proposed modifications in enhancing semantic\nsegmentation performance.",
+                    "authors": "Xi Chen, Yang Cai, Yuan Wu, Bo Xiong, Taesung Park",
+                    "published": "2024-02-07",
+                    "updated": "2024-02-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2110.04487v1",
+                    "title": "Colour augmentation for improved semi-supervised semantic segmentation",
+                    "abstract": "Consistency regularization describes a class of approaches that have yielded\nstate-of-the-art results for semi-supervised classification. While\nsemi-supervised semantic segmentation proved to be more challenging, a number\nof successful approaches have been recently proposed. Recent work explored the\nchallenges involved in using consistency regularization for segmentation\nproblems. In their self-supervised work Chen et al. found that colour\naugmentation prevents a classification network from using image colour\nstatistics as a short-cut for self-supervised learning via instance\ndiscrimination. Drawing inspiration from this we find that a similar problem\nimpedes semi-supervised semantic segmentation and offer colour augmentation as\na solution, improving semi-supervised semantic segmentation performance on\nchallenging photographic imagery.",
+                    "authors": "Geoff French, Michal Mackiewicz",
+                    "published": "2021-10-09",
+                    "updated": "2021-10-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2211.08352v1",
+                    "title": "Visual Semantic Segmentation Based on Few/Zero-Shot Learning: An Overview",
+                    "abstract": "Visual semantic segmentation aims at separating a visual sample into diverse\nblocks with specific semantic attributes and identifying the category for each\nblock, and it plays a crucial role in environmental perception. Conventional\nlearning-based visual semantic segmentation approaches count heavily on\nlarge-scale training data with dense annotations and consistently fail to\nestimate accurate semantic labels for unseen categories. This obstruction spurs\na craze for studying visual semantic segmentation with the assistance of\nfew/zero-shot learning. The emergence and rapid progress of few/zero-shot\nvisual semantic segmentation make it possible to learn unseen-category from a\nfew labeled or zero-labeled samples, which advances the extension to practical\napplications. Therefore, this paper focuses on the recently published\nfew/zero-shot visual semantic segmentation methods varying from 2D to 3D space\nand explores the commonalities and discrepancies of technical settlements under\ndifferent segmentation circumstances. Specifically, the preliminaries on\nfew/zero-shot visual semantic segmentation, including the problem definitions,\ntypical datasets, and technical remedies, are briefly reviewed and discussed.\nMoreover, three typical instantiations are involved to uncover the interactions\nof few/zero-shot learning with visual semantic segmentation, including image\nsemantic segmentation, video object segmentation, and 3D segmentation. Finally,\nthe future challenges of few/zero-shot visual semantic segmentation are\ndiscussed.",
+                    "authors": "Wenqi Ren, Yang Tang, Qiyu Sun, Chaoqiang Zhao, Qing-Long Han",
+                    "published": "2022-11-13",
+                    "updated": "2022-11-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2101.09642v2",
+                    "title": "Image Compression with Encoder-Decoder Matched Semantic Segmentation",
+                    "abstract": "In recent years, layered image compression is demonstrated to be a promising\ndirection, which encodes a compact representation of the input image and apply\nan up-sampling network to reconstruct the image. To further improve the quality\nof the reconstructed image, some works transmit the semantic segment together\nwith the compressed image data. Consequently, the compression ratio is also\ndecreased because extra bits are required for transmitting the semantic\nsegment. To solve this problem, we propose a new layered image compression\nframework with encoder-decoder matched semantic segmentation (EDMS). And then,\nfollowed by the semantic segmentation, a special convolution neural network is\nused to enhance the inaccurate semantic segment. As a result, the accurate\nsemantic segment can be obtained in the decoder without requiring extra bits.\nThe experimental results show that the proposed EDMS framework can get up to\n35.31% BD-rate reduction over the HEVC-based (BPG) codec, 5% bitrate, and 24%\nencoding time saving compare to the state-of-the-art semantic-based image\ncodec.",
+                    "authors": "Trinh Man Hoang, Jinjia Zhou, Yibo Fan",
+                    "published": "2021-01-24",
+                    "updated": "2021-01-30",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV",
+                        "cs.MM"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.11857v1",
+                    "title": "Joint Learning of Intrinsic Images and Semantic Segmentation",
+                    "abstract": "Semantic segmentation of outdoor scenes is problematic when there are\nvariations in imaging conditions. It is known that albedo (reflectance) is\ninvariant to all kinds of illumination effects. Thus, using reflectance images\nfor semantic segmentation task can be favorable. Additionally, not only\nsegmentation may benefit from reflectance, but also segmentation may be useful\nfor reflectance computation. Therefore, in this paper, the tasks of semantic\nsegmentation and intrinsic image decomposition are considered as a combined\nprocess by exploring their mutual relationship in a joint fashion. To that end,\nwe propose a supervised end-to-end CNN architecture to jointly learn intrinsic\nimage decomposition and semantic segmentation. We analyze the gains of\naddressing those two problems jointly. Moreover, new cascade CNN architectures\nfor intrinsic-for-segmentation and segmentation-for-intrinsic are proposed as\nsingle tasks. Furthermore, a dataset of 35K synthetic images of natural\nenvironments is created with corresponding albedo and shading (intrinsics), as\nwell as semantic labels (segmentation) assigned to each object/scene. The\nexperiments show that joint learning of intrinsic image decomposition and\nsemantic segmentation is beneficial for both tasks for natural scenes. Dataset\nand models are available at: https://ivi.fnwi.uva.nl/cv/intrinseg",
+                    "authors": "Anil S. Baslamisli, Thomas T. Groenestege, Partha Das, Hoang-An Le, Sezer Karaoglu, Theo Gevers",
+                    "published": "2018-07-31",
+                    "updated": "2018-07-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.01369v2",
+                    "title": "Exploring Limits of Diffusion-Synthetic Training with Weakly Supervised Semantic Segmentation",
+                    "abstract": "The advance of generative models for images has inspired various training\ntechniques for image recognition utilizing synthetic images. In semantic\nsegmentation, one promising approach is extracting pseudo-masks from attention\nmaps in text-to-image diffusion models, which enables\nreal-image-and-annotation-free training. However, the pioneering training\nmethod using the diffusion-synthetic images and pseudo-masks, i.e., DiffuMask\nhas limitations in terms of mask quality, scalability, and ranges of applicable\ndomains. To overcome these limitations, this work introduces three techniques\nfor diffusion-synthetic semantic segmentation training. First,\nreliability-aware robust training, originally used in weakly supervised\nlearning, helps segmentation with insufficient synthetic mask quality. %Second,\nlarge-scale pretraining of whole segmentation models, not only backbones, on\nsynthetic ImageNet-1k-class images with pixel-labels benefits downstream\nsegmentation tasks. Second, we introduce prompt augmentation, data augmentation\nto the prompt text set to scale up and diversify training images with a limited\ntext resources. Finally, LoRA-based adaptation of Stable Diffusion enables the\ntransfer to a distant domain, e.g., auto-driving images. Experiments in PASCAL\nVOC, ImageNet-S, and Cityscapes show that our method effectively closes gap\nbetween real and synthetic training in semantic segmentation.",
+                    "authors": "Ryota Yoshihashi, Yuya Otsuka, Kenji Doi, Tomohiro Tanaka, Hirokatsu Kataoka",
+                    "published": "2023-09-04",
+                    "updated": "2024-04-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2201.05869v2",
+                    "title": "Prototype Guided Network for Anomaly Segmentation",
+                    "abstract": "Semantic segmentation methods can not directly identify abnormal objects in\nimages. Anomaly Segmentation algorithm from this realistic setting can\ndistinguish between in-distribution objects and Out-Of-Distribution (OOD)\nobjects and output the anomaly probability for pixels. In this paper, a\nPrototype Guided Anomaly segmentation Network (PGAN) is proposed to extract\nsemantic prototypes for in-distribution training data from limited annotated\nimages. In the model, prototypes are used to model the hierarchical category\nsemantic information and distinguish OOD pixels. The proposed PGAN model\nincludes a semantic segmentation network and a prototype extraction network.\nSimilarity measures are adopted to optimize the prototypes. The learned\nsemantic prototypes are used as category semantics to compare the similarity\nwith features extracted from test images and then to generate semantic\nsegmentation prediction. The proposed prototype extraction network can also be\nintegrated into most semantic segmentation networks and recognize OOD pixels.\nOn the StreetHazards dataset, the proposed PGAN model produced mIoU of 53.4%\nfor anomaly segmentation. The experimental results demonstrate PGAN may achieve\nthe SOTA performance in the anomaly segmentation tasks.",
+                    "authors": "Yiqing Hao, Yi Jin, Gaoyun An",
+                    "published": "2022-01-15",
+                    "updated": "2022-03-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2007.02361v1",
+                    "title": "Self-supervised Depth Estimation to Regularise Semantic Segmentation in Knee Arthroscopy",
+                    "abstract": "Intra-operative automatic semantic segmentation of knee joint structures can\nassist surgeons during knee arthroscopy in terms of situational awareness.\nHowever, due to poor imaging conditions (e.g., low texture, overexposure,\netc.), automatic semantic segmentation is a challenging scenario, which\njustifies the scarce literature on this topic. In this paper, we propose a\nnovel self-supervised monocular depth estimation to regularise the training of\nthe semantic segmentation in knee arthroscopy. To further regularise the depth\nestimation, we propose the use of clean training images captured by the stereo\narthroscope of routine objects (presenting none of the poor imaging conditions\nand with rich texture information) to pre-train the model. We fine-tune such\nmodel to produce both the semantic segmentation and self-supervised monocular\ndepth using stereo arthroscopic images taken from inside the knee. Using a data\nset containing 3868 arthroscopic images captured during cadaveric knee\narthroscopy with semantic segmentation annotations, 2000 stereo image pairs of\ncadaveric knee arthroscopy, and 2150 stereo image pairs of routine objects, we\nshow that our semantic segmentation regularised by self-supervised depth\nestimation produces a more accurate segmentation than a state-of-the-art\nsemantic segmentation approach modeled exclusively with semantic segmentation\nannotation.",
+                    "authors": "Fengbei Liu, Yaqub Jonmohamadi, Gabriel Maicas, Ajay K. Pandey, Gustavo Carneiro",
+                    "published": "2020-07-05",
+                    "updated": "2020-07-05",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2004.10126v3",
+                    "title": "Generative Synthetic Augmentation using Label-to-Image Translation for Nuclei Image Segmentation",
+                    "abstract": "In medical image diagnosis, pathology image analysis using semantic\nsegmentation becomes important for efficient screening as a field of digital\npathology. The spatial augmentation is ordinary used for semantic segmentation.\nTumor images under malignant are rare and to annotate the labels of nuclei\nregion takes much time-consuming. We require an effective use of dataset to\nmaximize the segmentation accuracy. It is expected that some augmentation to\ntransform generalized images influence the segmentation performance. We propose\na synthetic augmentation using label-to-image translation, mapping from a\nsemantic label with the edge structure to a real image. Exactly this paper deal\nwith stain slides of nuclei in tumor. Actually, we demonstrate several\nsegmentation algorithms applied to the initial dataset that contains real\nimages and labels using synthetic augmentation in order to add their\ngeneralized images. We computes and reports that a proposed synthetic\naugmentation procedure improve their accuracy.",
+                    "authors": "Takato Yasuno",
+                    "published": "2020-04-21",
+                    "updated": "2021-03-02",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "eess.IV",
+                        "stat.ML",
+                        "I.4.6; I.2.6"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2106.04108v3",
+                    "title": "Fully Transformer Networks for Semantic Image Segmentation",
+                    "abstract": "Transformers have shown impressive performance in various natural language\nprocessing and computer vision tasks, due to the capability of modeling\nlong-range dependencies. Recent progress has demonstrated that combining such\nTransformers with CNN-based semantic image segmentation models is very\npromising. However, it is not well studied yet on how well a pure Transformer\nbased approach can achieve for image segmentation. In this work, we explore a\nnovel framework for semantic image segmentation, which is encoder-decoder based\nFully Transformer Networks (FTN). Specifically, we first propose a Pyramid\nGroup Transformer (PGT) as the encoder for progressively learning hierarchical\nfeatures, meanwhile reducing the computation complexity of the standard Visual\nTransformer (ViT). Then, we propose a Feature Pyramid Transformer (FPT) to fuse\nsemantic-level and spatial-level information from multiple levels of the PGT\nencoder for semantic image segmentation. Surprisingly, this simple baseline can\nachieve better results on multiple challenging semantic segmentation and face\nparsing benchmarks, including PASCAL Context, ADE20K, COCOStuff, and\nCelebAMask-HQ. The source code will be released on\nhttps://github.com/BR-IDL/PaddleViT.",
+                    "authors": "Sitong Wu, Tianyi Wu, Fangjian Lin, Shengwei Tian, Guodong Guo",
+                    "published": "2021-06-08",
+                    "updated": "2021-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.00679v3",
+                    "title": "Cooperative Semantic Segmentation and Image Restoration in Adverse Environmental Conditions",
+                    "abstract": "Most state-of-the-art semantic segmentation approaches only achieve high\naccuracy in good conditions. In practically-common but less-discussed adverse\nenvironmental conditions, their performance can decrease enormously. Existing\nstudies usually cast the handling of segmentation in adverse conditions as a\nseparate post-processing step after signal restoration, making the segmentation\nperformance largely depend on the quality of restoration. In this paper, we\npropose a novel deep-learning framework to tackle semantic segmentation and\nimage restoration in adverse environmental conditions in a holistic manner. The\nproposed approach contains two components: Semantically-Guided Adaptation,\nwhich exploits semantic information from degraded images to refine the\nsegmentation; and Exemplar-Guided Synthesis, which restores images from\nsemantic label maps given degraded exemplars as the guidance. Our method\ncooperatively leverages the complementarity and interdependence of low-level\nrestoration and high-level segmentation in adverse environmental conditions.\nExtensive experiments on various datasets demonstrate that our approach can not\nonly improve the accuracy of semantic segmentation with degradation cues, but\nalso boost the perceptual quality and structural similarity of image\nrestoration with semantic guidance.",
+                    "authors": "Weihao Xia, Zhanglin Cheng, Yujiu Yang, Jing-Hao Xue",
+                    "published": "2019-11-02",
+                    "updated": "2020-03-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2202.04754v2",
+                    "title": "Wireless Transmission of Images With The Assistance of Multi-level Semantic Information",
+                    "abstract": "Semantic-oriented communication has been considered as a promising to boost\nthe bandwidth efficiency by only transmitting the semantics of the data. In\nthis paper, we propose a multi-level semantic aware communication system for\nwireless image transmission, named MLSC-image, which is based on the deep\nlearning techniques and trained in an end to end manner. In particular, the\nproposed model includes a multilevel semantic feature extractor, that extracts\nboth the highlevel semantic information, such as the text semantics and the\nsegmentation semantics, and the low-level semantic information, such as local\nspatial details of the images. We employ a pretrained image caption to capture\nthe text semantics and a pretrained image segmentation model to obtain the\nsegmentation semantics. These high-level and low-level semantic features are\nthen combined and encoded by a joint semantic and channel encoder into symbols\nto transmit over the physical channel. The numerical results validate the\neffectiveness and efficiency of the proposed semantic communication system,\nespecially under the limited bandwidth condition, which indicates the\nadvantages of the high-level semantics in the compression of images.",
+                    "authors": "Zhenguo Zhang, Qianqian Yang, Shibo He, Mingyang Sun, Jiming Chen",
+                    "published": "2022-02-08",
+                    "updated": "2023-12-08",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.15608v1",
+                    "title": "Semantic Segmentation by Semantic Proportions",
+                    "abstract": "Semantic segmentation is a critical task in computer vision that aims to\nidentify and classify individual pixels in an image, with numerous applications\nfor example autonomous driving and medical image analysis. However, semantic\nsegmentation can be super challenging particularly due to the need for large\namounts of annotated data. Annotating images is a time-consuming and costly\nprocess, often requiring expert knowledge and significant effort. In this\npaper, we propose a novel approach for semantic segmentation by eliminating the\nneed of ground-truth segmentation maps. Instead, our approach requires only the\nrough information of individual semantic class proportions, shortened as\nsemantic proportions. It greatly simplifies the data annotation process and\nthus will significantly reduce the annotation time and cost, making it more\nfeasible for large-scale applications. Moreover, it opens up new possibilities\nfor semantic segmentation tasks where obtaining the full ground-truth\nsegmentation maps may not be feasible or practical. Extensive experimental\nresults demonstrate that our approach can achieve comparable and sometimes even\nbetter performance against the benchmark method that relies on the ground-truth\nsegmentation maps. Utilising semantic proportions suggested in this work offers\na promising direction for future research in the field of semantic\nsegmentation.",
+                    "authors": "Halil Ibrahim Aysel, Xiaohao Cai, Adam Pr\u00fcgel-Bennett",
+                    "published": "2023-05-24",
+                    "updated": "2023-05-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1804.04882v2",
+                    "title": "Learning to Exploit the Prior Network Knowledge for Weakly-Supervised Semantic Segmentation",
+                    "abstract": "Training a Convolutional Neural Network (CNN) for semantic segmentation\ntypically requires to collect a large amount of accurate pixel-level\nannotations, a hard and expensive task. In contrast, simple image tags are\neasier to gather. With this paper we introduce a novel weakly-supervised\nsemantic segmentation model able to learn from image labels, and just image\nlabels. Our model uses the prior knowledge of a network trained for image\nrecognition, employing these image annotations as an attention mechanism to\nidentify semantic regions in the images. We then present a methodology that\nbuilds accurate class-specific segmentation masks from these regions, where\nneither external objectness nor saliency algorithms are required. We describe\nhow to incorporate this mask generation strategy into a fully end-to-end\ntrainable process where the network jointly learns to classify and segment\nimages. Our experiments on PASCAL VOC 2012 dataset show that exploiting these\ngenerated class-specific masks in conjunction with our novel end-to-end\nlearning process outperforms several recent weakly-supervised semantic\nsegmentation methods that use image tags only, and even some models that\nleverage additional supervision or training data.",
+                    "authors": "Carolina Redondo-Cabrera, Marcos Baptista-R\u00edos, Roberto J. L\u00f3pez-Sastre",
+                    "published": "2018-04-13",
+                    "updated": "2019-02-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2304.11332v2",
+                    "title": "Input Augmentation with SAM: Boosting Medical Image Segmentation with Segmentation Foundation Model",
+                    "abstract": "The Segment Anything Model (SAM) is a recently developed large model for\ngeneral-purpose segmentation for computer vision tasks. SAM was trained using\n11 million images with over 1 billion masks and can produce segmentation\nresults for a wide range of objects in natural scene images. SAM can be viewed\nas a general perception model for segmentation (partitioning images into\nsemantically meaningful regions). Thus, how to utilize such a large foundation\nmodel for medical image segmentation is an emerging research target. This paper\nshows that although SAM does not immediately give high-quality segmentation for\nmedical image data, its generated masks, features, and stability scores are\nuseful for building and training better medical image segmentation models. In\nparticular, we demonstrate how to use SAM to augment image input for\ncommonly-used medical image segmentation models (e.g., U-Net). Experiments on\nthree segmentation tasks show the effectiveness of our proposed SAMAug method.\nThe code is available at \\url{https://github.com/yizhezhang2000/SAMAug}.",
+                    "authors": "Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Danny Z. Chen",
+                    "published": "2023-04-22",
+                    "updated": "2023-06-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.13297v5",
+                    "title": "GPS-GLASS: Learning Nighttime Semantic Segmentation Using Daytime Video and GPS data",
+                    "abstract": "Semantic segmentation for autonomous driving should be robust against various\nin-the-wild environments. Nighttime semantic segmentation is especially\nchallenging due to a lack of annotated nighttime images and a large domain gap\nfrom daytime images with sufficient annotation. In this paper, we propose a\nnovel GPS-based training framework for nighttime semantic segmentation. Given\nGPS-aligned pairs of daytime and nighttime images, we perform cross-domain\ncorrespondence matching to obtain pixel-level pseudo supervision. Moreover, we\nconduct flow estimation between daytime video frames and apply GPS-based\nscaling to acquire another pixel-level pseudo supervision. Using these pseudo\nsupervisions with a confidence map, we train a nighttime semantic segmentation\nnetwork without any annotation from nighttime images. Experimental results\ndemonstrate the effectiveness of the proposed method on several nighttime\nsemantic segmentation datasets. Our source code is available at\nhttps://github.com/jimmy9704/GPS-GLASS.",
+                    "authors": "Hongjae Lee, Changwoo Han, Jun-Sang Yoo, Seung-Won Jung",
+                    "published": "2022-07-27",
+                    "updated": "2023-08-18",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.04297v1",
+                    "title": "SHDM-NET: Heat Map Detail Guidance with Image Matting for Industrial Weld Semantic Segmentation Network",
+                    "abstract": "In actual industrial production, the assessment of the steel plate welding\neffect is an important task, and the segmentation of the weld section is the\nbasis of the assessment. This paper proposes an industrial weld segmentation\nnetwork based on a deep learning semantic segmentation algorithm fused with\nheatmap detail guidance and Image Matting to solve the automatic segmentation\nproblem of weld regions. In the existing semantic segmentation networks, the\nboundary information can be preserved by fusing the features of both high-level\nand low-level layers. However, this method can lead to insufficient expression\nof the spatial information in the low-level layer, resulting in inaccurate\nsegmentation boundary positioning. We propose a detailed guidance module based\non heatmaps to fully express the segmented region boundary information in the\nlow-level network to address this problem. Specifically, the expression of\nboundary information can be enhanced by adding a detailed branch to predict\nsegmented boundary and then matching it with the boundary heat map generated by\nmask labels to calculate the mean square error loss. In addition, although deep\nlearning has achieved great success in the field of semantic segmentation, the\nprecision of the segmentation boundary region is not high due to the loss of\ndetailed information caused by the classical segmentation network in the\nprocess of encoding and decoding process. This paper introduces a matting\nalgorithm to calibrate the boundary of the segmentation region of the semantic\nsegmentation network to solve this problem. Through many experiments on\nindustrial weld data sets, the effectiveness of our method is demonstrated, and\nthe MIOU reaches 97.93%. It is worth noting that this performance is comparable\nto human manual segmentation ( MIOU 97.96%).",
+                    "authors": "Qi Wang, Jingwu Mei",
+                    "published": "2022-07-09",
+                    "updated": "2022-07-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.04829v2",
+                    "title": "MixReorg: Cross-Modal Mixed Patch Reorganization is a Good Mask Learner for Open-World Semantic Segmentation",
+                    "abstract": "Recently, semantic segmentation models trained with image-level text\nsupervision have shown promising results in challenging open-world scenarios.\nHowever, these models still face difficulties in learning fine-grained semantic\nalignment at the pixel level and predicting accurate object masks. To address\nthis issue, we propose MixReorg, a novel and straightforward pre-training\nparadigm for semantic segmentation that enhances a model's ability to\nreorganize patches mixed across images, exploring both local visual relevance\nand global semantic coherence. Our approach involves generating fine-grained\npatch-text pairs data by mixing image patches while preserving the\ncorrespondence between patches and text. The model is then trained to minimize\nthe segmentation loss of the mixed images and the two contrastive losses of the\noriginal and restored features. With MixReorg as a mask learner, conventional\ntext-supervised semantic segmentation models can achieve highly generalizable\npixel-semantic alignment ability, which is crucial for open-world segmentation.\nAfter training with large-scale image-text data, MixReorg models can be applied\ndirectly to segment visual objects of arbitrary categories, without the need\nfor further fine-tuning. Our proposed framework demonstrates strong performance\non popular zero-shot semantic segmentation benchmarks, outperforming GroupViT\nby significant margins of 5.0%, 6.2%, 2.5%, and 3.4% mIoU on PASCAL VOC2012,\nPASCAL Context, MS COCO, and ADE20K, respectively.",
+                    "authors": "Kaixin Cai, Pengzhen Ren, Yi Zhu, Hang Xu, Jianzhuang Liu, Changlin Li, Guangrun Wang, Xiaodan Liang",
+                    "published": "2023-08-09",
+                    "updated": "2024-03-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1809.10245v1",
+                    "title": "Cylindrical Transform: 3D Semantic Segmentation of Kidneys With Limited Annotated Images",
+                    "abstract": "In this paper, we propose a novel technique for sampling sequential images\nusing a cylindrical transform in a cylindrical coordinate system for kidney\nsemantic segmentation in abdominal computed tomography (CT). The images\ngenerated from a cylindrical transform augment a limited annotated set of\nimages in three dimensions. This approach enables us to train contemporary\nclassification deep convolutional neural networks (DCNNs) instead of fully\nconvolutional networks (FCNs) for semantic segmentation. Typical semantic\nsegmentation models segment a sequential set of images (e.g. CT or video) by\nsegmenting each image independently. However, the proposed method not only\nconsiders the spatial dependency in the x-y plane, but also the spatial\nsequential dependency along the z-axis. The results show that classification\nDCNNs, trained on cylindrical transformed images, can achieve a higher\nsegmentation performance value than FCNs using a limited number of annotated\nimages.",
+                    "authors": "Hojjat Salehinejad, Sumeya Naqvi, Errol Colak, Joseph Barfett, Shahrokh Valaee",
+                    "published": "2018-09-24",
+                    "updated": "2018-09-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.NE"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.03212v2",
+                    "title": "Hierarchical Semantic Segmentation using Psychometric Learning",
+                    "abstract": "Assigning meaning to parts of image data is the goal of semantic image\nsegmentation. Machine learning methods, specifically supervised learning is\ncommonly used in a variety of tasks formulated as semantic segmentation. One of\nthe major challenges in the supervised learning approaches is expressing and\ncollecting the rich knowledge that experts have with respect to the meaning\npresent in the image data. Towards this, typically a fixed set of labels is\nspecified and experts are tasked with annotating the pixels, patches or\nsegments in the images with the given labels. In general, however, the set of\nclasses does not fully capture the rich semantic information present in the\nimages. For example, in medical imaging such as histology images, the different\nparts of cells could be grouped and sub-grouped based on the expertise of the\npathologist.\n  To achieve such a precise semantic representation of the concepts in the\nimage, we need access to the full depth of knowledge of the annotator. In this\nwork, we develop a novel approach to collect segmentation annotations from\nexperts based on psychometric testing. Our method consists of the psychometric\ntesting procedure, active query selection, query enhancement, and a deep metric\nlearning model to achieve a patch-level image embedding that allows for\nsemantic segmentation of images. We show the merits of our method with\nevaluation on the synthetically generated image, aerial image and histology\nimage.",
+                    "authors": "Lu Yin, Vlado Menkovski, Shiwei Liu, Mykola Pechenizkiy",
+                    "published": "2021-07-07",
+                    "updated": "2021-12-16",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.02095v1",
+                    "title": "Content-aware Token Sharing for Efficient Semantic Segmentation with Vision Transformers",
+                    "abstract": "This paper introduces Content-aware Token Sharing (CTS), a token reduction\napproach that improves the computational efficiency of semantic segmentation\nnetworks that use Vision Transformers (ViTs). Existing works have proposed\ntoken reduction approaches to improve the efficiency of ViT-based image\nclassification networks, but these methods are not directly applicable to\nsemantic segmentation, which we address in this work. We observe that, for\nsemantic segmentation, multiple image patches can share a token if they contain\nthe same semantic class, as they contain redundant information. Our approach\nleverages this by employing an efficient, class-agnostic policy network that\npredicts if image patches contain the same semantic class, and lets them share\na token if they do. With experiments, we explore the critical design choices of\nCTS and show its effectiveness on the ADE20K, Pascal Context and Cityscapes\ndatasets, various ViT backbones, and different segmentation decoders. With\nContent-aware Token Sharing, we are able to reduce the number of processed\ntokens by up to 44%, without diminishing the segmentation quality.",
+                    "authors": "Chenyang Lu, Daan de Geus, Gijs Dubbelman",
+                    "published": "2023-06-03",
+                    "updated": "2023-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.00242v1",
+                    "title": "3D Guided Weakly Supervised Semantic Segmentation",
+                    "abstract": "Pixel-wise clean annotation is necessary for fully-supervised semantic\nsegmentation, which is laborious and expensive to obtain. In this paper, we\npropose a weakly supervised 2D semantic segmentation model by incorporating\nsparse bounding box labels with available 3D information, which is much easier\nto obtain with advanced sensors. We manually labeled a subset of the 2D-3D\nSemantics(2D-3D-S) dataset with bounding boxes, and introduce our 2D-3D\ninference module to generate accurate pixel-wise segment proposal masks. Guided\nby 3D information, we first generate a point cloud of objects and calculate\nobjectness probability score for each point. Then we project the point cloud\nwith objectness probabilities back to 2D images followed by a refinement step\nto obtain segment proposals, which are treated as pseudo labels to train a\nsemantic segmentation network. Our method works in a recursive manner to\ngradually refine the above-mentioned segment proposals. Extensive experimental\nresults on the 2D-3D-S dataset show that the proposed method can generate\naccurate segment proposals when bounding box labels are available on only a\nsmall subset of training images. Performance comparison with recent\nstate-of-the-art methods further illustrates the effectiveness of our method.",
+                    "authors": "Weixuan Sun, Jing Zhang, Nick Barnes",
+                    "published": "2020-12-01",
+                    "updated": "2020-12-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1709.01956v1",
+                    "title": "Learning Dilation Factors for Semantic Segmentation of Street Scenes",
+                    "abstract": "Contextual information is crucial for semantic segmentation. However, finding\nthe optimal trade-off between keeping desired fine details and at the same time\nproviding sufficiently large receptive fields is non trivial. This is even more\nso, when objects or classes present in an image significantly vary in size.\nDilated convolutions have proven valuable for semantic segmentation, because\nthey allow to increase the size of the receptive field without sacrificing\nimage resolution. However, in current state-of-the-art methods, dilation\nparameters are hand-tuned and fixed. In this paper, we present an approach for\nlearning dilation parameters adaptively per channel, consistently improving\nsemantic segmentation results on street-scene datasets like Cityscapes and\nCamvid.",
+                    "authors": "Yang He, Margret Keuper, Bernt Schiele, Mario Fritz",
+                    "published": "2017-09-06",
+                    "updated": "2017-09-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1609.09220v1",
+                    "title": "CNN-aware Binary Map for General Semantic Segmentation",
+                    "abstract": "In this paper we introduce a novel method for general semantic segmentation\nthat can benefit from general semantics of Convolutional Neural Network (CNN).\nOur segmentation proposes visually and semantically coherent image segments. We\nuse binary encoding of CNN features to overcome the difficulty of the\nclustering on the high-dimensional CNN feature space. These binary codes are\nvery robust against noise and non-semantic changes in the image. These binary\nencoding can be embedded into the CNN as an extra layer at the end of the\nnetwork. This results in real-time segmentation. To the best of our knowledge\nour method is the first attempt on general semantic image segmentation using\nCNN. All the previous papers were limited to few number of category of the\nimages (e.g. PASCAL VOC). Experiments show that our segmentation algorithm\noutperform the state-of-the-art non-semantic segmentation methods by large\nmargin.",
+                    "authors": "Mahdyar Ravanbakhsh, Hossein Mousavi, Moin Nabi, Mohammad Rastegari, Carlo Regazzoni",
+                    "published": "2016-09-29",
+                    "updated": "2016-09-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2011.00674v1",
+                    "title": "Highway Driving Dataset for Semantic Video Segmentation",
+                    "abstract": "Scene understanding is an essential technique in semantic segmentation.\nAlthough there exist several datasets that can be used for semantic\nsegmentation, they are mainly focused on semantic image segmentation with large\ndeep neural networks. Therefore, these networks are not useful for real time\napplications, especially in autonomous driving systems. In order to solve this\nproblem, we make two contributions to semantic segmentation task. The first\ncontribution is that we introduce the semantic video dataset, the Highway\nDriving dataset, which is a densely annotated benchmark for a semantic video\nsegmentation task. The Highway Driving dataset consists of 20 video sequences\nhaving a 30Hz frame rate, and every frame is densely annotated. Secondly, we\npropose a baseline algorithm that utilizes a temporal correlation. Together\nwith our attempt to analyze the temporal correlation, we expect the Highway\nDriving dataset to encourage research on semantic video segmentation.",
+                    "authors": "Byungju Kim, Junho Yim, Junmo Kim",
+                    "published": "2020-11-02",
+                    "updated": "2020-11-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.13026v1",
+                    "title": "Weakly-Supervised Semantic Segmentation with Image-Level Labels: from Traditional Models to Foundation Models",
+                    "abstract": "The rapid development of deep learning has driven significant progress in the\nfield of image semantic segmentation - a fundamental task in computer vision.\nSemantic segmentation algorithms often depend on the availability of\npixel-level labels (i.e., masks of objects), which are expensive,\ntime-consuming, and labor-intensive. Weakly-supervised semantic segmentation\n(WSSS) is an effective solution to avoid such labeling. It utilizes only\npartial or incomplete annotations and provides a cost-effective alternative to\nfully-supervised semantic segmentation. In this paper, we focus on the WSSS\nwith image-level labels, which is the most challenging form of WSSS. Our work\nhas two parts. First, we conduct a comprehensive survey on traditional methods,\nprimarily focusing on those presented at premier research conferences. We\ncategorize them into four groups based on where their methods operate:\npixel-wise, image-wise, cross-image, and external data. Second, we investigate\nthe applicability of visual foundation models, such as the Segment Anything\nModel (SAM), in the context of WSSS. We scrutinize SAM in two intriguing\nscenarios: text prompting and zero-shot learning. We provide insights into the\npotential and challenges associated with deploying visual foundational models\nfor WSSS, facilitating future developments in this exciting research area.",
+                    "authors": "Zhaozheng Chen, Qianru Sun",
+                    "published": "2023-10-19",
+                    "updated": "2023-10-19",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.13395v3",
+                    "title": "ACDC: The Adverse Conditions Dataset with Correspondences for Semantic Driving Scene Understanding",
+                    "abstract": "Level 5 autonomy for self-driving cars requires a robust visual perception\nsystem that can parse input images under any visual condition. However,\nexisting semantic segmentation datasets are either dominated by images captured\nunder normal conditions or are small in scale. To address this, we introduce\nACDC, the Adverse Conditions Dataset with Correspondences for training and\ntesting semantic segmentation methods on adverse visual conditions. ACDC\nconsists of a large set of 4006 images which are equally distributed between\nfour common adverse conditions: fog, nighttime, rain, and snow. Each\nadverse-condition image comes with a high-quality fine pixel-level semantic\nannotation, a corresponding image of the same scene taken under normal\nconditions, and a binary mask that distinguishes between intra-image regions of\nclear and uncertain semantic content. Thus, ACDC supports both standard\nsemantic segmentation and the newly introduced uncertainty-aware semantic\nsegmentation. A detailed empirical study demonstrates the challenges that the\nadverse domains of ACDC pose to state-of-the-art supervised and unsupervised\napproaches and indicates the value of our dataset in steering future progress\nin the field. Our dataset and benchmark are publicly available.",
+                    "authors": "Christos Sakaridis, Dengxin Dai, Luc Van Gool",
+                    "published": "2021-04-27",
+                    "updated": "2021-09-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2403.01482v4",
+                    "title": "EAGLE: Eigen Aggregation Learning for Object-Centric Unsupervised Semantic Segmentation",
+                    "abstract": "Semantic segmentation has innately relied on extensive pixel-level annotated\ndata, leading to the emergence of unsupervised methodologies. Among them,\nleveraging self-supervised Vision Transformers for unsupervised semantic\nsegmentation (USS) has been making steady progress with expressive deep\nfeatures. Yet, for semantically segmenting images with complex objects, a\npredominant challenge remains: the lack of explicit object-level semantic\nencoding in patch-level features. This technical limitation often leads to\ninadequate segmentation of complex objects with diverse structures. To address\nthis gap, we present a novel approach, EAGLE, which emphasizes object-centric\nrepresentation learning for unsupervised semantic segmentation. Specifically,\nwe introduce EiCue, a spectral technique providing semantic and structural cues\nthrough an eigenbasis derived from the semantic similarity matrix of deep image\nfeatures and color affinity from an image. Further, by incorporating our\nobject-centric contrastive loss with EiCue, we guide our model to learn\nobject-level representations with intra- and inter-image object-feature\nconsistency, thereby enhancing semantic accuracy. Extensive experiments on\nCOCO-Stuff, Cityscapes, and Potsdam-3 datasets demonstrate the state-of-the-art\nUSS results of EAGLE with accurate and consistent semantic segmentation across\ncomplex scenes.",
+                    "authors": "Chanyoung Kim, Woojung Han, Dayun Ju, Seong Jae Hwang",
+                    "published": "2024-03-03",
+                    "updated": "2024-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.02094v1",
+                    "title": "Segment Anything Meets Semantic Communication",
+                    "abstract": "In light of the diminishing returns of traditional methods for enhancing\ntransmission rates, the domain of semantic communication presents promising new\nfrontiers. Focusing on image transmission, this paper explores the application\nof foundation models, particularly the Segment Anything Model (SAM) developed\nby Meta AI Research, to improve semantic communication. SAM is a promptable\nimage segmentation model that has gained attention for its ability to perform\nzero-shot segmentation tasks without explicit training or domain-specific\nknowledge. By employing SAM's segmentation capability and lightweight neural\nnetwork architecture for semantic coding, we propose a practical approach to\nsemantic communication. We demonstrate that this approach retains critical\nsemantic features, achieving higher image reconstruction quality and reducing\ncommunication overhead. This practical solution eliminates the\nresource-intensive stage of training a segmentation model and can be applied to\nany semantic coding architecture, paving the way for real-world applications.",
+                    "authors": "Shehbaz Tariq, Brian Estadimas Arfeto, Chaoning Zhang, Hyundong Shin",
+                    "published": "2023-06-03",
+                    "updated": "2023-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1511.06988v1",
+                    "title": "Learning High-level Prior with Convolutional Neural Networks for Semantic Segmentation",
+                    "abstract": "This paper proposes a convolutional neural network that can fuse high-level\nprior for semantic image segmentation. Motivated by humans' vision recognition\nsystem, our key design is a three-layer generative structure consisting of\nhigh-level coding, middle-level segmentation and low-level image to introduce\nglobal prior for semantic segmentation. Based on this structure, we proposed a\ngenerative model called conditional variational auto-encoder (CVAE) that can\nbuild up the links behind these three layers. These important links include an\nimage encoder that extracts high level info from image, a segmentation encoder\nthat extracts high level info from segmentation, and a hybrid decoder that\noutputs semantic segmentation from the high level prior and input image. We\ntheoretically derive the semantic segmentation as an optimization problem\nparameterized by these links. Finally, the optimization problem enables us to\ntake advantage of state-of-the-art fully convolutional network structure for\nthe implementation of the above encoders and decoder. Experimental results on\nseveral representative datasets demonstrate our supreme performance for\nsemantic segmentation.",
+                    "authors": "Haitian Zheng, Yebin Liu, Mengqi Ji, Feng Wu, Lu Fang",
+                    "published": "2015-11-22",
+                    "updated": "2015-11-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1610.01706v1",
+                    "title": "Exploiting Depth from Single Monocular Images for Object Detection and Semantic Segmentation",
+                    "abstract": "Augmenting RGB data with measured depth has been shown to improve the\nperformance of a range of tasks in computer vision including object detection\nand semantic segmentation. Although depth sensors such as the Microsoft Kinect\nhave facilitated easy acquisition of such depth information, the vast majority\nof images used in vision tasks do not contain depth information. In this paper,\nwe show that augmenting RGB images with estimated depth can also improve the\naccuracy of both object detection and semantic segmentation. Specifically, we\nfirst exploit the recent success of depth estimation from monocular images and\nlearn a deep depth estimation model. Then we learn deep depth features from the\nestimated depth and combine with RGB features for object detection and semantic\nsegmentation. Additionally, we propose an RGB-D semantic segmentation method\nwhich applies a multi-task training scheme: semantic label prediction and depth\nvalue regression. We test our methods on several datasets and demonstrate that\nincorporating information from estimated depth improves the performance of\nobject detection and semantic segmentation remarkably.",
+                    "authors": "Yuanzhouhan Cao, Chunhua Shen, Heng Tao Shen",
+                    "published": "2016-10-06",
+                    "updated": "2016-10-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2111.11103v1",
+                    "title": "Improving Semantic Image Segmentation via Label Fusion in Semantically Textured Meshes",
+                    "abstract": "Models for semantic segmentation require a large amount of hand-labeled\ntraining data which is costly and time-consuming to produce. For this purpose,\nwe present a label fusion framework that is capable of improving semantic pixel\nlabels of video sequences in an unsupervised manner. We make use of a 3D mesh\nrepresentation of the environment and fuse the predictions of different frames\ninto a consistent representation using semantic mesh textures. Rendering the\nsemantic mesh using the original intrinsic and extrinsic camera parameters\nyields a set of improved semantic segmentation images. Due to our optimized\nCUDA implementation, we are able to exploit the entire $c$-dimensional\nprobability distribution of annotations over $c$ classes in an\nuncertainty-aware manner. We evaluate our method on the Scannet dataset where\nwe improve annotations produced by the state-of-the-art segmentation network\nESANet from $52.05 \\%$ to $58.25 \\%$ pixel accuracy. We publish the source code\nof our framework online to foster future research in this area\n(\\url{https://github.com/fferflo/semantic-meshes}). To the best of our\nknowledge, this is the first publicly available label fusion framework for\nsemantic image segmentation based on meshes with semantic textures.",
+                    "authors": "Florian Fervers, Timo Breuer, Gregor Stachowiak, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens",
+                    "published": "2021-11-22",
+                    "updated": "2021-11-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2312.17243v1",
+                    "title": "Unsupervised Universal Image Segmentation",
+                    "abstract": "Several unsupervised image segmentation approaches have been proposed which\neliminate the need for dense manually-annotated segmentation masks; current\nmodels separately handle either semantic segmentation (e.g., STEGO) or\nclass-agnostic instance segmentation (e.g., CutLER), but not both (i.e.,\npanoptic segmentation). We propose an Unsupervised Universal Segmentation model\n(U2Seg) adept at performing various image segmentation tasks -- instance,\nsemantic and panoptic -- using a novel unified framework. U2Seg generates\npseudo semantic labels for these segmentation tasks via leveraging\nself-supervised models followed by clustering; each cluster represents\ndifferent semantic and/or instance membership of pixels. We then self-train the\nmodel on these pseudo semantic labels, yielding substantial performance gains\nover specialized methods tailored to each task: a +2.6 AP$^{\\text{box}}$ boost\nvs. CutLER in unsupervised instance segmentation on COCO and a +7.0 PixelAcc\nincrease (vs. STEGO) in unsupervised semantic segmentation on COCOStuff.\nMoreover, our method sets up a new baseline for unsupervised panoptic\nsegmentation, which has not been previously explored. U2Seg is also a strong\npretrained model for few-shot segmentation, surpassing CutLER by +5.0\nAP$^{\\text{mask}}$ when trained on a low-data regime, e.g., only 1% COCO\nlabels. We hope our simple yet effective method can inspire more research on\nunsupervised universal image segmentation.",
+                    "authors": "Dantong Niu, Xudong Wang, Xinyang Han, Long Lian, Roei Herzig, Trevor Darrell",
+                    "published": "2023-12-28",
+                    "updated": "2023-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1811.00174v4",
+                    "title": "Pixel Level Data Augmentation for Semantic Image Segmentation using Generative Adversarial Networks",
+                    "abstract": "Semantic segmentation is one of the basic topics in computer vision, it aims\nto assign semantic labels to every pixel of an image. Unbalanced semantic label\ndistribution could have a negative influence on segmentation accuracy. In this\npaper, we investigate using data augmentation approach to balance the semantic\nlabel distribution in order to improve segmentation performance. We propose\nusing generative adversarial networks (GANs) to generate realistic images for\nimproving the performance of semantic segmentation networks. Experimental\nresults show that the proposed method can not only improve segmentation\nperformance on those classes with low accuracy, but also obtain 1.3% to 2.1%\nincrease in average segmentation accuracy. It shows that this augmentation\nmethod can boost accuracy and be easily applicable to any other segmentation\nmodels.",
+                    "authors": "Shuangting Liu, Jiaqi Zhang, Yuxin Chen, Yifan Liu, Zengchang Qin, Tao Wan",
+                    "published": "2018-11-01",
+                    "updated": "2019-11-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2205.13278v1",
+                    "title": "Semantic Segmentation for Thermal Images: A Comparative Survey",
+                    "abstract": "Semantic segmentation is a challenging task since it requires excessively\nmore low-level spatial information of the image compared to other computer\nvision problems. The accuracy of pixel-level classification can be affected by\nmany factors, such as imaging limitations and the ambiguity of object\nboundaries in an image. Conventional methods exploit three-channel RGB images\ncaptured in the visible spectrum with deep neural networks (DNN). Thermal\nimages can significantly contribute during the segmentation since thermal\nimaging cameras are capable of capturing details despite the weather and\nillumination conditions. Using infrared spectrum in semantic segmentation has\nmany real-world use cases, such as autonomous driving, medical imaging,\nagriculture, defense industry, etc. Due to this wide range of use cases,\ndesigning accurate semantic segmentation algorithms with the help of infrared\nspectrum is an important challenge. One approach is to use both visible and\ninfrared spectrum images as inputs. These methods can accomplish higher\naccuracy due to enriched input information, with the cost of extra effort for\nthe alignment and processing of multiple inputs. Another approach is to use\nonly thermal images, enabling less hardware cost for smaller use cases. Even\nthough there are multiple surveys on semantic segmentation methods, the\nliterature lacks a comprehensive survey centered explicitly around semantic\nsegmentation using infrared spectrum. This work aims to fill this gap by\npresenting algorithms in the literature and categorizing them by their input\nimages.",
+                    "authors": "Z\u00fclfiye K\u00fct\u00fck, G\u00f6rkem Algan",
+                    "published": "2022-05-26",
+                    "updated": "2022-05-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2101.08418v2",
+                    "title": "Rethinking Semantic Segmentation Evaluation for Explainability and Model Selection",
+                    "abstract": "Semantic segmentation aims to robustly predict coherent class labels for\nentire regions of an image. It is a scene understanding task that powers\nreal-world applications (e.g., autonomous navigation). One important\napplication, the use of imagery for automated semantic understanding of\npedestrian environments, provides remote mapping of accessibility features in\nstreet environments. This application (and others like it) require detailed\ngeometric information of geographical objects. Semantic segmentation is a\nprerequisite for this task since it maps contiguous regions of the same class\nas single entities. Importantly, semantic segmentation uses like ours are not\npixel-wise outcomes; however, most of their quantitative evaluation metrics\n(e.g., mean Intersection Over Union) are based on pixel-wise similarities to a\nground-truth, which fails to emphasize over- and under-segmentation properties\nof a segmentation model. Here, we introduce a new metric to assess region-based\nover- and under-segmentation. We analyze and compare it to other metrics,\ndemonstrating that the use of our metric lends greater explainability to\nsemantic segmentation model performance in real-world applications.",
+                    "authors": "Yuxiang Zhang, Sachin Mehta, Anat Caspi",
+                    "published": "2021-01-21",
+                    "updated": "2023-02-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2006.07601v1",
+                    "title": "NoPeopleAllowed: The Three-Step Approach to Weakly Supervised Semantic Segmentation",
+                    "abstract": "We propose a novel approach to weakly supervised semantic segmentation, which\nconsists of three consecutive steps. The first two steps extract high-quality\npseudo masks from image-level annotated data, which are then used to train a\nsegmentation model on the third step. The presented approach also addresses two\nproblems in the data: class imbalance and missing labels. Using only\nimage-level annotations as supervision, our method is capable of segmenting\nvarious classes and complex objects. It achieves 37.34 mean IoU on the test\nset, placing 3rd at the LID Challenge in the task of weakly supervised semantic\nsegmentation.",
+                    "authors": "Mariia Dobko, Ostap Viniavskyi, Oles Dobosevych",
+                    "published": "2020-06-13",
+                    "updated": "2020-06-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2003.04404v1",
+                    "title": "FusionLane: Multi-Sensor Fusion for Lane Marking Semantic Segmentation Using Deep Neural Networks",
+                    "abstract": "It is a crucial step to achieve effective semantic segmentation of lane\nmarking during the construction of the lane level high-precision map. In recent\nyears, many image semantic segmentation methods have been proposed. These\nmethods mainly focus on the image from camera, due to the limitation of the\nsensor itself, the accurate three-dimensional spatial position of the lane\nmarking cannot be obtained, so the demand for the lane level high-precision map\nconstruction cannot be met. This paper proposes a lane marking semantic\nsegmentation method based on LIDAR and camera fusion deep neural network.\nDifferent from other methods, in order to obtain accurate position information\nof the segmentation results, the semantic segmentation object of this paper is\na bird's eye view converted from a LIDAR points cloud instead of an image\ncaptured by a camera. This method first uses the deeplabv3+ [\\ref{ref:1}]\nnetwork to segment the image captured by the camera, and the segmentation\nresult is merged with the point clouds collected by the LIDAR as the input of\nthe proposed network. In this neural network, we also add a long short-term\nmemory (LSTM) structure to assist the network for semantic segmentation of lane\nmarkings by using the the time series information. The experiments on more than\n14,000 image datasets which we have manually labeled and expanded have shown\nthe proposed method has better performance on the semantic segmentation of the\npoints cloud bird's eye view. Therefore, the automation of high-precision map\nconstruction can be significantly improved. Our code is available at\nhttps://github.com/rolandying/FusionLane.",
+                    "authors": "Ruochen Yin, Biao Yu, Huapeng Wu, Yutao Song, Runxin Niu",
+                    "published": "2020-03-09",
+                    "updated": "2020-03-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.05564v1",
+                    "title": "Hypergraph Convolutional Networks for Weakly-Supervised Semantic Segmentation",
+                    "abstract": "Semantic segmentation is a fundamental topic in computer vision. Several deep\nlearning methods have been proposed for semantic segmentation with outstanding\nresults. However, these models require a lot of densely annotated images. To\naddress this problem, we propose a new algorithm that uses HyperGraph\nConvolutional Networks for Weakly-supervised Semantic Segmentation\n(HyperGCN-WSS). Our algorithm constructs spatial and k-Nearest Neighbor (k-NN)\ngraphs from the images in the dataset to generate the hypergraphs. Then, we\ntrain a specialized HyperGraph Convolutional Network (HyperGCN) architecture\nusing some weak signals. The outputs of the HyperGCN are denominated\npseudo-labels, which are later used to train a DeepLab model for semantic\nsegmentation. HyperGCN-WSS is evaluated on the PASCAL VOC 2012 dataset for\nsemantic segmentation, using scribbles or clicks as weak signals. Our algorithm\nshows competitive performance against previous methods.",
+                    "authors": "Jhony H. Giraldo, Vincenzo Scarrica, Antonino Staiano, Francesco Camastra, Thierry Bouwmans",
+                    "published": "2022-10-11",
+                    "updated": "2022-10-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.11316v2",
+                    "title": "Generative Semantic Segmentation",
+                    "abstract": "We present Generative Semantic Segmentation (GSS), a generative learning\napproach for semantic segmentation. Uniquely, we cast semantic segmentation as\nan image-conditioned mask generation problem. This is achieved by replacing the\nconventional per-pixel discriminative learning with a latent prior learning\nprocess. Specifically, we model the variational posterior distribution of\nlatent variables given the segmentation mask. To that end, the segmentation\nmask is expressed with a special type of image (dubbed as maskige). This\nposterior distribution allows to generate segmentation masks unconditionally.\nTo achieve semantic segmentation on a given image, we further introduce a\nconditioning network. It is optimized by minimizing the divergence between the\nposterior distribution of maskige (i.e., segmentation masks) and the latent\nprior distribution of input training images. Extensive experiments on standard\nbenchmarks show that our GSS can perform competitively to prior art\nalternatives in the standard semantic segmentation setting, whilst achieving a\nnew state of the art in the more challenging cross-domain setting.",
+                    "authors": "Jiaqi Chen, Jiachen Lu, Xiatian Zhu, Li Zhang",
+                    "published": "2023-03-20",
+                    "updated": "2023-08-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.10782v2",
+                    "title": "Three Ways to Improve Semantic Segmentation with Self-Supervised Depth Estimation",
+                    "abstract": "Training deep networks for semantic segmentation requires large amounts of\nlabeled training data, which presents a major challenge in practice, as\nlabeling segmentation masks is a highly labor-intensive process. To address\nthis issue, we present a framework for semi-supervised semantic segmentation,\nwhich is enhanced by self-supervised monocular depth estimation from unlabeled\nimage sequences. In particular, we propose three key contributions: (1) We\ntransfer knowledge from features learned during self-supervised depth\nestimation to semantic segmentation, (2) we implement a strong data\naugmentation by blending images and labels using the geometry of the scene, and\n(3) we utilize the depth feature diversity as well as the level of difficulty\nof learning depth in a student-teacher framework to select the most useful\nsamples to be annotated for semantic segmentation. We validate the proposed\nmodel on the Cityscapes dataset, where all three modules demonstrate\nsignificant performance gains, and we achieve state-of-the-art results for\nsemi-supervised semantic segmentation. The implementation is available at\nhttps://github.com/lhoyer/improving_segmentation_with_selfsupervised_depth.",
+                    "authors": "Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian K\u00f6ring, Suman Saha, Luc Van Gool",
+                    "published": "2020-12-19",
+                    "updated": "2021-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.17874v1",
+                    "title": "SmooSeg: Smoothness Prior for Unsupervised Semantic Segmentation",
+                    "abstract": "Unsupervised semantic segmentation is a challenging task that segments images\ninto semantic groups without manual annotation. Prior works have primarily\nfocused on leveraging prior knowledge of semantic consistency or priori\nconcepts from self-supervised learning methods, which often overlook the\ncoherence property of image segments. In this paper, we demonstrate that the\nsmoothness prior, asserting that close features in a metric space share the\nsame semantics, can significantly simplify segmentation by casting unsupervised\nsemantic segmentation as an energy minimization problem. Under this paradigm,\nwe propose a novel approach called SmooSeg that harnesses self-supervised\nlearning methods to model the closeness relationships among observations as\nsmoothness signals. To effectively discover coherent semantic segments, we\nintroduce a novel smoothness loss that promotes piecewise smoothness within\nsegments while preserving discontinuities across different segments.\nAdditionally, to further enhance segmentation quality, we design an asymmetric\nteacher-student style predictor that generates smoothly updated pseudo labels,\nfacilitating an optimal fit between observations and labeling outputs. Thanks\nto the rich supervision cues of the smoothness prior, our SmooSeg significantly\noutperforms STEGO in terms of pixel accuracy on three datasets: COCOStuff\n(+14.9%), Cityscapes (+13.0%), and Potsdam-3 (+5.7%).",
+                    "authors": "Mengcheng Lan, Xinjiang Wang, Yiping Ke, Jiaxing Xu, Litong Feng, Wayne Zhang",
+                    "published": "2023-10-27",
+                    "updated": "2023-10-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.06974v1",
+                    "title": "A One Stop 3D Target Reconstruction and multilevel Segmentation Method",
+                    "abstract": "3D object reconstruction and multilevel segmentation are fundamental to\ncomputer vision research. Existing algorithms usually perform 3D scene\nreconstruction and target objects segmentation independently, and the\nperformance is not fully guaranteed due to the challenge of the 3D\nsegmentation. Here we propose an open-source one stop 3D target reconstruction\nand multilevel segmentation framework (OSTRA), which performs segmentation on\n2D images, tracks multiple instances with segmentation labels in the image\nsequence, and then reconstructs labelled 3D objects or multiple parts with\nMulti-View Stereo (MVS) or RGBD-based 3D reconstruction methods. We extend\nobject tracking and 3D reconstruction algorithms to support continuous\nsegmentation labels to leverage the advances in the 2D image segmentation,\nespecially the Segment-Anything Model (SAM) which uses the pretrained neural\nnetwork without additional training for new scenes, for 3D object segmentation.\nOSTRA supports most popular 3D object models including point cloud, mesh and\nvoxel, and achieves high performance for semantic segmentation, instance\nsegmentation and part segmentation on several 3D datasets. It even surpasses\nthe manual segmentation in scenes with complex structures and occlusions. Our\nmethod opens up a new avenue for reconstructing 3D targets embedded with rich\nmulti-scale segmentation information in complex scenes. OSTRA is available from\nhttps://github.com/ganlab/OSTRA.",
+                    "authors": "Jiexiong Xu, Weikun Zhao, Zhiyan Tang, Xiangchao Gan",
+                    "published": "2023-08-14",
+                    "updated": "2023-08-14",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.12417v2",
+                    "title": "SLAMs: Semantic Learning based Activation Map for Weakly Supervised Semantic Segmentation",
+                    "abstract": "Recent mainstream weakly-supervised semantic segmentation (WSSS) approaches\nmainly relies on image-level classification learning, which has limited\nrepresentation capacity. In this paper, we propose a novel semantic learning\nbased framework, named SLAMs (Semantic Learning based Activation Map), for\nWSSS.",
+                    "authors": "Junliang Chen, Xiaodong Zhao, Minmin Liu, Linlin Shen",
+                    "published": "2022-10-22",
+                    "updated": "2022-11-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.02840v1",
+                    "title": "Attention-based fusion of semantic boundary and non-boundary information to improve semantic segmentation",
+                    "abstract": "This paper introduces a method for image semantic segmentation grounded on a\nnovel fusion scheme, which takes place inside a deep convolutional neural\nnetwork. The main goal of our proposal is to explore object boundary\ninformation to improve the overall segmentation performance. Unlike previous\nworks that combine boundary and segmentation features, or those that use\nboundary information to regularize semantic segmentation, we instead propose a\nnovel approach that embodies boundary information onto segmentation. For that,\nour semantic segmentation method uses two streams, which are combined through\nan attention gate, forming an end-to-end Y-model. To the best of our knowledge,\nours is the first work to show that boundary detection can improve semantic\nsegmentation when fused through a semantic fusion gate (attention model). We\nperformed an extensive evaluation of our method over public data sets. We found\ncompetitive results on all data sets after comparing our proposed model with\nother twelve state-of-the-art segmenters, considering the same training\nconditions. Our proposed model achieved the best mIoU on the CityScapes,\nCamVid, and Pascal Context data sets, and the second best on Mapillary Vistas.",
+                    "authors": "Jefferson Fontinele, Gabriel Lefundes, Luciano Oliveira",
+                    "published": "2021-08-05",
+                    "updated": "2021-08-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.13118v1",
+                    "title": "Automatic Preprocessing and Ensemble Learning for Low Quality Cell Image Segmentation",
+                    "abstract": "We propose an automatic preprocessing and ensemble learning for segmentation\nof cell images with low quality. It is difficult to capture cells with strong\nlight. Therefore, the microscopic images of cells tend to have low image\nquality but these images are not good for semantic segmentation. Here we\npropose a method to translate an input image to the images that are easy to\nrecognize by deep learning. The proposed method consists of two deep neural\nnetworks. The first network is the usual training for semantic segmentation,\nand penultimate feature maps of the first network are used as filters to\ntranslate an input image to the images that emphasize each class. This is the\nautomatic preprocessing and translated cell images are easily classified. The\ninput cell image with low quality is translated by the feature maps in the\nfirst network, and the translated images are fed into the second network for\nsemantic segmentation. Since the outputs of the second network are multiple\nsegmentation results, we conduct the weighted ensemble of those segmentation\nimages. Two networks are trained by end-to-end manner, and we do not need to\nprepare images with high quality for the translation. We confirmed that our\nproposed method can translate cell images with low quality to the images that\nare easy to segment, and segmentation accuracy has improved using the weighted\nensemble learning.",
+                    "authors": "Sota Kato, Kazuhiro Hotta",
+                    "published": "2021-08-30",
+                    "updated": "2021-08-30",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2001.00335v1",
+                    "title": "Graph-FCN for image semantic segmentation",
+                    "abstract": "Semantic segmentation with deep learning has achieved great progress in\nclassifying the pixels in the image. However, the local location information is\nusually ignored in the high-level feature extraction by the deep learning,\nwhich is important for image semantic segmentation. To avoid this problem, we\npropose a graph model initialized by a fully convolutional network (FCN) named\nGraph-FCN for image semantic segmentation. Firstly, the image grid data is\nextended to graph structure data by a convolutional network, which transforms\nthe semantic segmentation problem into a graph node classification problem.\nThen we apply graph convolutional network to solve this graph node\nclassification problem. As far as we know, it is the first time that we apply\nthe graph convolutional network in image semantic segmentation. Our method\nachieves competitive performance in mean intersection over union (mIOU) on the\nVOC dataset(about 1.34% improvement), compared to the original FCN model.",
+                    "authors": "Yi Lu, Yaran Chen, Dongbin Zhao, Jianxin Chen",
+                    "published": "2020-01-02",
+                    "updated": "2020-01-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.17083v1",
+                    "title": "SegRCDB: Semantic Segmentation via Formula-Driven Supervised Learning",
+                    "abstract": "Pre-training is a strong strategy for enhancing visual models to efficiently\ntrain them with a limited number of labeled images. In semantic segmentation,\ncreating annotation masks requires an intensive amount of labor and time, and\ntherefore, a large-scale pre-training dataset with semantic labels is quite\ndifficult to construct. Moreover, what matters in semantic segmentation\npre-training has not been fully investigated. In this paper, we propose the\nSegmentation Radial Contour DataBase (SegRCDB), which for the first time\napplies formula-driven supervised learning for semantic segmentation. SegRCDB\nenables pre-training for semantic segmentation without real images or any\nmanual semantic labels. SegRCDB is based on insights about what is important in\npre-training for semantic segmentation and allows efficient pre-training.\nPre-training with SegRCDB achieved higher mIoU than the pre-training with\nCOCO-Stuff for fine-tuning on ADE-20k and Cityscapes with the same number of\ntraining images. SegRCDB has a high potential to contribute to semantic\nsegmentation pre-training and investigation by enabling the creation of large\ndatasets without manual annotation. The SegRCDB dataset will be released under\na license that allows research and commercial use. Code is available at:\nhttps://github.com/dahlian00/SegRCDB",
+                    "authors": "Risa Shinoda, Ryo Hayamizu, Kodai Nakashima, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka",
+                    "published": "2023-09-29",
+                    "updated": "2023-09-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1707.02432v2",
+                    "title": "Deep Semantic Segmentation for Automated Driving: Taxonomy, Roadmap and Challenges",
+                    "abstract": "Semantic segmentation was seen as a challenging computer vision problem few\nyears ago. Due to recent advancements in deep learning, relatively accurate\nsolutions are now possible for its use in automated driving. In this paper, the\nsemantic segmentation problem is explored from the perspective of automated\ndriving. Most of the current semantic segmentation algorithms are designed for\ngeneric images and do not incorporate prior structure and end goal for\nautomated driving. First, the paper begins with a generic taxonomic survey of\nsemantic segmentation algorithms and then discusses how it fits in the context\nof automated driving. Second, the particular challenges of deploying it into a\nsafety system which needs high level of accuracy and robustness are listed.\nThird, different alternatives instead of using an independent semantic\nsegmentation module are explored. Finally, an empirical evaluation of various\nsemantic segmentation architectures was performed on CamVid dataset in terms of\naccuracy and speed. This paper is a preliminary shorter version of a more\ndetailed survey which is work in progress.",
+                    "authors": "Mennatullah Siam, Sara Elkerdawy, Martin Jagersand, Senthil Yogamani",
+                    "published": "2017-07-08",
+                    "updated": "2017-08-03",
+                    "primary_cat": "stat.ML",
+                    "cats": [
+                        "stat.ML",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.05321v1",
+                    "title": "Image Segmentation Semantic Communication over Internet of Vehicles",
+                    "abstract": "In this paper, the problem of semantic-based efficient image transmission is\nstudied over the Internet of Vehicles (IoV). In the considered model, a vehicle\nshares massive amount of visual data perceived by its visual sensors to assist\nother vehicles in making driving decisions. However, it is hard to maintain a\nhigh reliable visual data transmission due to the limited spectrum resources.\nTo tackle this problem, a semantic communication approach is introduced to\nreduce the transmission data amount while ensuring the semantic-level accuracy.\nParticularly, an image segmentation semantic communication (ISSC) system is\nproposed, which can extract the semantic features from the perceived images and\ntransmit the features to the receiving vehicle that reconstructs the image\nsegmentations. The ISSC system consists of an encoder and a decoder at the\ntransmitter and the receiver, respectively. To accurately extract the image\nsemantic features, the ISSC system encoder employs a Swin Transformer based\nmulti-scale semantic feature extractor. Then, to resist the wireless noise and\nreconstruct the image segmentation, a semantic feature decoder and a\nreconstructor are designed at the receiver. Simulation results show that the\nproposed ISSC system can reconstruct the image segmentation accurately with a\nhigh compression ratio, and can achieve robust transmission performance against\nchannel noise, especially at the low signal-to-noise ratio (SNR). In terms of\nmean Intersection over Union (mIoU), the ISSC system can achieve an increase by\n75%, compared to the baselines using traditional coding method",
+                    "authors": "Qiang Pan, Haonan Tong, Jie Lv, Tao Luo, Zhilong Zhang, Changchuan Yin, Jianfeng Li",
+                    "published": "2022-10-11",
+                    "updated": "2022-10-11",
+                    "primary_cat": "cs.NI",
+                    "cats": [
+                        "cs.NI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.06753v1",
+                    "title": "3rd Place Solution for PVUW Challenge 2023: Video Panoptic Segmentation",
+                    "abstract": "In order to deal with the task of video panoptic segmentation in the wild, we\npropose a robust integrated video panoptic segmentation solution. In our\nsolution, we regard the video panoptic segmentation task as a segmentation\ntarget querying task, represent both semantic and instance targets as a set of\nqueries, and then combine these queries with video features extracted by neural\nnetworks to predict segmentation masks. In order to improve the learning\naccuracy and convergence speed of the solution, we add additional tasks of\nvideo semantic segmentation and video instance segmentation for joint training.\nIn addition, we also add an additional image semantic segmentation model to\nfurther improve the performance of semantic classes. In addition, we also add\nsome additional operations to improve the robustness of the model. Extensive\nexperiments on the VIPSeg dataset show that the proposed solution achieves\nstate-of-the-art performance with 50.04\\% VPQ on the VIPSeg test set, which is\n3rd place on the video panoptic segmentation track of the PVUW Challenge 2023.",
+                    "authors": "Jinming Su, Wangwang Yang, Junfeng Luo, Xiaolin Wei",
+                    "published": "2023-06-11",
+                    "updated": "2023-06-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.08988v1",
+                    "title": "Heterogeneous Feature Distillation Network for SAR Image Semantic Segmentation",
+                    "abstract": "Semantic segmentation for SAR (Synthetic Aperture Radar) images has attracted\nincreasing attention in the remote sensing community recently, due to SAR's\nall-time and all-weather imaging capability. However, SAR images are generally\nmore difficult to be segmented than their EO (Electro-Optical) counterparts,\nsince speckle noises and layovers are inevitably involved in SAR images. To\naddress this problem, we investigate how to introduce EO features to assist the\ntraining of a SAR-segmentation model, and propose a heterogeneous feature\ndistillation network for segmenting SAR images, called HFD-Net, where a\nSAR-segmentation student model gains knowledge from a pre-trained\nEO-segmentation teacher model. In the proposed HFD-Net, both the student and\nteacher models employ an identical architecture but different parameter\nconfigurations, and a heterogeneous feature distillation model is explored for\ntransferring latent EO features from the teacher model to the student model and\nthen enhancing the ability of the student model for SAR image segmentation. In\naddition, a heterogeneous feature alignment module is explored to aggregate\nmulti-scale features for segmentation in each of the student model and teacher\nmodel. Extensive experimental results on two public datasets demonstrate that\nthe proposed HFD-Net outperforms seven state-of-the-art SAR image semantic\nsegmentation methods.",
+                    "authors": "Gao Mengyu, Dong Qiulei",
+                    "published": "2022-10-17",
+                    "updated": "2022-10-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                }
+            ],
+            [
+                {
+                    "url": "http://arxiv.org/abs/2311.06242v1",
+                    "title": "Florence-2: Advancing a Unified Representation for a Variety of Vision Tasks",
+                    "abstract": "We introduce Florence-2, a novel vision foundation model with a unified,\nprompt-based representation for a variety of computer vision and\nvision-language tasks. While existing large vision models excel in transfer\nlearning, they struggle to perform a diversity of tasks with simple\ninstructions, a capability that implies handling the complexity of various\nspatial hierarchy and semantic granularity. Florence-2 was designed to take\ntext-prompt as task instructions and generate desirable results in text forms,\nwhether it be captioning, object detection, grounding or segmentation. This\nmulti-task learning setup demands large-scale, high-quality annotated data. To\nthis end, we co-developed FLD-5B that consists of 5.4 billion comprehensive\nvisual annotations on 126 million images, using an iterative strategy of\nautomated image annotation and model refinement. We adopted a\nsequence-to-sequence structure to train Florence-2 to perform versatile and\ncomprehensive vision tasks. Extensive evaluations on numerous tasks\ndemonstrated Florence-2 to be a strong vision foundation model contender with\nunprecedented zero-shot and fine-tuning capabilities.",
+                    "authors": "Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, Lu Yuan",
+                    "published": "2023-11-10",
+                    "updated": "2023-11-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Original Paper",
+                    "paper_cat": "Semantic AND Segmentation AND Image",
+                    "gt": "7.1. Vision-Language Foundation Models Recent vision-language pre-training models [29, 64, 95] have demonstrated impressive zero-shot transfer abilities to vision-language alignment and image classification tasks, thanks to the alignment of vision and text embeddings extracted from respective encoders through contrastive learning objectives [58, 74]. These models (e.g., [95]), trained on weakly large-scale image-text data, have been further extended to more downstream tasks such as object detection, achieving state-of-the-art performance with taskspecific adaptation heads. In contrast, other studies [2, 45, 78, 92] propose using a multi-modality decoder to predict text in an autoregressive manner with language modeling pre-training objectives. Techniques for fusing vision and language embeddings vary: GIT [78] concatenates vision and text tokens as decoder input and designs a casual attention mask, CoCa [92] uses attentional poolers with learnable queries to select task-specific vision representations which are then cross-attended via the decoder, and Flamingo [2] pools a fixed number of vision tokens with a Perceiver Resampler and adds new learnable cross-attention layers to the decoder while freezing the pre-trained vision encoder and text decoder. Beyond image captioning pre-training task, some research [15,55,79] attempts to formulate more vision tasks in a unified sequence-to-sequence learning paradigm, including object detection and image segmentation. Customized special tokens accommodate representations beyond pure text, such as bounding boxes [10, 55, 79]. This approach uses the same architecture for pre-training and downstream tasks, potentially using the same set of weights for all tasks. Our method, which falls into this category, aims to obtain foundation models that understand dense information beyond simple image-level captions. It shares the same encoder-decoder design as other multi-modality encoder14 decoder models [15, 55] adapted for sequence-to-sequence learning, but uses our built large-scale comprehensive annotation data instead of combining existing sparse annotated data. 7.2. Vision Datasets Comprehensive annotations. The quest for comprehensive understanding of visual scenes, the holy grail of computer vision [36], has evolved from focusing on individual datasets each targeting a single perspective, e.g., image classification [18], to providing multi-perspective [36, 40, 48], comprehensive annotations for every visual data point. Notable datasets like MS-COCO [13, 48] and Visual Genome [36] integrate various types of annotations, enabling richer understanding in spatial and semantic granularities and better model interactions across annotations. However, due to the high cost of human verification, these annotations are limited in size. Our datasets, while largescale, maintain comprehensive annotations covering text, region-text pairs, and text-phrase-region triplets, with reduced human involvement. Scalable annotations.: Over the past decade, vision datasets have rapidly scaled up from thousands [37, 42] to billion examples [29, 96] to encompass more visual concepts for better generalization. This shift is evident in recent foundation models that employ massive quantities of data [5]. These large datasets typically collect images from the web and parse noisy annotations from the corresponding metadata, such as category label from query [75, 96], short description from alt-text [29,64], as well as detailed description from interleaved text [2, 41]. Despite their diversity, these annotations suffer from randomness and limited types (i.e., texts only). Some works [32, 45] attempt to scale up annotations using pseudo-label generation with iteratively trained models, which offer higher quality without significant diversity loss. Our data pipeline extends these largescale, web-crawled noisy annotations with higher-quality, autonomous annotations generated from multiple specialist models. The pipeline iteratively refines labels and completes missing pieces, resulting in a scalable and comprehensive dataset for learning a unified visual representation.",
+                    "pre_questions": [],
+                    "main_content": "Introduction In the realm of Artificial General Intelligence (AGI) systems, there has been a notable shift towards utilizing pretrained, versatile representations, acknowledged for taskagnostic benefits accross diverse applications. This trend is evident in natural language processing (NLP), where advanced models [5, 6, 19, 43, 65, 66] show adaptability with comprehensive knowledge spanning various domains and tasks with simple instructions. The success of NLP motivates a parallel approach in computer vision. Universal representation for diverse vision-related tasks presents unique challenges, notably the need for comprehensive perceptual abilities. Unlike NLP, which deals Region-level Image-level Pixel-level None semantic Fine-grained semantic Coarse semantic A woman riding a bike down a street next to a red car. The image shows a person riding a red bicycle on a road with a red car in the background. The road is lined with trees on both sides and there is another person riding another bicycle in front of her. The date \" 9/22/2023\" is visible in the bottom. person car road red vintage car on street Spatial Hierarchy Semantic Granularity FLD-5B (Comprehensive Annotations) Florence-2 (Unified Architecture) classification caption detailed caption visual grounding & object detection regional proposal segmentation phrase segmentation visual grounding & object detection Figure 1. We aim to build a vision foundation model to enable extensive perception capabilities including spatial hierarchy and semantic granularity. To achieve this, a single unified model Florence-2 is pre-trained on our FLD-5B dataset encompassing a total of 5.4B comprehensive annotations across 126M images, which are collected by our Florence data engine. mainly with text, computer vision requires handling intricate visual data like object location, masked contours, and attributes. Attaining universal representation in computer vision demands adept management of a spectrum of complex tasks, organized two-dimensionally as illustrated in Figure 1: \u2022 Spatial Hierarchy: The model must discern spatial details across varying scales, understanding imagelevel concepts and fine-grained pixel specifics. Accommodating the intricate spatial hierarchy within vision demands the model\u2019s proficiency in handling diverse levels of granularity. \u2022 Semantic Granularity: Universal representation in computer vision should span a spectrum of semantic granularity. The model transitions from high-level captions to nuanced descriptions, enabling versatile understanding for diverse applications. 1 arXiv:2311.06242v1  [cs.CV]  10 Nov 2023 This pursuit is characterized by distinctiveness and substantial challenges. A key hurdle is the scarcity of comprehensive visual annotations, hindering the development of a foundational model capable of capturing the intricate nuances of spatial hierarchy and semantic granularity. Existing datasets, such as ImageNet [18], COCO [48], and Flickr30k Entities [61], tailored for specialized applications, are extensively labeled by humans. To overcome this constraint, it is imperative to generate extensive annotations for each image on a larger scale. Another challenge is the absence of a unified pretraining framework with a singular network architecture that seamlessly integrates spatial hierarchy and semantic granularity in computer vision. Traditional models excel in tasks like object detection [26, 97], semantic segmentation [16, 82], and image captioning [45, 78] with taskspecific design. However, it is essential to develop a comprehensive, unified model that is capable of adapting across various vision tasks in a task-agnostic manner, even accommodating new tasks with minimal or no task-specific finetuning. The model Florence [95] pioneers the integration of spatial, temporal, and multi-modal aspects in computer vision through unified pre-training and network architecture. The first evolutionary version [95] excels in transfer learning via pre-training with noisy text-image pairs and task-specific fine-tuning using specialized adapters. However, it relies on large task-specific datasets and adapters, leaving gaps in addressing the above dual key challenges. In this paper, we introduce Florence-2, a universal backbone achieved through multitask learning with extensive visual annotations. This results in a unified, prompt-based representation for diverse vision tasks, effectively addressing the challenges of limited comprehensive data and the absence of a unified architecture. Multitask learning necessitates large-scale, high-quality annotated data. Our data engine, instead of relying on labor-intensive manual annotation, autonomously generates a comprehensive visual dataset called FLD-5B, encompassing a total of 5.4B annotations for 126M images. This engine consists of two efficient processing modules. The first module uses specialized models to collaboratively and autonomously annotate images, moving away from the traditional single and manual annotation approach. Multiple models work together to reach a consensus, reminiscent of the wisdom of crowds concept [33, 80, 89], ensuring a more reliable and unbiased image understanding. The second module iteratively refines and filters these automated annotations using well-trained foundational models. By utilizing this extensive dataset, our model employs a sequence-to-sequence (seq2seq) architecture [17,19,66,76], which integrates an image encoder and a multi-modality encoder-decoder. This design accommodates a spectrum of vision tasks without the need for task-specific architectural modifications, aligning with the ethos of the NLP community for versatile model development with a consistent underlying structure. All annotations in the dataset FLD-5B, are uniformly standardized into textual outputs, facilitating a unified multi-task learning approach with consistent optimization with the same loss function as the objective. The outcome is a versatile vision foundation model, Florence-2, capable of performing a variety of tasks, such as object detection, captioning, and grounding, all within a single model governed by a uniform set of parameters. Task activation is achieved through textual prompts, reflecting the approach used by Large Language Models (LLMs) [65]. Our approach attains a universal representation, demonstrating broad applicability across various visual tasks. Key results include: \u2022 As a versatile vision foundation model, Florence-2 achieves new state-of-the-art zero-shot performance in tasks such as captioning on COCO [48], visual grounding on Flick30k [61], and referring expression comprehension on RefCOCO/+/g [31,56,93]. \u2022 After fine-tuning with public human-annotated data, Florence-2, despite its compact size, competes with larger specialist models. Notably, the fine-tuned Florence-2 establishes new state-of-the-art results on the benchmarks on RefCOCO/+/g. \u2022 The pre-trained Florence-2 backbone enhances performance on downstream tasks, e.g. COCO object detection and instance segmentation, and ADE20K semantic segmentation, surpassing both supervised and self-supervised models. Compared to pre-trained models on ImageNet, ours improves training efficiency by 4\u00d7 and achieves substantial improvements of 6.9, 5.5, and 5.9 points on COCO [48] and ADE20K [98] datasets, using Mask-RCNN [26], DINO [97], and UperNet [82] frameworks respectively. 2. Rethinking Vision Model Pre-training In pursuit of a versatile vision foundation model, we revisit three predominant pre-training paradigms: supervised (e.g., ImageNet classification [18]), self-supervised (e.g., SimCLR [9], MoCo [25], BEiT [4], MAE [24]), and weakly supervised (e.g., CLIP [64], Florence [95], SAM [32]). Each paradigm captures unique aspects of visual data but is inherently limited by the constraints of single-task learning frameworks. Supervised pre-training excels in object recognition but lacks adaptability [38]; self-supervised algorithms reveal intricate features but may overemphasize certain attributes [8]; weakly supervised methods leverage unstructured textual annotations but yield only image-level understanding [64]. To build a unified vision foundation model suitable for various applications, we must explore 2 Locate the objects in the image. Image Encoder Transformer Encoders Transformer Decoders What does the image describe? Locate the phrases in the caption:A woman riding a bike. What does the region (0.41, 0.15, 0.63, 0.73) describe? What is the polygon mask of region (0.41, 0.15, 0.63, 0.73)? The image shows a person riding a red bicycle on a road with a red car in the background. The person is wearing a white t-shirt, black pants, and a black hat. She has a backpack on her back and is pedaling with their feet on the pedals. The road is lined with trees on both sides and there is another person riding another bicycle in front of her. The date \"9/22/2023\" is visible in the bottom right corner of the image. A woman riding a bike A women riding a bike (0.41, 0.15, 0.63, 0.73 person riding red bicycle on road (0.48, 0.19, 0.48, 0.18, 0.49, 0.17, ...) person (0.41, 0.15, 0.63, 0.73) \u2026 car (0.58, 0.26, 0.89, 0.61) Person Person Car Figure 2. Florence-2 consists of an image encoder and standard multi-modality encoder-decoder. We train Florence-2 on our FLD-5B data in a unified multitask learning paradigm, resulting in a generaslist vision foundation model, which can perform various vision tasks. innovative pre-training strategies that overcome single-task limitations and integrate both textual and visual semantics. Image understanding necessitates capturing multiple levels of granularity, from global semantics to local details, and comprehending spatial relationships between objects and entities in their semantic context. To address these core aspects of image understanding, our approach incorporates a diverse set of annotations, effectively capturing visual understanding nuances and bridging the gap between vision and language understanding. 2.1. Comprehensive Multitask Learning To develop a versatile vision foundation model, we formulate a range of multitask learning objectives, each tailored to address specific aspects of visual comprehension. These objectives align with our predefined criteria: spatial hierarchy and semantic granularity, inspired by recent research on multitask learning [2,12,14,15,55,79]. Our multitask learning approach incorporates three distinct learning objectives, each addressing a different level of granularity and semantic understanding: \u2022 Image-level understanding tasks capture high-level semantics and foster a comprehensive understanding of images through linguistic descriptions [13, 18, 34, 91]. They enable the model to comprehend the overall context of an image and grasp semantic relationships and contextual nuances in the language domain. Exemplar tasks include image classification, captioning, and visual question answering. \u2022 Region/pixel-level recognition tasks facilitate detailed object and entity localization within images, capturing relationships between objects and their spatial context. Tasks include object detection, segmentation, and referring expression comprehension. \u2022 Fine-grained visual-semantic alignment tasks require fine-grained understanding of both text and image. It involves locating the image regions that correspond to the text phrases, such as objects, attributes, or relations. These tasks challenge the ability to capture the local details of visual entities and their semantic contexts, as well as the interactions between textual and visual elements. By combining these three learning objectives in a multitask learning framework, our foundation model learns to handle different levels of detail and semantic understanding. This strategic alignment enables our model to deal with various spatial details, distinguish levels of detail in understanding, and go beyond surface-level recognition\u2014ultimately learning a universal representation for vision understanding. 3 3. Model We present the foundation model Florence-2, designed for universal representation learning, capable of handling various vision tasks with a single set of weights and a unified architecture. As depicted in Figure 2, Florence-2 employs a sequence-to-sequence learning paradigm [77], integrating all tasks, described in Section 2, under a common language modeling objective. The model takes images coupled with task-prompt as task instructions, and generates the desirable results in text forms. It uses a vision encoder to convert images into visual token embeddings, which are then concatenated with text embeddings and processed by a transformer-based multi-modal encoder-decoder to generate the response. In the following sections, we will provide a detailed explanation of each model component. Task formulation. We adopt a sequence-to-sequence framework [10,15,55,77] to address various vision tasks in a unified manner. As shown in Table 13, we formulate each task as a translation problem: Given an input image and a task-specific prompt, we generate the corresponding output response. Depending on the task, the prompt and response can be either text or region: \u2022 Text: When the prompt or answer is plain text without special formatting, we maintain it in our final sequence-to-sequence format. \u2022 Region: For region-specific tasks, we add location tokens to the tokenizer\u2019s vocabulary list, representing quantized coordinates. We create 1, 000 bins, similar to [10,11,55,79], and represent regions using formats tailored to task requirements: \u2013 Box representation (x0, y0, x1, y1): Utilized in tasks such as object detection and dense region captioning, with location tokens corresponding to the box coordinates. The location tokens are the coordinates of the top-left and bottom-right corners of the box. \u2013 Quad box representation (x0, y0, ..., x3, y3): For text detection and recognition tasks, using location tokens for each coordinate of the quadrilateral enclosing the text. The location tokens are the coordinates of each corner of the quad box, starting from the top-left and going clockwise. \u2013 Polygon Representation (x0, y0, ..., xn, yn): For referring segmentation tasks, with location tokens representing the vertices of the polygon. The location tokens are the coordinates of the vertices of the polygon, in clockwise order. By extending the tokenizer\u2019s vocabulary to include location tokens, we enable the model to process region-specific information in a unified learning format. This eliminates the need to design task-specific heads for different tasks and allows for a more data-centric approach. Vision encoder. We employ DaViT [20] as the vision encoder. It processes an input image I \u2208RH\u00d7W \u00d73 (with H and W denoting height and width, respectively) into flattened visual token embeddings V \u2208RNv\u00d7Dv, where Nv and Dv represent the number and dimensionality of vision tokens, respectively. Multi-modality encoder decoder. We use a standard encoder-decoder transformer architecture to process visual and language token embeddings. We first obtain prompt text embeddings Tprompt \u2208RNt\u00d7D using our extended language tokenizer and word embedding layer [43]. Then, we concatenate vision token embeddings with prompt embeddings to form the multi-modality encoder module input, X = [V\u2032, Tprompt], where V\u2032 \u2208RNv\u00d7D is obtained by applying a linear projection and LayerNorm layer [3] to V for dimensionality alignment. Optimization objective. Given the input x combined from the image and the prompt, and the target y, we use the standard language modeling with cross-entropy loss for all the tasks. L = \u2212 |y| X i=1 logP\u03b8(yi|y<i, x), (1) where \u03b8 are the network parameters, |y| is the number of target tokens. 4. Data Engine To train our Florence-2 model, we require a comprehensive, large-scale, high-quality multitask dataset encompassing various image data aspects. Given the scarcity of such data, we have developed a new multitask image dataset. This dataset FLD-5B includes 126M images, 500M text annotations, and 1.3B text-region annotations, and 3.6B textphrase-region annotations across different tasks. We extensively explain our data collection and annotation procedures, encompassing adaptations for various annotation types. The data engine pipeline, shown in Figure 3, will be discussed in subsequent sections. 4.1. Image Collection We construct our data by gathering a diverse collection of images from various sources. We begin with the identification of three key tasks that act as primary sources for our image corpus: image classification, object detection, and image captioning. Consequently, we curate and combine five distinct datasets originating from the aforementioned tasks: ImageNet-22k [18], Object 365 [70], Open Images [40], Conceptual Captions [71], and LAION [68] 4 Segmentation Grounding Azure OCR API Segmentation Model Object / Proposal Det. Caption Model Iterative data refinement Initial annotation w/ specialists Data filtering and enhancement Region filtering Florence-2 Multi-round re-training Fine-tune w/ small-scale enhanced data Train w/ largescale data Grounding Model Image collection Text filter and enhancement Final annotations LMM Annotator LLM Aggregator Text Filter Florence-1 Region Score NMS Blacklist Brief Caption Detailed Caption More Detailed Caption OCR Object Detection Region Proposal Dense Caption Figure 3. Florence-2 data engine consists of three essential phrases: (1) initial annotation employing specialist models, (2) data filtering to correct errors and remove irrelevant annotations, and (3) an iterative process for data refinement. Our final dataset (FLD-5B) of over 5B annotations contains 126M images, 500M text annotations, 1.3B region-text annotations, and 3.6B text-phrase-region annotations. filtered by [45]. This combination results in a dataset of 126 million images in total. 4.2. Data Annotation Our primary objective is to generate comprehensive annotations that can support multitask learning effectively. Accordingly, our annotation endeavors span a comprehensive range of tasks, encapsulated within three discrete annotation categories: text, region-text pairs, and text-phraseregion triplets, which is illustrated in Figure 4. The data annotation workflow consists of three essential phases, each of which ensures the accuracy and quality of the annotations: (1) initial annotation employing specialist models, (2) data filtering to correct errors and remove irrelevant annotations, and (3) an iterative process for data refinement. Initial annotation with specialist models. To initiate the annotation process for each annotation type, we employ synthetic labels obtained from specialist models. These specialist models are a combination of offline models trained on a diverse range of publicly available datasets and online services hosted on cloud platforms. They are specifically tailored to excel in annotating their respective annotation types. It is worth noting that certain image datasets may already contain partial annotations for some annotation types. For instance, the Object 365 [70] dataset already includes human-annotated bounding boxes and corresponding categories as region-text annotations. In such cases, we merge the pre-existing annotations with the synthetic labels generated by the specialist models. This approach enhances the coverage and diversity of the annotations. Moreover, specific annotations, such as detailed descriptions in the text annotation type, are represented by datasets of a considerably small size. This inherently poses challenges in obtaining high-performance specialist models. Consequently, we opt to omit these tasks during the initial annotation phase. Annotations for these tasks are generated later during the iterative data refinement process. In summation, through the rigorous initial annotation procedures, we ensure that the aggregated dataset of 126 million images is comprehensively labeled across the majority of annotation types. Data filtering and enhancement. The initial annotations obtained from the specialist models, while comprehensive, are susceptible to noise and imprecision. In response to this challenge, we have implemented a multifaceted filtering process to refine and eliminate undesired annotations. Our general filtering protocol mainly focuses on two data types in the annotations: text and region data. First, pertaining to textual annotations, we are inspired by DiHT [63] and develop a parsing tool based on SpaCy [28] to extract objects, attributes, and actions. We filter out texts containing excessive objects, as they tend to introduce noise and may not accurately reflect the actual content in the corresponding images. Additionally, we assess the complexity of the actions and objects by measuring their degree of node in the dependency parsing tree. We retain texts with a certain minimum action and object complexity to ensure the richness of visual concepts in the images. Second, in relation to the region annotations, specifically bounding boxes, we remove the noisy boxes under a confidence score threshold. Complementing this, we also employ non-maximum suppression to reduce redundant or overlapping bounding boxes. Iterative data refinement. Using our filtered initial annotations, we trained a multitask model that processes sequences of data. Upon evaluating this model against our training images, we discerned a marked enhancement in its predictions, particularly in instances where original labels were marred by inaccuracies or extraneous noise, such as in alt-texts. Motivated by these findings, we integrated these updated annotations with our original ones and subjected the model to another training iteration. This cyclical re5 finement process incrementally improves the quality of our training dataset. In the case of tasks we initially bypassed due to insufficient data for the training of a robust specialist model, we leveraged the iteratively trained model for pre-training purposes. Subsequent fine-tuning of this pre-trained model with the sparse dataset showcased superior performance compared to a model trained from scratch on the same data. Thus, we harness the fine-tuned model as a specialist for annotating our expansive dataset comprising 126 million images, ensuring comprehensive annotation coverage. 4.3. Annotation-specific Variations In Section 4.2, we introduce our general annotation workflow. This section delves into each annotation type and the corresponding variations of the annotation procedure. Text. Text annotations categorize images using three types of granularities: brief, detailed, and more detailed. The brief text includes only one sentence that demonstrates the most salient objects and activities, which is similar to COCO caption [13]. In contrast, the detailed text and more detailed text contain multiple sentences that describe the image with richer objects, attributes, and actions. For the brief text, a Florence-2 model is trained as the specialist on publicly available image caption and imagetext datasets, creating an image-to-text model for initial annotations. Iterative refinement is used to minimize noise in these texts. For the detailed text, prompts including existing image annotations like the brief text and region-text annotations, are fed to large language models (LLMs) or large multimodal models (LMMs) to generate comprehensive descriptions. Due to the high cost of the large models, only a small set of detailed text and more detailed text are generated. These are used to fine-tune the caption specialist, developing a detailed description specialist for further annotations. Region-text pairs. The region-text pairs provide descriptive textual annotation for semantic regions in the image. Semantic regions include regions of visual objects as well as text regions. The region is represented by a tight bounding box surrounds the region. Moreover, each region can be annotated with varying degrees of granularity, including phrases and sentences, that contribute to a richer understanding of the region. Region-text pairs are annotated differently for text regions and visual object regions. Text regions are labeled using Azure AI Services\u2019 OCR API [1], while visual objects are initially annotated with a DINO object detector [97] trained on public datasets. Data filtering, including confidence thresholding and non-maximum suppression, removes noisy boxes. Textual annotations for the visual object regions are further enriched by brief text generated from an image-to-text model with cropped image regions. Each region then receives three textual annotations: phrase from object category, brief text, and noun phrase chunks from the brief text. The Florence-1 [95] model determines the most similar textual annotation to each image region. Text-phrase-region triplets. Text-phrase-region triplets consist of a descriptive text of the image, noun phrases in this text related to image objects, and region annotations for these objects. The text includes brief, detailed, and more detailed text generated earlier. For each text, the Grounding DINO model [50] identifies noun phrases and creates bounding boxes for them. Additionally, the SAM model [32] generates segmentation masks for each box, offering more precise object localization. During data filtering, a confidence score threshold is applied to both noun phrases and bounding boxes to ensure relevance. A blacklist is also used to exclude irrelevant noun phrases like pronouns and abstract concepts. 5. Dataset This section introduces the statistics and analysis of FLD-5B that we built using the data engine in Section 4. We begin with an overview of the dataset and compare it with the recent works. We then show further analyses of detailed annotation statistics, semantic coverage and spatial coverage in the established dataset. 5.1. Overview Following the data engine, we build a large-scale training set (FLD-5B) of 126M images, more than 500M text annotations, 1.3B region-text annotations, and 3.6B textphrase-region annotations. Each image is annotated with text, region-text pairs, and text-phrase-region triplets and each annotation type has multiple instances varying in diverse granularity. An illustrative example of an image and its corresponding annotations can be found in Figure 4. We provide a comparison between our data set and the existing data sets that are commonly used for training foundation models in Table 1. Our data set has several advantages over the previous ones, such as having more annotations in total and per image. Moreover, the annotations in our data set span multiple levels of spatial and semantic granularity, which allows for more diverse and comprehensive visual understanding tasks. 5.2. Data Analysis Annotation statistics. The statistics for each annotation type within our dataset are presented in Table 2. Firstly, we have around 500M text annotations, including brief, detailed, and more detailed texts with different lengths. It is noteworthy that our detailed and more detailed text has 4x and 9x number of tokens compared with the brief text that is similar to COCO captions [13]. These lengthy 6 Text annotations Less granular (image level) More granular (image level) Region-text pairs annotations None semantic Rich semantic Text-phrase-region annotations Less granular (region level) More granular (region level) The image showsa person riding a red bicycle on a road witha red car in the background. The person is wearing a white t-shirt, black pants, and a black hat. She has a backpack on her back and is pedaling with their feet on the pedals. The road is lined with trees on both sides and there isanother person riding another bicyclein front of her. The date \"9/22/2023 \" is visible in the bottom right corner of the image. A woman riding a bike down a street next to a red car. The image shows a person riding a red bicycle on a road with a red car in the background. The person is wearing a white t-shirt, black pants, and a black hat. She has a backpack on her back and is pedaling with their feet on the pedals. The road is lined with trees on both sides and there is another person riding another bicycle in front of her. The date \"9/22/2023\" is visible in the bottom right corner of the image. person, car, road car person person red vintage car on street woman riding red bicycle with black backpack on city street person riding bicycle on street 9/22/2023 a person riding a red bicycle a red car another person riding another bicycle 9/22/2023 A woman riding a bike a red car A woman riding a bikedown a street next toa red car. Image level Region level Figure 4. An illustrative example of an image and its corresponding annotations in FLD-5B dataset. Each image in FLD-5B is annotated with text, region-text pairs, and text-phrase-region triplets by Florence data engine, which covers multiple spatial hierarchies, brief-to-detailed progressive granularity, and a wide semantics spectrum, enabling more comprehensive visual understanding from diverse perspectives. Dataset Rep. Model #Images #Annotations Spatial hierarchy Semantics granularity JFT300M [21] ViT 300M 300M Image-level Coarse WIT [64] CLIP 400M 400M Image-level Coarse SA-1B [32] SAM 11M 1B Region-level Non-semantic GrIT [60] Kosmos-2 91M 137M Image & Region-level Fine-grained M3W [2] Flamingo 185M 43.3M* Multi-image-level Fine-grained FLD-5B (ours) Florence-2 (ours) 126M 5B Image & Region-level Coarse to fine-grained Table 1. Comparison with datasets in vision foundation model training. *Flamingo\u2019s annotations are counted in the number of documents, where each document may have multiple images. annotations provide much richer information for comphrensive visual understanding. In addition, our dataset has around 1.3B region-text annotations, which is more than 30x larger than the academic object detection datasets such as OpenImages [40] and Object 365 [70]. On average, each image has around 5 regions, and each region is annotated with either a phrase or a relatively longer brief text. Note that the regional brief text (2.55 avg tokens) is shorter than typical brief text annotation (7.95 avg tokens), as the regional brief text annotation actually includes a mixture of phrase, noun chunks, and brief text based on the Florence-1 score. More details can be found from Section 4.3 region-text pairs. Moreover, we collect text-phrase-region annotations that include more than 3.6B phrase-region pairs for the 500M text annotations. Specifically, the brief text annotation has 4.27 average phrase-region pairs, while detailed and more detailed text annotation has more than 10 pairs, indicating that the richer text annotation covers more objects and their corresponding phrases in the text. Semantic coverage. Our text annotations comprise various text types, addressing different levels of detail. To assess 7 Annotation Type Text Type #Image Annotations #Avg Tokens #Regions #Avg Regions #Avg Regional Tokens Text Brief 235M 7.95 Detailed 126M 31.65 More detailed 126M 70.53 Region-Text Phrase 126M 681M 5.42 1.19 Brief 126M 681M 5.42 2.55 Text-Phrase-Region Brief 235M 7.95 1007M 4.27 1.93 Detailed 126M 31.65 1289M 10.25 1.49 More detailed 126M 70.53 1278M 10.17 1.35 Table 2. Annotation statistics of FLD-5B dataset. semantic coverage, we employ SpaCy [28] for tokenization and parsing, inspired by DiHT [63]. This process yields part-of-speech (POS) tags and the dependency parsing tree among tokens. We establish heuristic rules based on POS tags, categorizing tokens into semantic element types, e.g., objects, attributes, actions, and proper nouns. Additionally, we introduce the concept of token complexity, measured by the total degrees of the token in the dependency parsing tree when treated as an undirected graph. This complexity reflects the richness of semantic connections. In our study, we focus on measuring the complexity of objects and actions. Table 3 presents the statistics on the average number of semantic elements and their corresponding complexity. The results indicate that all measurements increase with the inclusion of more details in text annotations. Notably, average actions experience the most significant boost, with detailed and more detailed text exhibiting 7\u00d7 and 15\u00d7 increases, respectively, compared to brief text. This highlights the limitations of traditional brief text annotations in describing image actions. Conversely, the increment in proper nouns is relatively low, potentially because specialists often describe objects more generally than using specific proper nouns. In terms of complexity measurements, both objects and actions show more semantic connections in detailed text annotations. The complexity of actions exhibits a higher improvement, aligning with our observation of the increasing number of actions. Spatial coverage. Our region-text and text-phrase-region annotations, represented by bounding boxes and masks, capture the location of visual concepts within images. The distribution of box areas, as shown in Figure 5a, reveals more small boxes in region-text pairs and a uniform box size distribution in text-phrase-region triplets. This difference stems from the the divergent origins of these boxes: object detectors for region-text pairs and a grounding model for text-phrase-region triplets, which aligns boxes to textual phrases representing both localized and overarching Text Type Brief Detailed More detailed #Image Annotations 235M 126M 126M #Avg Tokens 7.95 31.65 70.53 #Avg Objects 3.23 13.31 28.06 #Avg Attributes 2.80 7.27 16.25 #Avg Actions 0.58 4.21 8.76 #Proper Nouns 1.10 2.40 2.41 Avg Object Complexity 2.80 4.00 4.02 Avg Action Complexity 1.14 3.63 4.38 Table 3. Statistics of the average number of semantic elements and corresponding complexity in FLD-5B dataset. image concepts. In Figure 5b, the log-format distribution of aspect ratios is illustrated. Region-text pairs and textphrase-region triplets exhibit similar symmetric distributions, covering a wide range of aspect ratios. Heatmaps of the box center for each annotation type, shown in Figures. 5c and 5d, indicate a center bias, with region-text pairs displaying a more uniform distribution than text-phraseregion triplets. 6. Experiments Our Florence-2 models are trained on FLD-5B to learn a universal image representation. We conduct our experiments in three main parts: (1) We evaluate the zero-shot performance of our method on various tasks to show its inherent ability to handle multiple tasks without any extra fine-tuning on task-specific data using one single generalist model. (2) We show the adaptability of our method by further training one single generalist model with additional supervised data on a wide range of tasks, achieving competitive state-of-the-art performance. (3) We examine the performance of the learned visual representation on the downstream tasks as the backbone to show the superiority of our pre-training method over previous approaches. 8 0.0 0.2 0.4 0.6 0.8 1.0 square-root normalized area 0 1 2 3 4 5 probability density Distribution of area of bounding boxes region-text text-phrase-region (a) 3 2 1 0 1 2 3 log aspect ratio 0.00 0.25 0.50 0.75 1.00 1.25 1.50 probability density Distribution of aspect ratio of bounding boxes region-text text-phrase-region (b) 0.0 0.2 0.4 0.6 0.8 1.0 normalized x 0.0 0.2 0.4 0.6 0.8 1.0 normalized y Distribution of box center  (region-text) 10 3 10 2 probability (c) 0.0 0.2 0.4 0.6 0.8 1.0 normalized x 0.0 0.2 0.4 0.6 0.8 1.0 normalized y Distribution of box center  (text-phrase-region) 10 4 10 3 10 2 probability (d) Figure 5. Distributions of bounding boxes in FLD-5B dataset. 6.1. Setup We investigate two model variants with different sizes: Florence-2-B model with 232 million parameters and Florence-2-L model with 771 million parameters. The detailed architectures of each model are given in Table 15. We initialize the weights of the image encoder and multi-modality encoder-decoder from UniCL [87] and BART [43], respectively. We adopt AdamW [54] with cosine learning rate decay [53] for training our models. We leverage Deepspeed [67] and mixed precision to improve the training efficiency. The maximum learning rate is set at 1e \u22124 for the base model and 1e \u22125 for the large model. A linear warmup to the maximum learning rate is applied during the first 5,000 optimization steps. We train our models with a mini-batch size of 2048/3072 (base/large) and an image size of 384\u00d7384 until reaching 3 billion effective training samples. Similar to [15, 29, 64, 92, 95], we further conduct high-resolution tuning with an image size of 768\u00d7768 for 0.5 billion samples for the base model and 0.1 billion samples for the large model. 6.2. Zero-shot Evaluation Across Tasks We present a powerful vision foundation model that does not require task-specific supervised annotations for finetuning. The zero-shot performance of our model is shown in Table 4. For image-level tasks, Florence-2-L achieves a 135.6 CIDEr score on the COCO caption benchmark [48], utilizing less than 1% of the parameters compared to the 80B Flamingo [2] model (which has an 84.3 CIDEr score). For region-level grounding and referring expression comprehension tasks, Florence-2-L establishes a new record in zero-shot performance achieving a 5.7 improvement in Flickr30k [61] Recall@1, and approximately 4%, 8%, and 8% absolute improvements on Refcoco, Refcoco+, and Refcocog [94], respectively, compared to the Kosmos-2 [60] model, which has 1.6B parameters. Additionally, our pretrained model attains a 35.8% mIOU in the Refcoco referring expression segmentation (RES) [94] task, a capability not supported by prior foundation models. 6.3. Generalist Model with Public Supervised Data We demonstrate the versatility and effectiveness of our model as a vision foundation that can be transferred to various downstream tasks. We fine-tune Florence-2 models by adding a collection of public datasets that cover imagelevel, region-level, pixel-level tasks, yielding one generalist model for various vision tasks. The details of the dataset collection are provided in Appendix B. Tables 5 and 6 compare our model with other state-of-the-art models. Our key findings are: Simple design for strong performance. Florence-2 demonstrates strong performance with standard multimodality Transformer encoder-decoder without special designs, particularly for region-level and pixel-level tasks. For example, Florence-2-L outperforms PolyFormer [49] on both RefCOCO REC task and RES task by 3.0 Accuracy@0.5 and 3.54 mIOU respectively, where PolyFormer [49] adapts specifically designed regression-based prediction head for coordinates. Florence-2-L also outperforms previous SOTA method UNINEXT [84] on RefCOCO by 0.8 Accuracy@0.5, where UNINEXT [84] is based on advanced object detector Deformable DETR [100] and DINO [97]. Competitive performance with fewer parameters. Florence-2-L achieves competitive performance without the need for LLMs, showcasing efficiency in handling diverse tasks while maintaining a compact size. For instance, Florence-2-L attains a CIDEr score of 140.0 on the COCO Caption karpathy test split [30], outperforming models with significantly more parameters, such as Flamingo (80B parameters, 138.1 CIDEr score). Adaptable generalization across task levels. Florence-2 demonstrates competitive performance across image-level, pixel-level, and region-level tasks, emphasizing its adaptability and effectiveness in addressing various challenges in computer vision and natural language processing. For example, in the TextVQA task, Florence-2-L sets a new stateof-the-art performance with an accuracy of 81.5 without any external OCR token input, surpassing previous SOTA meth9 Method #params COCO Cap. NoCaps TextCaps COCO Det. Flickr30k Refcoco Refcoco+ Refcocog Refcoco RES test val val val2017 test val test-A test-B val test-A test-B val test val CIDEr CIDEr CIDEr mAP R@1 Accuracy Accuracy Accuracy mIoU Flamingo [2] 80B 84.3 Kosmos-2 [60] 1.6B 78.7 52.3 57.4 47.3 45.5 50.7 42.2 60.6 61.7 Florence-2-B 0.23B 133.0 118.7 70.1 34.7 83.6 53.9 58.4 49.7 51.5 56.4 47.9 66.3 65.1 34.6 Florence-2-L 0.77B 135.6 120.8 72.8 37.5 84.4 56.3 61.6 51.4 53.6 57.9 49.9 68.0 67.0 35.8 Table 4. Zero-shot performance of generalist vision foundation models. The models do not see the training data of the evaluation tasks during training. Florence-2 models are pre-trained on FLD-5B dataset. Karpathy test split is used for COCO caption evaluation. Method #params COCO Caption NoCaps TextCaps VQAv2 TextVQA VizWiz VQA Karpathy test val val test-dev test-dev test-dev CIDEr CIDEr CIDEr Acc Acc Acc Specialist Models CoCa [92] 2.1B 143.6 122.4 82.3 BLIP-2 [44] 7.8B 144.5 121.6 82.2 GIT2 [78] 5.1B 145 126.9 148.6 81.7 67.3 71.0 Flamingo [2] 80B 138.1 82.0 54.1 65.7 PaLI [15] 17B 149.1 127.0 160.0\u25b3 84.3 58.8 / 73.1\u25b3 71.6 / 74.4\u25b3 PaLI-X [12] 55B 149.2 126.3 147 / 163.7\u25b3 86.0 71.4 / 80.8\u25b3 70.9 / 74.6\u25b3 Generalist Models Unified-IO [55] 2.9B 100 77.9 57.4 Florence-2-B 0.23B 140.0 116.7 143.9 79.7 63.6 63.6 Florence-2-L 0.77B 143.3 124.9 151.1 81.7 73.5 72.6 Table 5. Performance of specialist and generalist models on captioning and VQA tasks. Specialist Models refer to those that are fine-tuned specifically for each task, while Generalist Models denote a single model fine-tuned in a task-agnostic manner, applicable across all tasks. \u25b3indicates usage of external OCR as input. ods [12,15]. These achievements emphasize Florence-2\u2019s efficiency in handling diverse tasks while maintaining a compact size, making it a unique and valuable asset in the ever-evolving landscape of AI research and applications. 6.4. Downstream Tasks Fine-tuning In this section, we investigate the performance of our single model fine-tuning on downstream tasks. This experiment highlights the superiority of Florence-2 pre-training over previous approaches, as it demonstrates the effectiveness of the learned universal image representation. We use the base size model with about 80M parameters in our experiments to ensure fair comparison with other methods. Object detection and segmentation. We conduct COCO object detection and instance segmentation [48] experiments with Mask R-CNN [26], and COCO object detection [48] experiments with DINO [97] to further demonstrate the effectiveness of Florence-2 pre-training. We train on the train2017 split and evaluate on the val2017 split. For Mask R-CNN [26] experiments, we follow the common setup used in [51, 97], we use the standard 1\u00d7 (12 epochs) schedule with multi-scale training for all experiments. The learning rate is stepped down by a factor of 0.1 at the 67% and 89% of training epochs. We do not use any additional augmentation (such as random crop, mosaic, etc) or optimization techniques (such as EMA, weight normalization) during training to ensure a fair comparison. We do not use any test time augmentation (TTA) either. Thanks to the strong universal representation learned by Florence-2 pre-training, we do not require longer training epochs, such as 36 epochs in [51, 81, 85, 86], or 100 epochs in [46], to achieve better results. For DINO [97] experiments, we train DINO-4scale [97] detector for 12 epochs (1\u00d7) using the same data augmentation strategy as employed by [7]. 10 Method #params COCO Det. Flickr30k Refcoco Refcoco+ Refcocog Refcoco RES val2017 test val test-A test-B val test-A test-B val test val mAP R@1 Accuracy Accuracy Accuracy mIoU Specialist Models SeqTR [99] 83.7 86.5 81.2 71.5 76.3 64.9 74.9 74.2 PolyFormer [49] 90.4 92.9 87.2 85.0 89.8 78.0 85.8 85.9 76.9 UNINEXT [84] 0.74B 60.6 92.6 94.3 91.5 85.2 89.6 79.8 88.7 89.4 Ferret [90] 13B 89.5 92.4 84.4 82.8 88.1 75.2 85.8 86.3 Generalist Models UniTAB [88] 88.6 91.1 83.8 81.0 85.4 71.6 84.6 84.7 Florence-2-B 0.23B 41.4 84.0 92.6 94.8 91.5 86.8 91.7 82.2 89.8 82.2 78.0 Florence-2-L 0.77B 43.4 85.2 93.4 95.3 92.0 88.3 92.9 83.6 91.2 91.7 80.5 Table 6. Performance of specialist and generalist models on region-level tasks. Specialist Models refer to those that are fine-tuned specifically for each task, while Generalist Models denote a single model fine-tuned in a task-agnostic manner, applicable across all tasks. 4X efficiency gain 6.9 points improvements (a) Mask-RCNN on COCO detection. 5.5 points improvements 4X efficiency gain (b) DINO on COCO detection. 4X efficiency gain 5.9 points improvements (c) UpperNet on ADE20K. Figure 6. Training efficiency on COCO object detection and segmentation, and ADE20K semantic segmentation tasks. First, our base model achieves a strong performance improvement compared to other approaches. As shown in Table 7, our DaViT-B model pre-trained by Florence-2 surpasses previous best base model (ConvNext v2-B), which is pre-trained by FCMAE [81], by 0.7 APb using Mask RCNN. Importantly, while ConvNeXt v2-B leverages a 3\u00d7 schedule (36 epochs), our model efficiently employs a 1\u00d7 schedule (12 epochs) thanks to our powerful pre-trained universal representation. For DINO framework, our model significantly outperforms the ViT-B, achieving a notable improvement of 4.2 AP. Second, our pre-training demonstrates higher training efficiency. As shown in Table 8 and Figure 6, compared to the model with supervised ImageNet-1k pre-training, our model with Florence-2 pre-training achieves 4x efficiency and a significant improvement of 6.9 AP and 5.5 AP with Mask-RCNN and DINO framework, respectively. Third, our pre-training provides a good generic representation without extensive fine-tuning. Table 8 indicates that the models with Florence-2 pre-training maintains competitive performances when the first two stages are frozen with only 0.3 and 0.2 drops for Mask-RCNN and DINO, respectively. Moreover, our approach with completely frozen backbone can outperform the model with supervised ImageNet-1k pre-training by 1.6 and 2.4 for MaskRCNN and DINO. Semantic segmentation. We conduct semantic segmentation experiments with UperNet [82] framework on ADE20k [98] dataset. We mostly follow the training and evaluation protocols from Swin [51]. Specifically, we use input size 512\u00d7512 and train the model for 40k iterations with a batch size of 64. We adopt the AdamW [54] optimizer with the optimal learning rate searched from {8e4,4e-4,2e-4,1e-4}. Our results show a similar trend to the object detection experiments. As illustrated in Table 9, our base model outperforms the previous SoTA model, which is BEiT pretrained ViT-B [4], by 1.3 and 1.4 points in single-scale and multi-scale testing protocol, respectively. With the same backbone architecture of DaViT-B [20], Florence-2 pretrained model achieves a remarkable improvement of 4.9 points and 4\u00d7 efficiency compared to the ImageNet-1k pre11 Mask R-CNN DINO Backbone Pretrain APb APm AP ViT-B [46] MAE, IN-1k 51.6 45.9 55.0 Swin-B [51] Sup IN-1k 50.2 53.4 Swin-B [51] SimMIM [83] 52.3 FocalAtt-B [86] Sup IN-1k 49.0 43.7 FocalNet-B [85] Sup IN-1k 49.8 44.1 54.4 ConvNeXt v1-B [52] Sup IN-1k 50.3 44.9 52.6 ConvNeXt v2-B [81] Sup IN-1k 51.0 45.6 ConvNeXt v2-B [81] FCMAE 52.9 46.6 DaViT-B [20] Florence-2 53.6 46.4 59.2 Table 7. COCO object detection and instance segmentation results using Mask-RCNN framework, and COCO object detection results using DINO-4scale framework. All the entries use a base size model to ensure a fair comparison. For Mask-RCNN experiments, our method utilizes 1\u00d7 schedule (12 epochs), ViT-B use 100 epochs, all others use 3\u00d7 (36 epochs). For DINO experiments, all the entries use 1\u00d7 schedule except for ViT-B which uses 50 epochs. Pretrain Frozen stages Mask R-CNN DINO UperNet APb APm AP mIoU Sup IN1k n/a 46.7 42.0 53.7 49 UniCL [87] n/a 50.4 45.0 57.3 53.6 Florence-2 n/a 53.6 46.4 59.2 54.9 Florence-2 [1] 53.6 46.3 59.2 54.1 Florence-2 [1, 2] 53.3 46.1 59.0 54.4 Florence-2 [1, 2, 3] 49.5 42.9 56.7 49.6 Florence-2 [1, 2, 3, 4] 48.3 44.5 56.1 45.9 Table 8. Downstream task fine-tuning on COCO and ADE20K dataset. COCO object detection using Mask R-CNN and DINO. ADE20K semantic segmentation using UperNet. All entries use DaViT-B with 80M parameters as the backbone and standard 1\u00d7 schedule. trained counterpart as demonstrated in Table 8 and Figure 6. 6.5. Ablation Studies Multitask transfer. In this study, we aimed to identify the most effective pre-trained model for transfer learning across various downstream tasks in computer vision. We compared three different models, each pre-trained on a different combination of tasks: \u2022 Image-level Model: pre-trained on image-level tasks only \u2022 Image-Region Model: pre-trained on image-level and region-level tasks \u2022 Image-Region-Pixel Model: pre-trained on imagelevel, region-level, and pixel-level tasks Backbone Pretrain mIoU ms-mIoU ViT-B [24] Sup IN-1k 47.4 ViT-B [24] MAE IN-1k 48.1 ViT-B [4] BEiT 53.6 54.1 ViT-B [59] BEiTv2 IN-1k 53.1 ViT-B [59] BEiTv2 IN-22k 53.5 Swin-B [51] Sup IN-1k 48.1 49.7 Swin-B [51] Sup IN-22k 51.8 Swin-B [51] SimMIM [83] 52.8 FocalAtt-B [86] Sup IN-1k 49.0 50.5 FocalNet-B [85] Sup IN-1k 50.5 51.4 ConvNeXt v1-B [52] Sup IN-1k 49.9 ConvNeXt v2-B [81] Sup IN-1k 50.5 ConvNeXt v2-B [81] FCMAE 52.1 DaViT-B [20] Florence-2 54.9 55.5 Table 9. ADE20K semantic segmentation results using UperNet. The input size is 512\u00d7512 for all the entries, except for models with BEiT pre-trained, which use the input size of 640 \u00d7 640. For pre-training, we optimize all models for the same number of effective samples (72M) on a subset of our FLD5B dataset. These models are then transferred to a combined dataset with four downstream tasks, each representing a different level of task granularity: COCO caption (image-level task), COCO object detection (region-level task), Flickr30k grounding (region-level task), RefCOCO referring segmentation (pixel-level task). The results are shown in Figure 7. The results demonstrate that Image-Region-Pixel Model, pre-trained on all three levels of tasks, consistently demonstrated competitive performance across the four downstream tasks. For the COCO caption task, Image-Region-Pixel Model initially performs worse than Image-level Model and Image-Region Model but eventually achieve a final performance (133.4 CIDEr) that is only slightly worse than the other models (134.6 CIDEr). For the COCO object detection task, Image-RegionPixel Model outperforms Image-level Model by a significant margin (28.3 vs. 0.1) and was only slightly worse than Image-Region Model (29.7). For the Flickr30k grounding task, Image-Region-Pixel Model shows strong performance (78.1 recall@1), comparable to Image-Region Model (79.1 recall@1) and significantly better than Image-level Model (62.0 recall@1). For the RefCOCO referring segmentation task, ImageRegion-Pixel Model clearly outperforms both Image-level Model and Image-Region Model, achieving the highest performance (31.6 mIoU) compared to the other models (28.4 and 18.2 mIoU). Our findings suggest that the Image-Region-Pixel Model, which is pre-trained on tasks at the image, region, 12 2500 5000 7500 10000 12500 15000 17500 20000 Optimization Steps 1.29 1.30 1.31 1.32 1.33 1.34 1.35 CIDEr COCO Caption 2500 5000 7500 10000 12500 15000 17500 20000 Optimization Steps 0.05 0.10 0.15 0.20 0.25 0.30 AP COCO Object Detection 2500 5000 7500 10000 12500 15000 17500 20000 Optimization Steps 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Recall@1 Flickr30k Grounding 2500 5000 7500 10000 12500 15000 17500 20000 Optimization Steps 0 5 10 15 20 25 30 mIOU RefCOCO Referring Segmentation image-level image-level + region-level image-level + region-level + pixel-level Figure 7. Multitask transfer. We conduct experiments with three different versions of Florence-2 models, each trained on a different level of image annotation: image level, image and region level, and image, region, and pixel level. We then evaluate the transfer learning performance of these models on four downstream tasks: COCO caption, COCO object detection, Flickr30k grounding, and Refcoco referring segmentation. Model Caption Detection Grounding RES CIDEr AP Recall@1 mIOU oIOU Base 118.7 19.7 76.3 18.6 17.8 Large 124.4 22.6 78.2 21.5 19.1 Table 10. Model scaling. Zero-shot performance on COCO caption and COCO object detection, Flickr30k grounding, RefCOCO referring expression segmentation(RES). and pixel levels, is the most effective base model for transfer learning across various computer vision tasks. This model shows strong performance on all four downstream tasks we evaluated, and consistently outperforms the Image-level Model and matches or exceeds the Image-Region Model in performance. By pre-training a model on tasks at different levels of granularity, we can ensure that the base model is better prepared to handle a diverse range of downstream tasks, offering a versatile and robust solution for transfer learning in computer vision. Model scaling. We aimed to investigate the impact of increasing model capacity on zero-shot performance on various downstream tasks in computer vision. We compared two models: Florence-2-B and Florence-2-L, which have 232M and 771M parameters, respectively. The model architectures are described in Table 15. We show the zero-shot performance on four downstream tasks in Table 10. The large model clearly outperforms the base model across various downstream tasks. Data scaling. We conducted experiments to study how zero-shot performance on various computer vision tasks is affected by the scale of pre-training data. We used four different data sizes for pre-training: 0.12M, 0.36M, 1.2M, and 12M images. All models were trained with the same effective sample size (72M) on a subset of FLD-5B data. Table 11 presents the zero-shot performance results on Data Caption Detection Grounding RES size CIDEr AP Recall@1 mIOU oIOU 0.12M 102.8 16.1 74.0 15.9 16.6 0.36M 114.3 18.7 75.8 16.6 16.4 1.2M 118.1 18.9 76.3 19.3 18.4 12M 118.7 19.7 76.3 18.6 17.8 Table 11. Data scaling. Zero-shot performance on COCO caption, COCO object detection, Flickr30k grounding, COCORef referring segmentation. COCO caption, COCO object detection, Flickr30k grounding, and RefCoco referring segmentation (RES) tasks. We can observe a trend of improved zero-shot performance on the downstream tasks as the pre-training data size increases (except for RES, 1.2M data has slightly better performance compared to 12M). Our experiments on data scaling demonstrate that larger pre-training data sizes generally lead to improved zero-shot performance across a variety of downstream tasks in computer vision. This finding suggests that investing in larger pre-training datasets can provide a more effective and versatile foundation for handling a wide range of downstream tasks. Our approach to scaling data is significantly more efficient than relying solely on human annotations, as most of the annotation generation is performed using model inference. By leveraging specialist models to generate annotations, we can substantially reduce the time and cost associated with manual annotation efforts, which often involve labor-intensive processes and may be subject to human errors or inconsistencies. Furthermore, utilizing model-generated annotations enables us to scale the pre-training datasets more rapidly and efficiently, allowing us to explore the impact of larger data 13 Caption Detection Grounding RES V Pre L Pre CIDEr AP Recall@1 mIOU oIOU Freeze Vision Encoder \u2713 \u2713 120.0 6.9 66.3 9.9 13.6 Unfreeze Vision Encoder \u2713 81.3 4.9 69.0 15.3 15.6 \u2713 117.4 19.6 75.2 21.5 19.3 \u2713 \u2713 118.7 19.7 76.3 18.6 17.8 Table 12. Basic components. Zero-shot performance on COCO caption, COCO object detection, Flickr30k grounding, and COCORef referring segmentation. V Pre and L Pre indicate that using vision and language pre-training initialization, respectively. sizes on model performance across various downstream tasks in computer vision. This not only facilitates the development of more effective and versatile foundation models but also ensures that the annotation process remains sustainable and scalable as the need for high-quality labeled data continues to grow. In summary, our data scaling approach offers a more efficient alternative to traditional human annotation methods by harnessing the power of specialist models for annotation generation. This strategy enables us to accelerate the pretraining process, optimize model performance, and effectively manage the ever-increasing demand for labeled data in the field of computer vision. Training settings. We analyze the basic model training settings for the two primary components of our model, namely the vision encoder and the multi-modality encoder-decoder. The experiment results are presented in Table 12 We observe that freezing the vision encoders does not affect the performance on tasks that require image-level understanding, but it significantly degrades the performance on tasks that require region-level or pixel-level understanding (e.g., AP on COCO object detection drops from 19.7 to 6.9). Previous methods for pre-training vision foundation models mainly focus on image-level tasks (e.g., image classification [27,38], image-text contrastive learning [64,95]), which may not provide them with sufficient region-level and pixel-level skills for downstream tasks. Therefore, it is important to unfreeze the vision backbone, enabling it to learn region-level and pixel-level features for various downstream tasks. The effect of language pre-training weights on multimodal encoder-decoder tasks varies depending on the task. Tasks that require more text understanding, such as captioning and grounding, benefit slightly from using language pretraining weights (e.g., COCO caption, Flickr30k grounding). Tasks that are mostly vision-focused, such as object detection and region segmentation, do not gain much from using language pre-training weights (for COCO object detection, the gain is only 0.1; for RES tasks, which use only localization tokens, the drop is 2.91 mIOU). We investigate the effects of different training configurations on the performance of a foundation model in regionlevel and pixel-level tasks. We find that unfreezing the vision backbone is crucial for enhancing the model\u2019s ability to learn from regions and pixels, which is beneficial for transferring to various downstream tasks. Moreover, we observe that using language pre-training weights can help the model in tasks that require text understanding, but have less impact on tasks that are purely vision-based. These results offer useful guidance for choosing the best training settings for different computer vision tasks. The Florence Project endeavors to develop a foundational vision model endowed with a diverse array of perceptual capabilities, encompassing spatial hierarchy and semantic granularity. To this end, we construct FLD-5B dataset containing an extensive collection of 126M images paired with 5B comprehensive annotations, which are collected by the Florence data engine. Subsequently, we pretrain Florence-2 on this rich dataset through comprehensive multitask learning in a unified manner. Florence-2 has exhibited remarkable zero-shot capabilities that extend across a wide spectrum of visual tasks, such as captioning, object detection, visual grounding, and referring segmentation, among others. The experimental findings underscore the potency of the universal representation pre-trained by Florence-2, revealing its substantial contributions to the enhancement of a multitude of downstream tasks. Acknowledgment. We would like to express our heartfelt gratitude to all the contributors from the Azure AI team who worked on the Florence project. We sincerely appreciate Misha Bilenko for the invaluable guidance and support. Our thanks are extended to Yi-Ling Chen, Mengchen Liu, Yen-Chun Chen and Dongdong Chen for engaging in helpful discussions and to Yunsheng Li for their assistance with segmentation annotations. Deep appreciation is also expressed to Qingfen Lin, Ryan Menezes, Kuan Lu, Gabe Blanco, Shohei Ono, Ping Jin, Jiahe Zhou, Xiong Qiao, Tong Bai, Xingchao Peng, Pei Guo, Lihang Li for providing valuable feedback in downstream applications discussions. Special thanks to Cha Zhang, Jinyu Li, Min Gao, Christina Sun, Oliver Ernst, Kevin Pan, Mei Gao for their work on data annotation support and insightful discussions in data pipeline. Furthermore, we would like to thank Thomas Soemo, Nguyen Bach for their constructive feedback."
+                },
+                {
+                    "url": "http://arxiv.org/abs/1707.02968v2",
+                    "title": "Revisiting Unreasonable Effectiveness of Data in Deep Learning Era",
+                    "abstract": "The success of deep learning in vision can be attributed to: (a) models with\nhigh capacity; (b) increased computational power; and (c) availability of\nlarge-scale labeled data. Since 2012, there have been significant advances in\nrepresentation capabilities of the models and computational capabilities of\nGPUs. But the size of the biggest dataset has surprisingly remained constant.\nWhat will happen if we increase the dataset size by 10x or 100x? This paper\ntakes a step towards clearing the clouds of mystery surrounding the\nrelationship between `enormous data' and visual deep learning. By exploiting\nthe JFT-300M dataset which has more than 375M noisy labels for 300M images, we\ninvestigate how the performance of current vision tasks would change if this\ndata was used for representation learning. Our paper delivers some surprising\n(and some expected) findings. First, we find that the performance on vision\ntasks increases logarithmically based on volume of training data size. Second,\nwe show that representation learning (or pre-training) still holds a lot of\npromise. One can improve performance on many vision tasks by just training a\nbetter base model. Finally, as expected, we present new state-of-the-art\nresults for different vision tasks including image classification, object\ndetection, semantic segmentation and human pose estimation. Our sincere hope is\nthat this inspires vision community to not undervalue the data and develop\ncollective efforts in building larger datasets.",
+                    "authors": "Chen Sun, Abhinav Shrivastava, Saurabh Singh, Abhinav Gupta",
+                    "published": "2017-07-10",
+                    "updated": "2017-08-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1811.00982v2",
+                    "title": "The Open Images Dataset V4: Unified image classification, object detection, and visual relationship detection at scale",
+                    "abstract": "We present Open Images V4, a dataset of 9.2M images with unified annotations\nfor image classification, object detection and visual relationship detection.\nThe images have a Creative Commons Attribution license that allows to share and\nadapt the material, and they have been collected from Flickr without a\npredefined list of class names or tags, leading to natural class statistics and\navoiding an initial design bias. Open Images V4 offers large scale across\nseveral dimensions: 30.1M image-level labels for 19.8k concepts, 15.4M bounding\nboxes for 600 object classes, and 375k visual relationship annotations\ninvolving 57 classes. For object detection in particular, we provide 15x more\nbounding boxes than the next largest datasets (15.4M boxes on 1.9M images). The\nimages often show complex scenes with several objects (8 annotated objects per\nimage on average). We annotated visual relationships between them, which\nsupport visual relationship detection, an emerging task that requires\nstructured reasoning. We provide in-depth comprehensive statistics about the\ndataset, we validate the quality of the annotations, we study how the\nperformance of several modern models evolves with increasing amounts of\ntraining data, and we demonstrate two applications made possible by having\nunified annotations of multiple types coexisting in the same images. We hope\nthat the scale, quality, and variety of Open Images V4 will foster further\nresearch and innovation even beyond the areas of image classification, object\ndetection, and visual relationship detection.",
+                    "authors": "Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, Vittorio Ferrari",
+                    "published": "2018-11-02",
+                    "updated": "2020-02-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2205.01917v2",
+                    "title": "CoCa: Contrastive Captioners are Image-Text Foundation Models",
+                    "abstract": "Exploring large-scale pretrained foundation models is of significant interest\nin computer vision because these models can be quickly transferred to many\ndownstream tasks. This paper presents Contrastive Captioner (CoCa), a\nminimalist design to pretrain an image-text encoder-decoder foundation model\njointly with contrastive loss and captioning loss, thereby subsuming model\ncapabilities from contrastive approaches like CLIP and generative methods like\nSimVLM. In contrast to standard encoder-decoder transformers where all decoder\nlayers attend to encoder outputs, CoCa omits cross-attention in the first half\nof decoder layers to encode unimodal text representations, and cascades the\nremaining decoder layers which cross-attend to the image encoder for multimodal\nimage-text representations. We apply a contrastive loss between unimodal image\nand text embeddings, in addition to a captioning loss on the multimodal decoder\noutputs which predicts text tokens autoregressively. By sharing the same\ncomputational graph, the two training objectives are computed efficiently with\nminimal overhead. CoCa is pretrained end-to-end and from scratch on both\nweb-scale alt-text data and annotated images by treating all labels simply as\ntext, seamlessly unifying natural language supervision for representation\nlearning. Empirically, CoCa achieves state-of-the-art performance with\nzero-shot transfer or minimal task-specific adaptation on a broad range of\ndownstream tasks, spanning visual recognition (ImageNet, Kinetics-400/600/700,\nMoments-in-Time), crossmodal retrieval (MSCOCO, Flickr30K, MSR-VTT), multimodal\nunderstanding (VQA, SNLI-VE, NLVR2), and image captioning (MSCOCO, NoCaps).\nNotably on ImageNet classification, CoCa obtains 86.3% zero-shot top-1\naccuracy, 90.6% with a frozen encoder and learned classification head, and new\nstate-of-the-art 91.0% top-1 accuracy on ImageNet with a finetuned encoder.",
+                    "authors": "Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, Yonghui Wu",
+                    "published": "2022-05-04",
+                    "updated": "2022-06-14",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG",
+                        "cs.MM"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.07258v3",
+                    "title": "On the Opportunities and Risks of Foundation Models",
+                    "abstract": "AI is undergoing a paradigm shift with the rise of models (e.g., BERT,\nDALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a\nwide range of downstream tasks. We call these models foundation models to\nunderscore their critically central yet incomplete character. This report\nprovides a thorough account of the opportunities and risks of foundation\nmodels, ranging from their capabilities (e.g., language, vision, robotics,\nreasoning, human interaction) and technical principles(e.g., model\narchitectures, training procedures, data, systems, security, evaluation,\ntheory) to their applications (e.g., law, healthcare, education) and societal\nimpact (e.g., inequity, misuse, economic and environmental impact, legal and\nethical considerations). Though foundation models are based on standard deep\nlearning and transfer learning, their scale results in new emergent\ncapabilities,and their effectiveness across so many tasks incentivizes\nhomogenization. Homogenization provides powerful leverage but demands caution,\nas the defects of the foundation model are inherited by all the adapted models\ndownstream. Despite the impending widespread deployment of foundation models,\nwe currently lack a clear understanding of how they work, when they fail, and\nwhat they are even capable of due to their emergent properties. To tackle these\nquestions, we believe much of the critical research on foundation models will\nrequire deep interdisciplinary collaboration commensurate with their\nfundamentally sociotechnical nature.",
+                    "authors": "Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf Roohani, Camilo Ruiz, Jack Ryan, Christopher R\u00e9, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tram\u00e8r, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga, Jiaxuan You, Matei Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, Percy Liang",
+                    "published": "2021-08-16",
+                    "updated": "2022-07-12",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.AI",
+                        "cs.CY"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.16527v2",
+                    "title": "OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents",
+                    "abstract": "Large multimodal models trained on natural documents, which interleave images\nand text, outperform models trained on image-text pairs on various multimodal\nbenchmarks. However, the datasets used to train these models have not been\nreleased, and the collection process has not been fully specified. We introduce\nthe OBELICS dataset, an open web-scale filtered dataset of interleaved\nimage-text documents comprising 141 million web pages extracted from Common\nCrawl, 353 million associated images, and 115 billion text tokens. We describe\nthe dataset creation process, present comprehensive filtering rules, and\nprovide an analysis of the dataset's content. To show the viability of OBELICS,\nwe train vision and language models of 9 and 80 billion parameters named\nIDEFICS, and obtain competitive performance on different multimodal benchmarks.\nWe release our dataset, models and code.",
+                    "authors": "Hugo Lauren\u00e7on, Lucile Saulnier, L\u00e9o Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh",
+                    "published": "2023-06-21",
+                    "updated": "2023-08-21",
+                    "primary_cat": "cs.IR",
+                    "cats": [
+                        "cs.IR",
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2304.02643v1",
+                    "title": "Segment Anything",
+                    "abstract": "We introduce the Segment Anything (SA) project: a new task, model, and\ndataset for image segmentation. Using our efficient model in a data collection\nloop, we built the largest segmentation dataset to date (by far), with over 1\nbillion masks on 11M licensed and privacy respecting images. The model is\ndesigned and trained to be promptable, so it can transfer zero-shot to new\nimage distributions and tasks. We evaluate its capabilities on numerous tasks\nand find that its zero-shot performance is impressive -- often competitive with\nor even superior to prior fully supervised results. We are releasing the\nSegment Anything Model (SAM) and corresponding dataset (SA-1B) of 1B masks and\n11M images at https://segment-anything.com to foster research into foundation\nmodels for computer vision.",
+                    "authors": "Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Doll\u00e1r, Ross Girshick",
+                    "published": "2023-04-05",
+                    "updated": "2023-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2201.12086v2",
+                    "title": "BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation",
+                    "abstract": "Vision-Language Pre-training (VLP) has advanced the performance for many\nvision-language tasks. However, most existing pre-trained models only excel in\neither understanding-based tasks or generation-based tasks. Furthermore,\nperformance improvement has been largely achieved by scaling up the dataset\nwith noisy image-text pairs collected from the web, which is a suboptimal\nsource of supervision. In this paper, we propose BLIP, a new VLP framework\nwhich transfers flexibly to both vision-language understanding and generation\ntasks. BLIP effectively utilizes the noisy web data by bootstrapping the\ncaptions, where a captioner generates synthetic captions and a filter removes\nthe noisy ones. We achieve state-of-the-art results on a wide range of\nvision-language tasks, such as image-text retrieval (+2.7% in average\nrecall@1), image captioning (+2.8% in CIDEr), and VQA (+1.6% in VQA score).\nBLIP also demonstrates strong generalization ability when directly transferred\nto video-language tasks in a zero-shot manner. Code, models, and datasets are\nreleased at https://github.com/salesforce/BLIP.",
+                    "authors": "Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi",
+                    "published": "2022-01-28",
+                    "updated": "2022-02-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1405.0312v3",
+                    "title": "Microsoft COCO: Common Objects in Context",
+                    "abstract": "We present a new dataset with the goal of advancing the state-of-the-art in\nobject recognition by placing the question of object recognition in the context\nof the broader question of scene understanding. This is achieved by gathering\nimages of complex everyday scenes containing common objects in their natural\ncontext. Objects are labeled using per-instance segmentations to aid in precise\nobject localization. Our dataset contains photos of 91 objects types that would\nbe easily recognizable by a 4 year old. With a total of 2.5 million labeled\ninstances in 328k images, the creation of our dataset drew upon extensive crowd\nworker involvement via novel user interfaces for category detection, instance\nspotting and instance segmentation. We present a detailed statistical analysis\nof the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide\nbaseline performance analysis for bounding box and segmentation detection\nresults using a Deformable Parts Model.",
+                    "authors": "Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, Piotr Doll\u00e1r",
+                    "published": "2014-05-01",
+                    "updated": "2015-02-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2109.10852v2",
+                    "title": "Pix2seq: A Language Modeling Framework for Object Detection",
+                    "abstract": "We present Pix2Seq, a simple and generic framework for object detection.\nUnlike existing approaches that explicitly integrate prior knowledge about the\ntask, we cast object detection as a language modeling task conditioned on the\nobserved pixel inputs. Object descriptions (e.g., bounding boxes and class\nlabels) are expressed as sequences of discrete tokens, and we train a neural\nnetwork to perceive the image and generate the desired sequence. Our approach\nis based mainly on the intuition that if a neural network knows about where and\nwhat the objects are, we just need to teach it how to read them out. Beyond the\nuse of task-specific data augmentations, our approach makes minimal assumptions\nabout the task, yet it achieves competitive results on the challenging COCO\ndataset, compared to highly specialized and well optimized detection\nalgorithms.",
+                    "authors": "Ting Chen, Saurabh Saxena, Lala Li, David J. Fleet, Geoffrey Hinton",
+                    "published": "2021-09-22",
+                    "updated": "2022-03-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.CL",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2102.05918v2",
+                    "title": "Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision",
+                    "abstract": "Pre-trained representations are becoming crucial for many NLP and perception\ntasks. While representation learning in NLP has transitioned to training on raw\ntext without human annotations, visual and vision-language representations\nstill rely heavily on curated training datasets that are expensive or require\nexpert knowledge. For vision applications, representations are mostly learned\nusing datasets with explicit class labels such as ImageNet or OpenImages. For\nvision-language, popular datasets like Conceptual Captions, MSCOCO, or CLIP all\ninvolve a non-trivial data collection (and cleaning) process. This costly\ncuration process limits the size of datasets and hence hinders the scaling of\ntrained models. In this paper, we leverage a noisy dataset of over one billion\nimage alt-text pairs, obtained without expensive filtering or post-processing\nsteps in the Conceptual Captions dataset. A simple dual-encoder architecture\nlearns to align visual and language representations of the image and text pairs\nusing a contrastive loss. We show that the scale of our corpus can make up for\nits noise and leads to state-of-the-art representations even with such a simple\nlearning scheme. Our visual representation achieves strong performance when\ntransferred to classification tasks such as ImageNet and VTAB. The aligned\nvisual and language representations enables zero-shot image classification and\nalso set new state-of-the-art results on Flickr30K and MSCOCO image-text\nretrieval benchmarks, even when compared with more sophisticated\ncross-attention models. The representations also enable cross-modality search\nwith complex text and text + image queries.",
+                    "authors": "Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig",
+                    "published": "2021-02-11",
+                    "updated": "2021-06-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.CL",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2206.08916v2",
+                    "title": "Unified-IO: A Unified Model for Vision, Language, and Multi-Modal Tasks",
+                    "abstract": "We propose Unified-IO, a model that performs a large variety of AI tasks\nspanning classical computer vision tasks, including pose estimation, object\ndetection, depth estimation and image generation, vision-and-language tasks\nsuch as region captioning and referring expression, to natural language\nprocessing tasks such as question answering and paraphrasing. Developing a\nsingle unified model for such a large variety of tasks poses unique challenges\ndue to the heterogeneous inputs and outputs pertaining to each task, including\nRGB images, per-pixel maps, binary masks, bounding boxes, and language. We\nachieve this unification by homogenizing every supported input and output into\na sequence of discrete vocabulary tokens. This common representation across all\ntasks allows us to train a single transformer-based architecture, jointly on\nover 90 diverse datasets in the vision and language fields. Unified-IO is the\nfirst model capable of performing all 7 tasks on the GRIT benchmark and\nproduces strong results across 16 diverse benchmarks like NYUv2-Depth,\nImageNet, VQA2.0, OK-VQA, Swig, VizWizGround, BoolQ, and SciTail, with no\ntask-specific fine-tuning. Code and demos for Unified-IO are available at:\nhttps://unified-io.allenai.org.",
+                    "authors": "Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, Aniruddha Kembhavi",
+                    "published": "2022-06-17",
+                    "updated": "2022-10-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1602.07332v1",
+                    "title": "Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations",
+                    "abstract": "Despite progress in perceptual tasks such as image classification, computers\nstill perform poorly on cognitive tasks such as image description and question\nanswering. Cognition is core to tasks that involve not just recognizing, but\nreasoning about our visual world. However, models used to tackle the rich\ncontent in images for cognitive tasks are still being trained using the same\ndatasets designed for perceptual tasks. To achieve success at cognitive tasks,\nmodels need to understand the interactions and relationships between objects in\nan image. When asked \"What vehicle is the person riding?\", computers will need\nto identify the objects in an image as well as the relationships riding(man,\ncarriage) and pulling(horse, carriage) in order to answer correctly that \"the\nperson is riding a horse-drawn carriage\".\n  In this paper, we present the Visual Genome dataset to enable the modeling of\nsuch relationships. We collect dense annotations of objects, attributes, and\nrelationships within each image to learn these models. Specifically, our\ndataset contains over 100K images where each image has an average of 21\nobjects, 18 attributes, and 18 pairwise relationships between objects. We\ncanonicalize the objects, attributes, relationships, and noun phrases in region\ndescriptions and questions answer pairs to WordNet synsets. Together, these\nannotations represent the densest and largest dataset of image descriptions,\nobjects, attributes, relationships, and question answers.",
+                    "authors": "Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, Fei-Fei Li",
+                    "published": "2016-02-23",
+                    "updated": "2016-02-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2106.04560v2",
+                    "title": "Scaling Vision Transformers",
+                    "abstract": "Attention-based neural networks such as the Vision Transformer (ViT) have\nrecently attained state-of-the-art results on many computer vision benchmarks.\nScale is a primary ingredient in attaining excellent results, therefore,\nunderstanding a model's scaling properties is a key to designing future\ngenerations effectively. While the laws for scaling Transformer language models\nhave been studied, it is unknown how Vision Transformers scale. To address\nthis, we scale ViT models and data, both up and down, and characterize the\nrelationships between error rate, data, and compute. Along the way, we refine\nthe architecture and training of ViT, reducing memory consumption and\nincreasing accuracy of the resulting models. As a result, we successfully train\na ViT model with two billion parameters, which attains a new state-of-the-art\non ImageNet of 90.45% top-1 accuracy. The model also performs well for few-shot\ntransfer, for example, reaching 84.86% top-1 accuracy on ImageNet with only 10\nexamples per class.",
+                    "authors": "Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, Lucas Beyer",
+                    "published": "2021-06-08",
+                    "updated": "2022-06-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2205.14100v5",
+                    "title": "GIT: A Generative Image-to-text Transformer for Vision and Language",
+                    "abstract": "In this paper, we design and train a Generative Image-to-text Transformer,\nGIT, to unify vision-language tasks such as image/video captioning and question\nanswering. While generative models provide a consistent network architecture\nbetween pre-training and fine-tuning, existing work typically contains complex\nstructures (uni/multi-modal encoder/decoder) and depends on external modules\nsuch as object detectors/taggers and optical character recognition (OCR). In\nGIT, we simplify the architecture as one image encoder and one text decoder\nunder a single language modeling task. We also scale up the pre-training data\nand the model size to boost the model performance. Without bells and whistles,\nour GIT establishes new state of the arts on 12 challenging benchmarks with a\nlarge margin. For instance, our model surpasses the human performance for the\nfirst time on TextCaps (138.2 vs. 125.5 in CIDEr). Furthermore, we present a\nnew scheme of generation-based image classification and scene text recognition,\nachieving decent performance on standard benchmarks. Codes are released at\n\\url{https://github.com/microsoft/GenerativeImage2Text}.",
+                    "authors": "Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, Lijuan Wang",
+                    "published": "2022-05-27",
+                    "updated": "2022-12-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2202.03052v2",
+                    "title": "OFA: Unifying Architectures, Tasks, and Modalities Through a Simple Sequence-to-Sequence Learning Framework",
+                    "abstract": "In this work, we pursue a unified paradigm for multimodal pretraining to\nbreak the scaffolds of complex task/modality-specific customization. We propose\nOFA, a Task-Agnostic and Modality-Agnostic framework that supports Task\nComprehensiveness. OFA unifies a diverse set of cross-modal and unimodal tasks,\nincluding image generation, visual grounding, image captioning, image\nclassification, language modeling, etc., in a simple sequence-to-sequence\nlearning framework. OFA follows the instruction-based learning in both\npretraining and finetuning stages, requiring no extra task-specific layers for\ndownstream tasks. In comparison with the recent state-of-the-art vision &\nlanguage models that rely on extremely large cross-modal datasets, OFA is\npretrained on only 20M publicly available image-text pairs. Despite its\nsimplicity and relatively small-scale training data, OFA achieves new SOTAs in\na series of cross-modal tasks while attaining highly competitive performances\non uni-modal tasks. Our further analysis indicates that OFA can also\neffectively transfer to unseen tasks and unseen domains. Our code and models\nare publicly available at https://github.com/OFA-Sys/OFA.",
+                    "authors": "Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, Hongxia Yang",
+                    "published": "2022-02-07",
+                    "updated": "2022-06-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.CL"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2209.06794v4",
+                    "title": "PaLI: A Jointly-Scaled Multilingual Language-Image Model",
+                    "abstract": "Effective scaling and a flexible task interface enable large language models\nto excel at many tasks. We present PaLI (Pathways Language and Image model), a\nmodel that extends this approach to the joint modeling of language and vision.\nPaLI generates text based on visual and textual inputs, and with this interface\nperforms many vision, language, and multimodal tasks, in many languages. To\ntrain PaLI, we make use of large pre-trained encoder-decoder language models\nand Vision Transformers (ViTs). This allows us to capitalize on their existing\ncapabilities and leverage the substantial cost of training them. We find that\njoint scaling of the vision and language components is important. Since\nexisting Transformers for language are much larger than their vision\ncounterparts, we train a large, 4-billion parameter ViT (ViT-e) to quantify the\nbenefits from even larger-capacity vision models. To train PaLI, we create a\nlarge multilingual mix of pretraining tasks, based on a new image-text training\nset containing 10B images and texts in over 100 languages. PaLI achieves\nstate-of-the-art in multiple vision and language tasks (such as captioning,\nvisual question-answering, scene-text understanding), while retaining a simple,\nmodular, and scalable design.",
+                    "authors": "Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, Radu Soricut",
+                    "published": "2022-09-14",
+                    "updated": "2023-06-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.CL"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2103.00020v1",
+                    "title": "Learning Transferable Visual Models From Natural Language Supervision",
+                    "abstract": "State-of-the-art computer vision systems are trained to predict a fixed set\nof predetermined object categories. This restricted form of supervision limits\ntheir generality and usability since additional labeled data is needed to\nspecify any other visual concept. Learning directly from raw text about images\nis a promising alternative which leverages a much broader source of\nsupervision. We demonstrate that the simple pre-training task of predicting\nwhich caption goes with which image is an efficient and scalable way to learn\nSOTA image representations from scratch on a dataset of 400 million (image,\ntext) pairs collected from the internet. After pre-training, natural language\nis used to reference learned visual concepts (or describe new ones) enabling\nzero-shot transfer of the model to downstream tasks. We study the performance\nof this approach by benchmarking on over 30 different existing computer vision\ndatasets, spanning tasks such as OCR, action recognition in videos,\ngeo-localization, and many types of fine-grained object classification. The\nmodel transfers non-trivially to most tasks and is often competitive with a\nfully supervised baseline without the need for any dataset specific training.\nFor instance, we match the accuracy of the original ResNet-50 on ImageNet\nzero-shot without needing to use any of the 1.28 million training examples it\nwas trained on. We release our code and pre-trained model weights at\nhttps://github.com/OpenAI/CLIP.",
+                    "authors": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever",
+                    "published": "2021-02-26",
+                    "updated": "2021-02-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1504.00325v2",
+                    "title": "Microsoft COCO Captions: Data Collection and Evaluation Server",
+                    "abstract": "In this paper we describe the Microsoft COCO Caption dataset and evaluation\nserver. When completed, the dataset will contain over one and a half million\ncaptions describing over 330,000 images. For the training and validation\nimages, five independent human generated captions will be provided. To ensure\nconsistency in evaluation of automatic caption generation algorithms, an\nevaluation server is used. The evaluation server receives candidate captions\nand scores them using several popular metrics, including BLEU, METEOR, ROUGE\nand CIDEr. Instructions for using the evaluation server are provided.",
+                    "authors": "Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, C. Lawrence Zitnick",
+                    "published": "2015-04-01",
+                    "updated": "2015-04-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.CL"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.03748v2",
+                    "title": "Representation Learning with Contrastive Predictive Coding",
+                    "abstract": "While supervised learning has enabled great progress in many applications,\nunsupervised learning has not seen such widespread adoption, and remains an\nimportant and challenging endeavor for artificial intelligence. In this work,\nwe propose a universal unsupervised learning approach to extract useful\nrepresentations from high-dimensional data, which we call Contrastive\nPredictive Coding. The key insight of our model is to learn such\nrepresentations by predicting the future in latent space by using powerful\nautoregressive models. We use a probabilistic contrastive loss which induces\nthe latent space to capture information that is maximally useful to predict\nfuture samples. It also makes the model tractable by using negative sampling.\nWhile most prior work has focused on evaluating representations for a\nparticular modality, we demonstrate that our approach is able to learn useful\nrepresentations achieving strong performance on four distinct domains: speech,\nimages, text and reinforcement learning in 3D environments.",
+                    "authors": "Aaron van den Oord, Yazhe Li, Oriol Vinyals",
+                    "published": "2018-07-10",
+                    "updated": "2019-01-22",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2204.14198v2",
+                    "title": "Flamingo: a Visual Language Model for Few-Shot Learning",
+                    "abstract": "Building models that can be rapidly adapted to novel tasks using only a\nhandful of annotated examples is an open challenge for multimodal machine\nlearning research. We introduce Flamingo, a family of Visual Language Models\n(VLM) with this ability. We propose key architectural innovations to: (i)\nbridge powerful pretrained vision-only and language-only models, (ii) handle\nsequences of arbitrarily interleaved visual and textual data, and (iii)\nseamlessly ingest images or videos as inputs. Thanks to their flexibility,\nFlamingo models can be trained on large-scale multimodal web corpora containing\narbitrarily interleaved text and images, which is key to endow them with\nin-context few-shot learning capabilities. We perform a thorough evaluation of\nour models, exploring and measuring their ability to rapidly adapt to a variety\nof image and video tasks. These include open-ended tasks such as visual\nquestion-answering, where the model is prompted with a question which it has to\nanswer; captioning tasks, which evaluate the ability to describe a scene or an\nevent; and close-ended tasks such as multiple-choice visual question-answering.\nFor tasks lying anywhere on this spectrum, a single Flamingo model can achieve\na new state of the art with few-shot learning, simply by prompting the model\nwith task-specific examples. On numerous benchmarks, Flamingo outperforms\nmodels fine-tuned on thousands of times more task-specific data.",
+                    "authors": "Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, Karen Simonyan",
+                    "published": "2022-04-29",
+                    "updated": "2022-11-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2111.11432v1",
+                    "title": "Florence: A New Foundation Model for Computer Vision",
+                    "abstract": "Automated visual understanding of our diverse and open world demands computer\nvision models to generalize well with minimal customization for specific tasks,\nsimilar to human vision. Computer vision foundation models, which are trained\non diverse, large-scale dataset and can be adapted to a wide range of\ndownstream tasks, are critical for this mission to solve real-world computer\nvision applications. While existing vision foundation models such as CLIP,\nALIGN, and Wu Dao 2.0 focus mainly on mapping images and textual\nrepresentations to a cross-modal shared representation, we introduce a new\ncomputer vision foundation model, Florence, to expand the representations from\ncoarse (scene) to fine (object), from static (images) to dynamic (videos), and\nfrom RGB to multiple modalities (caption, depth). By incorporating universal\nvisual-language representations from Web-scale image-text data, our Florence\nmodel can be easily adapted for various computer vision tasks, such as\nclassification, retrieval, object detection, VQA, image caption, video\nretrieval and action recognition. Moreover, Florence demonstrates outstanding\nperformance in many types of transfer learning: fully sampled fine-tuning,\nlinear probing, few-shot transfer and zero-shot transfer for novel images and\nobjects. All of these properties are critical for our vision foundation model\nto serve general purpose vision tasks. Florence achieves new state-of-the-art\nresults in majority of 44 representative benchmarks, e.g., ImageNet-1K\nzero-shot classification with top-1 accuracy of 83.74 and the top-5 accuracy of\n97.18, 62.4 mAP on COCO fine tuning, 80.36 on VQA, and 87.8 on Kinetics-600.",
+                    "authors": "Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, Pengchuan Zhang",
+                    "published": "2021-11-22",
+                    "updated": "2021-11-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.10782v2",
+                    "title": "Three Ways to Improve Semantic Segmentation with Self-Supervised Depth Estimation",
+                    "abstract": "Training deep networks for semantic segmentation requires large amounts of\nlabeled training data, which presents a major challenge in practice, as\nlabeling segmentation masks is a highly labor-intensive process. To address\nthis issue, we present a framework for semi-supervised semantic segmentation,\nwhich is enhanced by self-supervised monocular depth estimation from unlabeled\nimage sequences. In particular, we propose three key contributions: (1) We\ntransfer knowledge from features learned during self-supervised depth\nestimation to semantic segmentation, (2) we implement a strong data\naugmentation by blending images and labels using the geometry of the scene, and\n(3) we utilize the depth feature diversity as well as the level of difficulty\nof learning depth in a student-teacher framework to select the most useful\nsamples to be annotated for semantic segmentation. We validate the proposed\nmodel on the Cityscapes dataset, where all three modules demonstrate\nsignificant performance gains, and we achieve state-of-the-art results for\nsemi-supervised semantic segmentation. The implementation is available at\nhttps://github.com/lhoyer/improving_segmentation_with_selfsupervised_depth.",
+                    "authors": "Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian K\u00f6ring, Suman Saha, Luc Van Gool",
+                    "published": "2020-12-19",
+                    "updated": "2021-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2211.08352v1",
+                    "title": "Visual Semantic Segmentation Based on Few/Zero-Shot Learning: An Overview",
+                    "abstract": "Visual semantic segmentation aims at separating a visual sample into diverse\nblocks with specific semantic attributes and identifying the category for each\nblock, and it plays a crucial role in environmental perception. Conventional\nlearning-based visual semantic segmentation approaches count heavily on\nlarge-scale training data with dense annotations and consistently fail to\nestimate accurate semantic labels for unseen categories. This obstruction spurs\na craze for studying visual semantic segmentation with the assistance of\nfew/zero-shot learning. The emergence and rapid progress of few/zero-shot\nvisual semantic segmentation make it possible to learn unseen-category from a\nfew labeled or zero-labeled samples, which advances the extension to practical\napplications. Therefore, this paper focuses on the recently published\nfew/zero-shot visual semantic segmentation methods varying from 2D to 3D space\nand explores the commonalities and discrepancies of technical settlements under\ndifferent segmentation circumstances. Specifically, the preliminaries on\nfew/zero-shot visual semantic segmentation, including the problem definitions,\ntypical datasets, and technical remedies, are briefly reviewed and discussed.\nMoreover, three typical instantiations are involved to uncover the interactions\nof few/zero-shot learning with visual semantic segmentation, including image\nsemantic segmentation, video object segmentation, and 3D segmentation. Finally,\nthe future challenges of few/zero-shot visual semantic segmentation are\ndiscussed.",
+                    "authors": "Wenqi Ren, Yang Tang, Qiyu Sun, Chaoqiang Zhao, Qing-Long Han",
+                    "published": "2022-11-13",
+                    "updated": "2022-11-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.00242v1",
+                    "title": "3D Guided Weakly Supervised Semantic Segmentation",
+                    "abstract": "Pixel-wise clean annotation is necessary for fully-supervised semantic\nsegmentation, which is laborious and expensive to obtain. In this paper, we\npropose a weakly supervised 2D semantic segmentation model by incorporating\nsparse bounding box labels with available 3D information, which is much easier\nto obtain with advanced sensors. We manually labeled a subset of the 2D-3D\nSemantics(2D-3D-S) dataset with bounding boxes, and introduce our 2D-3D\ninference module to generate accurate pixel-wise segment proposal masks. Guided\nby 3D information, we first generate a point cloud of objects and calculate\nobjectness probability score for each point. Then we project the point cloud\nwith objectness probabilities back to 2D images followed by a refinement step\nto obtain segment proposals, which are treated as pseudo labels to train a\nsemantic segmentation network. Our method works in a recursive manner to\ngradually refine the above-mentioned segment proposals. Extensive experimental\nresults on the 2D-3D-S dataset show that the proposed method can generate\naccurate segment proposals when bounding box labels are available on only a\nsmall subset of training images. Performance comparison with recent\nstate-of-the-art methods further illustrates the effectiveness of our method.",
+                    "authors": "Weixuan Sun, Jing Zhang, Nick Barnes",
+                    "published": "2020-12-01",
+                    "updated": "2020-12-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2404.04807v1",
+                    "title": "D2SL: Decouple Defogging and Semantic Learning for Foggy Domain-Adaptive Segmentation",
+                    "abstract": "We investigated domain adaptive semantic segmentation in foggy weather\nscenarios, which aims to enhance the utilization of unlabeled foggy data and\nimprove the model's adaptability to foggy conditions. Current methods rely on\nclear images as references, jointly learning defogging and segmentation for\nfoggy images. Despite making some progress, there are still two main drawbacks:\n(1) the coupling of segmentation and defogging feature representations,\nresulting in a decrease in semantic representation capability, and (2) the\nfailure to leverage real fog priors in unlabeled foggy data, leading to\ninsufficient model generalization ability. To address these issues, we propose\na novel training framework, Decouple Defogging and Semantic learning, called\nD2SL, aiming to alleviate the adverse impact of defogging tasks on the final\nsegmentation task. In this framework, we introduce a domain-consistent transfer\nstrategy to establish a connection between defogging and segmentation tasks.\nFurthermore, we design a real fog transfer strategy to improve defogging\neffects by fully leveraging the fog priors from real foggy images. Our approach\nenhances the semantic representations required for segmentation during the\ndefogging learning process and maximizes the representation capability of fog\ninvariance by effectively utilizing real fog data. Comprehensive experiments\nvalidate the effectiveness of the proposed method.",
+                    "authors": "Xuan Sun, Zhanfu An, Yuyu Liu",
+                    "published": "2024-04-07",
+                    "updated": "2024-04-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.MM"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2404.04231v1",
+                    "title": "Image-Text Co-Decomposition for Text-Supervised Semantic Segmentation",
+                    "abstract": "This paper addresses text-supervised semantic segmentation, aiming to learn a\nmodel capable of segmenting arbitrary visual concepts within images by using\nonly image-text pairs without dense annotations. Existing methods have\ndemonstrated that contrastive learning on image-text pairs effectively aligns\nvisual segments with the meanings of texts. We notice that there is a\ndiscrepancy between text alignment and semantic segmentation: A text often\nconsists of multiple semantic concepts, whereas semantic segmentation strives\nto create semantically homogeneous segments. To address this issue, we propose\na novel framework, Image-Text Co-Decomposition (CoDe), where the paired image\nand text are jointly decomposed into a set of image regions and a set of word\nsegments, respectively, and contrastive learning is developed to enforce\nregion-word alignment. To work with a vision-language model, we present a\nprompt learning mechanism that derives an extra representation to highlight an\nimage segment or a word segment of interest, with which more effective features\ncan be extracted from that segment. Comprehensive experimental results\ndemonstrate that our method performs favorably against existing text-supervised\nsemantic segmentation methods on six benchmark datasets.",
+                    "authors": "Ji-Jia Wu, Andy Chia-Hao Chang, Chieh-Yu Chuang, Chun-Pei Chen, Yu-Lun Liu, Min-Hung Chen, Hou-Ning Hu, Yung-Yu Chuang, Yen-Yu Lin",
+                    "published": "2024-04-05",
+                    "updated": "2024-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2203.15943v1",
+                    "title": "Self-Supervised Leaf Segmentation under Complex Lighting Conditions",
+                    "abstract": "As an essential prerequisite task in image-based plant phenotyping, leaf\nsegmentation has garnered increasing attention in recent years. While\nself-supervised learning is emerging as an effective alternative to various\ncomputer vision tasks, its adaptation for image-based plant phenotyping remains\nrather unexplored. In this work, we present a self-supervised leaf segmentation\nframework consisting of a self-supervised semantic segmentation model, a\ncolor-based leaf segmentation algorithm, and a self-supervised color correction\nmodel. The self-supervised semantic segmentation model groups the semantically\nsimilar pixels by iteratively referring to the self-contained information,\nallowing the pixels of the same semantic object to be jointly considered by the\ncolor-based leaf segmentation algorithm for identifying the leaf regions.\nAdditionally, we propose to use a self-supervised color correction model for\nimages taken under complex illumination conditions. Experimental results on\ndatasets of different plant species demonstrate the potential of the proposed\nself-supervised framework in achieving effective and generalizable leaf\nsegmentation.",
+                    "authors": "Xufeng Lin, Chang-Tsun Li, Scott Adams, Abbas Kouzani, Richard Jiang, Ligang He, Yongjian Hu, Michael Vernon, Egan Doeven, Lawrence Webb, Todd Mcclellan, Adam Guskic",
+                    "published": "2022-03-29",
+                    "updated": "2022-03-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1812.10885v1",
+                    "title": "Coarse-to-fine Semantic Segmentation from Image-level Labels",
+                    "abstract": "Deep neural network-based semantic segmentation generally requires\nlarge-scale cost extensive annotations for training to obtain better\nperformance. To avoid pixel-wise segmentation annotations which are needed for\nmost methods, recently some researchers attempted to use object-level labels\n(e.g. bounding boxes) or image-level labels (e.g. image categories). In this\npaper, we propose a novel recursive coarse-to-fine semantic segmentation\nframework based on only image-level category labels. For each image, an initial\ncoarse mask is first generated by a convolutional neural network-based\nunsupervised foreground segmentation model and then is enhanced by a graph\nmodel. The enhanced coarse mask is fed to a fully convolutional neural network\nto be recursively refined. Unlike existing image-level label-based semantic\nsegmentation methods which require to label all categories for images contain\nmultiple types of objects, our framework only needs one label for each image\nand can handle images contains multi-category objects. With only trained on\nImageNet, our framework achieves comparable performance on PASCAL VOC dataset\nas other image-level label-based state-of-the-arts of semantic segmentation.\nFurthermore, our framework can be easily extended to foreground object\nsegmentation task and achieves comparable performance with the state-of-the-art\nsupervised methods on the Internet Object dataset.",
+                    "authors": "Longlong Jing, Yucheng Chen, Yingli Tian",
+                    "published": "2018-12-28",
+                    "updated": "2018-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2102.12095v1",
+                    "title": "Synergy Between Semantic Segmentation and Image Denoising via Alternate Boosting",
+                    "abstract": "The capability of image semantic segmentation may be deteriorated due to\nnoisy input image, where image denoising prior to segmentation helps. Both\nimage denoising and semantic segmentation have been developed significantly\nwith the advance of deep learning. Thus, we are interested in the synergy\nbetween them by using a holistic deep model. We observe that not only denoising\nhelps combat the drop of segmentation accuracy due to noise, but also\npixel-wise semantic information boosts the capability of denoising. We then\npropose a boosting network to perform denoising and segmentation alternately.\nThe proposed network is composed of multiple segmentation and denoising blocks\n(SDBs), each of which estimates semantic map then uses the map to regularize\ndenoising. Experimental results show that the denoised image quality is\nimproved substantially and the segmentation accuracy is improved to close to\nthat of clean images. Our code and models will be made publicly available.",
+                    "authors": "Shunxin Xu, Ke Sun, Dong Liu, Zhiwei Xiong, Zheng-Jun Zha",
+                    "published": "2021-02-24",
+                    "updated": "2021-02-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.04618v1",
+                    "title": "Multi-Scale Semantic Segmentation with Modified MBConv Blocks",
+                    "abstract": "Recently, MBConv blocks, initially designed for efficiency in\nresource-limited settings and later adapted for cutting-edge image\nclassification performances, have demonstrated significant potential in image\nclassification tasks. Despite their success, their application in semantic\nsegmentation has remained relatively unexplored. This paper introduces a novel\nadaptation of MBConv blocks specifically tailored for semantic segmentation.\nOur modification stems from the insight that semantic segmentation requires the\nextraction of more detailed spatial information than image classification. We\nargue that to effectively perform multi-scale semantic segmentation, each\nbranch of a U-Net architecture, regardless of its resolution, should possess\nequivalent segmentation capabilities. By implementing these changes, our\napproach achieves impressive mean Intersection over Union (IoU) scores of 84.5%\nand 84.0% on the Cityscapes test and validation datasets, respectively,\ndemonstrating the efficacy of our proposed modifications in enhancing semantic\nsegmentation performance.",
+                    "authors": "Xi Chen, Yang Cai, Yuan Wu, Bo Xiong, Taesung Park",
+                    "published": "2024-02-07",
+                    "updated": "2024-02-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.13026v1",
+                    "title": "Weakly-Supervised Semantic Segmentation with Image-Level Labels: from Traditional Models to Foundation Models",
+                    "abstract": "The rapid development of deep learning has driven significant progress in the\nfield of image semantic segmentation - a fundamental task in computer vision.\nSemantic segmentation algorithms often depend on the availability of\npixel-level labels (i.e., masks of objects), which are expensive,\ntime-consuming, and labor-intensive. Weakly-supervised semantic segmentation\n(WSSS) is an effective solution to avoid such labeling. It utilizes only\npartial or incomplete annotations and provides a cost-effective alternative to\nfully-supervised semantic segmentation. In this paper, we focus on the WSSS\nwith image-level labels, which is the most challenging form of WSSS. Our work\nhas two parts. First, we conduct a comprehensive survey on traditional methods,\nprimarily focusing on those presented at premier research conferences. We\ncategorize them into four groups based on where their methods operate:\npixel-wise, image-wise, cross-image, and external data. Second, we investigate\nthe applicability of visual foundation models, such as the Segment Anything\nModel (SAM), in the context of WSSS. We scrutinize SAM in two intriguing\nscenarios: text prompting and zero-shot learning. We provide insights into the\npotential and challenges associated with deploying visual foundational models\nfor WSSS, facilitating future developments in this exciting research area.",
+                    "authors": "Zhaozheng Chen, Qianru Sun",
+                    "published": "2023-10-19",
+                    "updated": "2023-10-19",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.11506v1",
+                    "title": "LCCo: Lending CLIP to Co-Segmentation",
+                    "abstract": "This paper studies co-segmenting the common semantic object in a set of\nimages. Existing works either rely on carefully engineered networks to mine the\nimplicit semantic information in visual features or require extra data (i.e.,\nclassification labels) for training. In this paper, we leverage the contrastive\nlanguage-image pre-training framework (CLIP) for the task. With a backbone\nsegmentation network that independently processes each image from the set, we\nintroduce semantics from CLIP into the backbone features, refining them in a\ncoarse-to-fine manner with three key modules: i) an image set feature\ncorrespondence module, encoding global consistent semantic information of the\nimage set; ii) a CLIP interaction module, using CLIP-mined common semantics of\nthe image set to refine the backbone feature; iii) a CLIP regularization\nmodule, drawing CLIP towards this co-segmentation task, identifying the best\nCLIP semantic and using it to regularize the backbone feature. Experiments on\nfour standard co-segmentation benchmark datasets show that the performance of\nour method outperforms state-of-the-art methods.",
+                    "authors": "Xin Duan, Yan Yang, Liyuan Pan, Xiabi Liu",
+                    "published": "2023-08-22",
+                    "updated": "2023-08-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2003.04404v1",
+                    "title": "FusionLane: Multi-Sensor Fusion for Lane Marking Semantic Segmentation Using Deep Neural Networks",
+                    "abstract": "It is a crucial step to achieve effective semantic segmentation of lane\nmarking during the construction of the lane level high-precision map. In recent\nyears, many image semantic segmentation methods have been proposed. These\nmethods mainly focus on the image from camera, due to the limitation of the\nsensor itself, the accurate three-dimensional spatial position of the lane\nmarking cannot be obtained, so the demand for the lane level high-precision map\nconstruction cannot be met. This paper proposes a lane marking semantic\nsegmentation method based on LIDAR and camera fusion deep neural network.\nDifferent from other methods, in order to obtain accurate position information\nof the segmentation results, the semantic segmentation object of this paper is\na bird's eye view converted from a LIDAR points cloud instead of an image\ncaptured by a camera. This method first uses the deeplabv3+ [\\ref{ref:1}]\nnetwork to segment the image captured by the camera, and the segmentation\nresult is merged with the point clouds collected by the LIDAR as the input of\nthe proposed network. In this neural network, we also add a long short-term\nmemory (LSTM) structure to assist the network for semantic segmentation of lane\nmarkings by using the the time series information. The experiments on more than\n14,000 image datasets which we have manually labeled and expanded have shown\nthe proposed method has better performance on the semantic segmentation of the\npoints cloud bird's eye view. Therefore, the automation of high-precision map\nconstruction can be significantly improved. Our code is available at\nhttps://github.com/rolandying/FusionLane.",
+                    "authors": "Ruochen Yin, Biao Yu, Huapeng Wu, Yutao Song, Runxin Niu",
+                    "published": "2020-03-09",
+                    "updated": "2020-03-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2307.13215v1",
+                    "title": "Image Segmentation Keras : Implementation of Segnet, FCN, UNet, PSPNet and other models in Keras",
+                    "abstract": "Semantic segmentation plays a vital role in computer vision tasks, enabling\nprecise pixel-level understanding of images. In this paper, we present a\ncomprehensive library for semantic segmentation, which contains implementations\nof popular segmentation models like SegNet, FCN, UNet, and PSPNet. We also\nevaluate and compare these models on several datasets, offering researchers and\npractitioners a powerful toolset for tackling diverse segmentation challenges.",
+                    "authors": "Divam Gupta",
+                    "published": "2023-07-25",
+                    "updated": "2023-07-25",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1811.00174v4",
+                    "title": "Pixel Level Data Augmentation for Semantic Image Segmentation using Generative Adversarial Networks",
+                    "abstract": "Semantic segmentation is one of the basic topics in computer vision, it aims\nto assign semantic labels to every pixel of an image. Unbalanced semantic label\ndistribution could have a negative influence on segmentation accuracy. In this\npaper, we investigate using data augmentation approach to balance the semantic\nlabel distribution in order to improve segmentation performance. We propose\nusing generative adversarial networks (GANs) to generate realistic images for\nimproving the performance of semantic segmentation networks. Experimental\nresults show that the proposed method can not only improve segmentation\nperformance on those classes with low accuracy, but also obtain 1.3% to 2.1%\nincrease in average segmentation accuracy. It shows that this augmentation\nmethod can boost accuracy and be easily applicable to any other segmentation\nmodels.",
+                    "authors": "Shuangting Liu, Jiaqi Zhang, Yuxin Chen, Yifan Liu, Zengchang Qin, Tao Wan",
+                    "published": "2018-11-01",
+                    "updated": "2019-11-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2101.08418v2",
+                    "title": "Rethinking Semantic Segmentation Evaluation for Explainability and Model Selection",
+                    "abstract": "Semantic segmentation aims to robustly predict coherent class labels for\nentire regions of an image. It is a scene understanding task that powers\nreal-world applications (e.g., autonomous navigation). One important\napplication, the use of imagery for automated semantic understanding of\npedestrian environments, provides remote mapping of accessibility features in\nstreet environments. This application (and others like it) require detailed\ngeometric information of geographical objects. Semantic segmentation is a\nprerequisite for this task since it maps contiguous regions of the same class\nas single entities. Importantly, semantic segmentation uses like ours are not\npixel-wise outcomes; however, most of their quantitative evaluation metrics\n(e.g., mean Intersection Over Union) are based on pixel-wise similarities to a\nground-truth, which fails to emphasize over- and under-segmentation properties\nof a segmentation model. Here, we introduce a new metric to assess region-based\nover- and under-segmentation. We analyze and compare it to other metrics,\ndemonstrating that the use of our metric lends greater explainability to\nsemantic segmentation model performance in real-world applications.",
+                    "authors": "Yuxiang Zhang, Sachin Mehta, Anat Caspi",
+                    "published": "2021-01-21",
+                    "updated": "2023-02-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2304.02713v1",
+                    "title": "NUMSnet: Nested-U Multi-class Segmentation network for 3D Medical Image Stacks",
+                    "abstract": "Semantic segmentation for medical 3D image stacks enables accurate volumetric\nreconstructions, computer-aided diagnostics and follow up treatment planning.\nIn this work, we present a novel variant of the Unet model called the NUMSnet\nthat transmits pixel neighborhood features across scans through nested layers\nto achieve accurate multi-class semantic segmentations with minimal training\ndata. We analyze the semantic segmentation performance of the NUMSnet model in\ncomparison with several Unet model variants to segment 3-7 regions of interest\nusing only 10% of images for training per Lung-CT and Heart-CT volumetric image\nstacks. The proposed NUMSnet model achieves up to 20% improvement in\nsegmentation recall with 4-9% improvement in Dice scores for Lung-CT stacks and\n2.5-10% improvement in Dice scores for Heart-CT stacks when compared to the\nUnet++ model. The NUMSnet model needs to be trained by ordered images around\nthe central scan of each volumetric stack. Propagation of image feature\ninformation from the 6 nested layers of the Unet++ model are found to have\nbetter computation and segmentation performances than propagation of all\nup-sampling layers in a Unet++ model. The NUMSnet model achieves comparable\nsegmentation performances to existing works, while being trained on as low as\n5\\% of the training images. Also, transfer learning allows faster convergence\nof the NUMSnet model for multi-class semantic segmentation from pathology in\nLung-CT images to cardiac segmentations in Heart-CT stacks. Thus, the proposed\nmodel can standardize multi-class semantic segmentation on a variety of\nvolumetric image stacks with minimal training dataset. This can significantly\nreduce the cost, time and inter-observer variabilities associated with\ncomputer-aided detections and treatment.",
+                    "authors": "Sohini Roychowdhury",
+                    "published": "2023-04-05",
+                    "updated": "2023-04-05",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.00949v3",
+                    "title": "Synthetic Instance Segmentation from Semantic Image Segmentation Masks",
+                    "abstract": "In recent years, the development of instance segmentation has garnered\nsignificant attention in a wide range of applications. However, the training of\na fully-supervised instance segmentation model requires costly both\ninstance-level and pixel-level annotations. In contrast, weakly-supervised\ninstance segmentation methods (i.e., with image-level class labels or point\nlabels) struggle to satisfy the accuracy and recall requirements of practical\nscenarios. In this paper, we propose a novel paradigm called synthetic instance\nsegmentation (SISeg), which achieves Instance Segmentation results from image\nmasks predicted using off-the-shelf semantic segmentation models. SISeg does\nnot require training a semantic or/and instance segmentation model and avoids\nthe need for instance-level image annotations. Therefore, it is highly\nefficient. Specifically, we first obtain a semantic segmentation mask of the\ninput image via a trained semantic segmentation model. Then, we calculate a\ndisplacement field vector for each pixel based on the segmentation mask, which\ncan indicate representations belonging to the same class but different\ninstances, i.e., obtaining the instance-level object information. Finally,\ninstance segmentation results are obtained after being refined by a learnable\ncategory-agnostic object boundary branch. Extensive experimental results on two\nchallenging datasets and representative semantic segmentation baselines\n(including CNNs and Transformers) demonstrate that SISeg can achieve\ncompetitive results compared to the state-of-the-art fully-supervised instance\nsegmentation methods without the need for additional human resources or\nincreased computational costs. The code is available at: SISeg",
+                    "authors": "Yuchen Shen, Dong Zhang, Yuhui Zheng, Zechao Li, Liyong Fu, Qiaolin Ye",
+                    "published": "2023-08-02",
+                    "updated": "2023-10-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1705.09052v3",
+                    "title": "Weakly Supervised Semantic Segmentation Based on Web Image Co-segmentation",
+                    "abstract": "Training a Fully Convolutional Network (FCN) for semantic segmentation\nrequires a large number of masks with pixel level labelling, which involves a\nlarge amount of human labour and time for annotation. In contrast, web images\nand their image-level labels are much easier and cheaper to obtain. In this\nwork, we propose a novel method for weakly supervised semantic segmentation\nwith only image-level labels. The method utilizes the internet to retrieve a\nlarge number of images and uses a large scale co-segmentation framework to\ngenerate masks for the retrieved images. We first retrieve images from search\nengines, e.g. Flickr and Google, using semantic class names as queries, e.g.\nclass names in the dataset PASCAL VOC 2012. We then use high quality masks\nproduced by co-segmentation on the retrieved images as well as the target\ndataset images with image level labels to train segmentation networks. We\nobtain an IoU score of 56.9 on test set of PASCAL VOC 2012, which reaches the\nstate-of-the-art performance.",
+                    "authors": "Tong Shen, Guosheng Lin, Lingqiao Liu, Chunhua Shen, Ian Reid",
+                    "published": "2017-05-25",
+                    "updated": "2017-08-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.05490v1",
+                    "title": "Learning Semantic Segmentation with Query Points Supervision on Aerial Images",
+                    "abstract": "Semantic segmentation is crucial in remote sensing, where high-resolution\nsatellite images are segmented into meaningful regions. Recent advancements in\ndeep learning have significantly improved satellite image segmentation.\nHowever, most of these methods are typically trained in fully supervised\nsettings that require high-quality pixel-level annotations, which are expensive\nand time-consuming to obtain. In this work, we present a weakly supervised\nlearning algorithm to train semantic segmentation algorithms that only rely on\nquery point annotations instead of full mask labels. Our proposed approach\nperforms accurate semantic segmentation and improves efficiency by\nsignificantly reducing the cost and time required for manual annotation.\nSpecifically, we generate superpixels and extend the query point labels into\nthose superpixels that group similar meaningful semantics. Then, we train\nsemantic segmentation models, supervised with images partially labeled with the\nsuperpixels pseudo-labels. We benchmark our weakly supervised training approach\non an aerial image dataset and different semantic segmentation architectures,\nshowing that we can reach competitive performance compared to fully supervised\ntraining while reducing the annotation effort.",
+                    "authors": "Santiago Rivier, Carlos Hinojosa, Silvio Giancola, Bernard Ghanem",
+                    "published": "2023-09-11",
+                    "updated": "2023-09-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.11857v1",
+                    "title": "Joint Learning of Intrinsic Images and Semantic Segmentation",
+                    "abstract": "Semantic segmentation of outdoor scenes is problematic when there are\nvariations in imaging conditions. It is known that albedo (reflectance) is\ninvariant to all kinds of illumination effects. Thus, using reflectance images\nfor semantic segmentation task can be favorable. Additionally, not only\nsegmentation may benefit from reflectance, but also segmentation may be useful\nfor reflectance computation. Therefore, in this paper, the tasks of semantic\nsegmentation and intrinsic image decomposition are considered as a combined\nprocess by exploring their mutual relationship in a joint fashion. To that end,\nwe propose a supervised end-to-end CNN architecture to jointly learn intrinsic\nimage decomposition and semantic segmentation. We analyze the gains of\naddressing those two problems jointly. Moreover, new cascade CNN architectures\nfor intrinsic-for-segmentation and segmentation-for-intrinsic are proposed as\nsingle tasks. Furthermore, a dataset of 35K synthetic images of natural\nenvironments is created with corresponding albedo and shading (intrinsics), as\nwell as semantic labels (segmentation) assigned to each object/scene. The\nexperiments show that joint learning of intrinsic image decomposition and\nsemantic segmentation is beneficial for both tasks for natural scenes. Dataset\nand models are available at: https://ivi.fnwi.uva.nl/cv/intrinseg",
+                    "authors": "Anil S. Baslamisli, Thomas T. Groenestege, Partha Das, Hoang-An Le, Sezer Karaoglu, Theo Gevers",
+                    "published": "2018-07-31",
+                    "updated": "2018-07-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2201.05869v2",
+                    "title": "Prototype Guided Network for Anomaly Segmentation",
+                    "abstract": "Semantic segmentation methods can not directly identify abnormal objects in\nimages. Anomaly Segmentation algorithm from this realistic setting can\ndistinguish between in-distribution objects and Out-Of-Distribution (OOD)\nobjects and output the anomaly probability for pixels. In this paper, a\nPrototype Guided Anomaly segmentation Network (PGAN) is proposed to extract\nsemantic prototypes for in-distribution training data from limited annotated\nimages. In the model, prototypes are used to model the hierarchical category\nsemantic information and distinguish OOD pixels. The proposed PGAN model\nincludes a semantic segmentation network and a prototype extraction network.\nSimilarity measures are adopted to optimize the prototypes. The learned\nsemantic prototypes are used as category semantics to compare the similarity\nwith features extracted from test images and then to generate semantic\nsegmentation prediction. The proposed prototype extraction network can also be\nintegrated into most semantic segmentation networks and recognize OOD pixels.\nOn the StreetHazards dataset, the proposed PGAN model produced mIoU of 53.4%\nfor anomaly segmentation. The experimental results demonstrate PGAN may achieve\nthe SOTA performance in the anomaly segmentation tasks.",
+                    "authors": "Yiqing Hao, Yi Jin, Gaoyun An",
+                    "published": "2022-01-15",
+                    "updated": "2022-03-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2103.00286v1",
+                    "title": "A Novel Adaptive Deep Network for Building Footprint Segmentation",
+                    "abstract": "Building footprint segmentations for high resolution images are increasingly\ndemanded for many remote sensing applications. By the emerging deep learning\napproaches, segmentation networks have made significant advances in the\nsemantic segmentation of objects. However, these advances and the increased\naccess to satellite images require the generation of accurate object boundaries\nin satellite images. In the current paper, we propose a novel network-based on\nPix2Pix methodology to solve the problem of inaccurate boundaries obtained by\nconverting satellite images into maps using segmentation networks in order to\nsegment building footprints. To define the new network named G2G, our framework\nincludes two generators where the first generator extracts localization\nfeatures in order to merge them with the boundary features extracted from the\nsecond generator to segment all detailed building edges. Moreover, different\nstrategies are implemented to enhance the quality of the proposed networks'\nresults, implying that the proposed network outperforms state-of-the-art\nnetworks in segmentation accuracy with a large margin for all evaluation\nmetrics. The implementation is available at\nhttps://github.com/A2Amir/A-Novel-Adaptive-Deep-Network-for-Building-Footprint-Segmentation.",
+                    "authors": "A. Ziaee, R. Dehbozorgi, M. D\u00f6ller",
+                    "published": "2021-02-27",
+                    "updated": "2021-02-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG",
+                        "cs.NA",
+                        "eess.IV",
+                        "math.NA"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2106.04108v3",
+                    "title": "Fully Transformer Networks for Semantic Image Segmentation",
+                    "abstract": "Transformers have shown impressive performance in various natural language\nprocessing and computer vision tasks, due to the capability of modeling\nlong-range dependencies. Recent progress has demonstrated that combining such\nTransformers with CNN-based semantic image segmentation models is very\npromising. However, it is not well studied yet on how well a pure Transformer\nbased approach can achieve for image segmentation. In this work, we explore a\nnovel framework for semantic image segmentation, which is encoder-decoder based\nFully Transformer Networks (FTN). Specifically, we first propose a Pyramid\nGroup Transformer (PGT) as the encoder for progressively learning hierarchical\nfeatures, meanwhile reducing the computation complexity of the standard Visual\nTransformer (ViT). Then, we propose a Feature Pyramid Transformer (FPT) to fuse\nsemantic-level and spatial-level information from multiple levels of the PGT\nencoder for semantic image segmentation. Surprisingly, this simple baseline can\nachieve better results on multiple challenging semantic segmentation and face\nparsing benchmarks, including PASCAL Context, ADE20K, COCOStuff, and\nCelebAMask-HQ. The source code will be released on\nhttps://github.com/BR-IDL/PaddleViT.",
+                    "authors": "Sitong Wu, Tianyi Wu, Fangjian Lin, Shengwei Tian, Guodong Guo",
+                    "published": "2021-06-08",
+                    "updated": "2021-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2011.00674v1",
+                    "title": "Highway Driving Dataset for Semantic Video Segmentation",
+                    "abstract": "Scene understanding is an essential technique in semantic segmentation.\nAlthough there exist several datasets that can be used for semantic\nsegmentation, they are mainly focused on semantic image segmentation with large\ndeep neural networks. Therefore, these networks are not useful for real time\napplications, especially in autonomous driving systems. In order to solve this\nproblem, we make two contributions to semantic segmentation task. The first\ncontribution is that we introduce the semantic video dataset, the Highway\nDriving dataset, which is a densely annotated benchmark for a semantic video\nsegmentation task. The Highway Driving dataset consists of 20 video sequences\nhaving a 30Hz frame rate, and every frame is densely annotated. Secondly, we\npropose a baseline algorithm that utilizes a temporal correlation. Together\nwith our attempt to analyze the temporal correlation, we expect the Highway\nDriving dataset to encourage research on semantic video segmentation.",
+                    "authors": "Byungju Kim, Junho Yim, Junmo Kim",
+                    "published": "2020-11-02",
+                    "updated": "2020-11-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2401.11739v1",
+                    "title": "EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models",
+                    "abstract": "Diffusion models have recently received increasing research attention for\ntheir remarkable transfer abilities in semantic segmentation tasks. However,\ngenerating fine-grained segmentation masks with diffusion models often requires\nadditional training on annotated datasets, leaving it unclear to what extent\npre-trained diffusion models alone understand the semantic relations of their\ngenerated images. To address this question, we leverage the semantic knowledge\nextracted from Stable Diffusion (SD) and aim to develop an image segmentor\ncapable of generating fine-grained segmentation maps without any additional\ntraining. The primary difficulty stems from the fact that semantically\nmeaningful feature maps typically exist only in the spatially lower-dimensional\nlayers, which poses a challenge in directly extracting pixel-level semantic\nrelations from these feature maps. To overcome this issue, our framework\nidentifies semantic correspondences between image pixels and spatial locations\nof low-dimensional feature maps by exploiting SD's generation process and\nutilizes them for constructing image-resolution segmentation maps. In extensive\nexperiments, the produced segmentation maps are demonstrated to be well\ndelineated and capture detailed parts of the images, indicating the existence\nof highly accurate pixel-level semantic knowledge in diffusion models.",
+                    "authors": "Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, Seung Wook Kim",
+                    "published": "2024-01-22",
+                    "updated": "2024-01-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.17091v1",
+                    "title": "SSSegmenation: An Open Source Supervised Semantic Segmentation Toolbox Based on PyTorch",
+                    "abstract": "This paper presents SSSegmenation, which is an open source supervised\nsemantic image segmentation toolbox based on PyTorch. The design of this\ntoolbox is motivated by MMSegmentation while it is easier to use because of\nfewer dependencies and achieves superior segmentation performance under a\ncomparable training and testing setup. Moreover, the toolbox also provides\nplenty of trained weights for popular and contemporary semantic segmentation\nmethods, including Deeplab, PSPNet, OCRNet, MaskFormer, \\emph{etc}. We expect\nthat this toolbox can contribute to the future development of semantic\nsegmentation. Codes and model zoos are available at\n\\href{https://github.com/SegmentationBLWX/sssegmentation/}{SSSegmenation}.",
+                    "authors": "Zhenchao Jin",
+                    "published": "2023-05-26",
+                    "updated": "2023-05-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.02094v1",
+                    "title": "Segment Anything Meets Semantic Communication",
+                    "abstract": "In light of the diminishing returns of traditional methods for enhancing\ntransmission rates, the domain of semantic communication presents promising new\nfrontiers. Focusing on image transmission, this paper explores the application\nof foundation models, particularly the Segment Anything Model (SAM) developed\nby Meta AI Research, to improve semantic communication. SAM is a promptable\nimage segmentation model that has gained attention for its ability to perform\nzero-shot segmentation tasks without explicit training or domain-specific\nknowledge. By employing SAM's segmentation capability and lightweight neural\nnetwork architecture for semantic coding, we propose a practical approach to\nsemantic communication. We demonstrate that this approach retains critical\nsemantic features, achieving higher image reconstruction quality and reducing\ncommunication overhead. This practical solution eliminates the\nresource-intensive stage of training a segmentation model and can be applied to\nany semantic coding architecture, paving the way for real-world applications.",
+                    "authors": "Shehbaz Tariq, Brian Estadimas Arfeto, Chaoning Zhang, Hyundong Shin",
+                    "published": "2023-06-03",
+                    "updated": "2023-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.12903v1",
+                    "title": "Land Cover Change Detection via Semantic Segmentation",
+                    "abstract": "This paper presents a change detection method that identifies land cover\nchanges from aerial imagery, using semantic segmentation, a machine learning\napproach. We present a land cover classification training pipeline with Deeplab\nv3+, state-of-the-art semantic segmentation technology, including data\npreparation, model training for seven land cover types, and model exporting\nmodules. In the land cover change detection system, the inputs are images\nretrieved from Google Earth at the same location but from different times. The\nsystem then predicts semantic segmentation results on these images using the\ntrained model and calculates the land cover class percentage for each input\nimage. We see an improvement in the accuracy of the land cover semantic\nsegmentation model, with a mean IoU of 0.756 compared to 0.433, as reported in\nthe DeepGlobe land cover classification challenge. The land cover change\ndetection system that leverages the state-of-the-art semantic segmentation\ntechnology is proposed and can be used for deforestation analysis, land\nmanagement, and urban planning.",
+                    "authors": "Renee Su, Rong Chen",
+                    "published": "2019-11-28",
+                    "updated": "2019-11-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.05840v2",
+                    "title": "Self-Correlation and Cross-Correlation Learning for Few-Shot Remote Sensing Image Semantic Segmentation",
+                    "abstract": "Remote sensing image semantic segmentation is an important problem for remote\nsensing image interpretation. Although remarkable progress has been achieved,\nexisting deep neural network methods suffer from the reliance on massive\ntraining data. Few-shot remote sensing semantic segmentation aims at learning\nto segment target objects from a query image using only a few annotated support\nimages of the target class. Most existing few-shot learning methods stem\nprimarily from their sole focus on extracting information from support images,\nthereby failing to effectively address the large variance in appearance and\nscales of geographic objects. To tackle these challenges, we propose a\nSelf-Correlation and Cross-Correlation Learning Network for the few-shot remote\nsensing image semantic segmentation. Our model enhances the generalization by\nconsidering both self-correlation and cross-correlation between support and\nquery images to make segmentation predictions. To further explore the\nself-correlation with the query image, we propose to adopt a classical spectral\nmethod to produce a class-agnostic segmentation mask based on the basic visual\ninformation of the image. Extensive experiments on two remote sensing image\ndatasets demonstrate the effectiveness and superiority of our model in few-shot\nremote sensing image semantic segmentation. Code and models will be accessed at\nhttps://github.com/linhanwang/SCCNet.",
+                    "authors": "Linhan Wang, Shuo Lei, Jianfeng He, Shengkun Wang, Min Zhang, Chang-Tien Lu",
+                    "published": "2023-09-11",
+                    "updated": "2023-09-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2101.09642v2",
+                    "title": "Image Compression with Encoder-Decoder Matched Semantic Segmentation",
+                    "abstract": "In recent years, layered image compression is demonstrated to be a promising\ndirection, which encodes a compact representation of the input image and apply\nan up-sampling network to reconstruct the image. To further improve the quality\nof the reconstructed image, some works transmit the semantic segment together\nwith the compressed image data. Consequently, the compression ratio is also\ndecreased because extra bits are required for transmitting the semantic\nsegment. To solve this problem, we propose a new layered image compression\nframework with encoder-decoder matched semantic segmentation (EDMS). And then,\nfollowed by the semantic segmentation, a special convolution neural network is\nused to enhance the inaccurate semantic segment. As a result, the accurate\nsemantic segment can be obtained in the decoder without requiring extra bits.\nThe experimental results show that the proposed EDMS framework can get up to\n35.31% BD-rate reduction over the HEVC-based (BPG) codec, 5% bitrate, and 24%\nencoding time saving compare to the state-of-the-art semantic-based image\ncodec.",
+                    "authors": "Trinh Man Hoang, Jinjia Zhou, Yibo Fan",
+                    "published": "2021-01-24",
+                    "updated": "2021-01-30",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV",
+                        "cs.MM"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.11316v2",
+                    "title": "Generative Semantic Segmentation",
+                    "abstract": "We present Generative Semantic Segmentation (GSS), a generative learning\napproach for semantic segmentation. Uniquely, we cast semantic segmentation as\nan image-conditioned mask generation problem. This is achieved by replacing the\nconventional per-pixel discriminative learning with a latent prior learning\nprocess. Specifically, we model the variational posterior distribution of\nlatent variables given the segmentation mask. To that end, the segmentation\nmask is expressed with a special type of image (dubbed as maskige). This\nposterior distribution allows to generate segmentation masks unconditionally.\nTo achieve semantic segmentation on a given image, we further introduce a\nconditioning network. It is optimized by minimizing the divergence between the\nposterior distribution of maskige (i.e., segmentation masks) and the latent\nprior distribution of input training images. Extensive experiments on standard\nbenchmarks show that our GSS can perform competitively to prior art\nalternatives in the standard semantic segmentation setting, whilst achieving a\nnew state of the art in the more challenging cross-domain setting.",
+                    "authors": "Jiaqi Chen, Jiachen Lu, Xiatian Zhu, Li Zhang",
+                    "published": "2023-03-20",
+                    "updated": "2023-08-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.04829v2",
+                    "title": "MixReorg: Cross-Modal Mixed Patch Reorganization is a Good Mask Learner for Open-World Semantic Segmentation",
+                    "abstract": "Recently, semantic segmentation models trained with image-level text\nsupervision have shown promising results in challenging open-world scenarios.\nHowever, these models still face difficulties in learning fine-grained semantic\nalignment at the pixel level and predicting accurate object masks. To address\nthis issue, we propose MixReorg, a novel and straightforward pre-training\nparadigm for semantic segmentation that enhances a model's ability to\nreorganize patches mixed across images, exploring both local visual relevance\nand global semantic coherence. Our approach involves generating fine-grained\npatch-text pairs data by mixing image patches while preserving the\ncorrespondence between patches and text. The model is then trained to minimize\nthe segmentation loss of the mixed images and the two contrastive losses of the\noriginal and restored features. With MixReorg as a mask learner, conventional\ntext-supervised semantic segmentation models can achieve highly generalizable\npixel-semantic alignment ability, which is crucial for open-world segmentation.\nAfter training with large-scale image-text data, MixReorg models can be applied\ndirectly to segment visual objects of arbitrary categories, without the need\nfor further fine-tuning. Our proposed framework demonstrates strong performance\non popular zero-shot semantic segmentation benchmarks, outperforming GroupViT\nby significant margins of 5.0%, 6.2%, 2.5%, and 3.4% mIoU on PASCAL VOC2012,\nPASCAL Context, MS COCO, and ADE20K, respectively.",
+                    "authors": "Kaixin Cai, Pengzhen Ren, Yi Zhu, Hang Xu, Jianzhuang Liu, Changlin Li, Guangrun Wang, Xiaodan Liang",
+                    "published": "2023-08-09",
+                    "updated": "2024-03-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.04297v1",
+                    "title": "SHDM-NET: Heat Map Detail Guidance with Image Matting for Industrial Weld Semantic Segmentation Network",
+                    "abstract": "In actual industrial production, the assessment of the steel plate welding\neffect is an important task, and the segmentation of the weld section is the\nbasis of the assessment. This paper proposes an industrial weld segmentation\nnetwork based on a deep learning semantic segmentation algorithm fused with\nheatmap detail guidance and Image Matting to solve the automatic segmentation\nproblem of weld regions. In the existing semantic segmentation networks, the\nboundary information can be preserved by fusing the features of both high-level\nand low-level layers. However, this method can lead to insufficient expression\nof the spatial information in the low-level layer, resulting in inaccurate\nsegmentation boundary positioning. We propose a detailed guidance module based\non heatmaps to fully express the segmented region boundary information in the\nlow-level network to address this problem. Specifically, the expression of\nboundary information can be enhanced by adding a detailed branch to predict\nsegmented boundary and then matching it with the boundary heat map generated by\nmask labels to calculate the mean square error loss. In addition, although deep\nlearning has achieved great success in the field of semantic segmentation, the\nprecision of the segmentation boundary region is not high due to the loss of\ndetailed information caused by the classical segmentation network in the\nprocess of encoding and decoding process. This paper introduces a matting\nalgorithm to calibrate the boundary of the segmentation region of the semantic\nsegmentation network to solve this problem. Through many experiments on\nindustrial weld data sets, the effectiveness of our method is demonstrated, and\nthe MIOU reaches 97.93%. It is worth noting that this performance is comparable\nto human manual segmentation ( MIOU 97.96%).",
+                    "authors": "Qi Wang, Jingwu Mei",
+                    "published": "2022-07-09",
+                    "updated": "2022-07-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2304.11332v2",
+                    "title": "Input Augmentation with SAM: Boosting Medical Image Segmentation with Segmentation Foundation Model",
+                    "abstract": "The Segment Anything Model (SAM) is a recently developed large model for\ngeneral-purpose segmentation for computer vision tasks. SAM was trained using\n11 million images with over 1 billion masks and can produce segmentation\nresults for a wide range of objects in natural scene images. SAM can be viewed\nas a general perception model for segmentation (partitioning images into\nsemantically meaningful regions). Thus, how to utilize such a large foundation\nmodel for medical image segmentation is an emerging research target. This paper\nshows that although SAM does not immediately give high-quality segmentation for\nmedical image data, its generated masks, features, and stability scores are\nuseful for building and training better medical image segmentation models. In\nparticular, we demonstrate how to use SAM to augment image input for\ncommonly-used medical image segmentation models (e.g., U-Net). Experiments on\nthree segmentation tasks show the effectiveness of our proposed SAMAug method.\nThe code is available at \\url{https://github.com/yizhezhang2000/SAMAug}.",
+                    "authors": "Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Danny Z. Chen",
+                    "published": "2023-04-22",
+                    "updated": "2023-06-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.15608v1",
+                    "title": "Semantic Segmentation by Semantic Proportions",
+                    "abstract": "Semantic segmentation is a critical task in computer vision that aims to\nidentify and classify individual pixels in an image, with numerous applications\nfor example autonomous driving and medical image analysis. However, semantic\nsegmentation can be super challenging particularly due to the need for large\namounts of annotated data. Annotating images is a time-consuming and costly\nprocess, often requiring expert knowledge and significant effort. In this\npaper, we propose a novel approach for semantic segmentation by eliminating the\nneed of ground-truth segmentation maps. Instead, our approach requires only the\nrough information of individual semantic class proportions, shortened as\nsemantic proportions. It greatly simplifies the data annotation process and\nthus will significantly reduce the annotation time and cost, making it more\nfeasible for large-scale applications. Moreover, it opens up new possibilities\nfor semantic segmentation tasks where obtaining the full ground-truth\nsegmentation maps may not be feasible or practical. Extensive experimental\nresults demonstrate that our approach can achieve comparable and sometimes even\nbetter performance against the benchmark method that relies on the ground-truth\nsegmentation maps. Utilising semantic proportions suggested in this work offers\na promising direction for future research in the field of semantic\nsegmentation.",
+                    "authors": "Halil Ibrahim Aysel, Xiaohao Cai, Adam Pr\u00fcgel-Bennett",
+                    "published": "2023-05-24",
+                    "updated": "2023-05-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.13978v3",
+                    "title": "Personalized Image Semantic Segmentation",
+                    "abstract": "Semantic segmentation models trained on public datasets have achieved great\nsuccess in recent years. However, these models didn't consider the\npersonalization issue of segmentation though it is important in practice. In\nthis paper, we address the problem of personalized image segmentation. The\nobjective is to generate more accurate segmentation results on unlabeled\npersonalized images by investigating the data's personalized traits. To open up\nfuture research in this area, we collect a large dataset containing various\nusers' personalized images called PIS (Personalized Image Semantic\nSegmentation). We also survey some recent researches related to this problem\nand report their performance on our dataset. Furthermore, by observing the\ncorrelation among a user's personalized images, we propose a baseline method\nthat incorporates the inter-image context when segmenting certain images.\nExtensive experiments show that our method outperforms the existing methods on\nthe proposed dataset. The code and the PIS dataset will be made publicly\navailable.",
+                    "authors": "Yu Zhang, Chang-Bin Zhang, Peng-Tao Jiang, Ming-Ming Cheng, Feng Mao",
+                    "published": "2021-07-24",
+                    "updated": "2021-09-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.03983v1",
+                    "title": "Weakly Supervised Semantic Segmentation of Satellite Images",
+                    "abstract": "When one wants to train a neural network to perform semantic segmentation,\ncreating pixel-level annotations for each of the images in the database is a\ntedious task. If he works with aerial or satellite images, which are usually\nvery large, it is even worse. With that in mind, we investigate how to use\nimage-level annotations in order to perform semantic segmentation. Image-level\nannotations are much less expensive to acquire than pixel-level annotations,\nbut we lose a lot of information for the training of the model. From the\nannotations of the images, the model must find by itself how to classify the\ndifferent regions of the image. In this work, we use the method proposed by Anh\nand Kwak [1] to produce pixel-level annotation from image level annotation. We\ncompare the overall quality of our generated dataset with the original dataset.\nIn addition, we propose an adaptation of the AffinityNet that allows us to\ndirectly perform a semantic segmentation. Our results show that the generated\nlabels lead to the same performances for the training of several segmentation\nnetworks. Also, the quality of semantic segmentation performed directly by the\nAffinityNet and the Random Walk is close to the one of the best\nfully-supervised approaches.",
+                    "authors": "Adrien Nivaggioli, Hicham Randrianarivo",
+                    "published": "2019-04-08",
+                    "updated": "2019-04-08",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.05321v1",
+                    "title": "Image Segmentation Semantic Communication over Internet of Vehicles",
+                    "abstract": "In this paper, the problem of semantic-based efficient image transmission is\nstudied over the Internet of Vehicles (IoV). In the considered model, a vehicle\nshares massive amount of visual data perceived by its visual sensors to assist\nother vehicles in making driving decisions. However, it is hard to maintain a\nhigh reliable visual data transmission due to the limited spectrum resources.\nTo tackle this problem, a semantic communication approach is introduced to\nreduce the transmission data amount while ensuring the semantic-level accuracy.\nParticularly, an image segmentation semantic communication (ISSC) system is\nproposed, which can extract the semantic features from the perceived images and\ntransmit the features to the receiving vehicle that reconstructs the image\nsegmentations. The ISSC system consists of an encoder and a decoder at the\ntransmitter and the receiver, respectively. To accurately extract the image\nsemantic features, the ISSC system encoder employs a Swin Transformer based\nmulti-scale semantic feature extractor. Then, to resist the wireless noise and\nreconstruct the image segmentation, a semantic feature decoder and a\nreconstructor are designed at the receiver. Simulation results show that the\nproposed ISSC system can reconstruct the image segmentation accurately with a\nhigh compression ratio, and can achieve robust transmission performance against\nchannel noise, especially at the low signal-to-noise ratio (SNR). In terms of\nmean Intersection over Union (mIoU), the ISSC system can achieve an increase by\n75%, compared to the baselines using traditional coding method",
+                    "authors": "Qiang Pan, Haonan Tong, Jie Lv, Tao Luo, Zhilong Zhang, Changchuan Yin, Jianfeng Li",
+                    "published": "2022-10-11",
+                    "updated": "2022-10-11",
+                    "primary_cat": "cs.NI",
+                    "cats": [
+                        "cs.NI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1509.02441v1",
+                    "title": "Semantic Video Segmentation : Exploring Inference Efficiency",
+                    "abstract": "We explore the efficiency of the CRF inference beyond image level semantic\nsegmentation and perform joint inference in video frames. The key idea is to\ncombine best of two worlds: semantic co-labeling and more expressive models.\nOur formulation enables us to perform inference over ten thousand images within\nseconds and makes the system amenable to perform video semantic segmentation\nmost effectively. On CamVid dataset, with TextonBoost unaries, our proposed\nmethod achieves up to 8% improvement in accuracy over individual semantic image\nsegmentation without additional time overhead. The source code is available at\nhttps://github.com/subtri/video_inference",
+                    "authors": "Subarna Tripathi, Serge Belongie, Youngbae Hwang, Truong Nguyen",
+                    "published": "2015-09-04",
+                    "updated": "2015-09-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.02840v1",
+                    "title": "Attention-based fusion of semantic boundary and non-boundary information to improve semantic segmentation",
+                    "abstract": "This paper introduces a method for image semantic segmentation grounded on a\nnovel fusion scheme, which takes place inside a deep convolutional neural\nnetwork. The main goal of our proposal is to explore object boundary\ninformation to improve the overall segmentation performance. Unlike previous\nworks that combine boundary and segmentation features, or those that use\nboundary information to regularize semantic segmentation, we instead propose a\nnovel approach that embodies boundary information onto segmentation. For that,\nour semantic segmentation method uses two streams, which are combined through\nan attention gate, forming an end-to-end Y-model. To the best of our knowledge,\nours is the first work to show that boundary detection can improve semantic\nsegmentation when fused through a semantic fusion gate (attention model). We\nperformed an extensive evaluation of our method over public data sets. We found\ncompetitive results on all data sets after comparing our proposed model with\nother twelve state-of-the-art segmenters, considering the same training\nconditions. Our proposed model achieved the best mIoU on the CityScapes,\nCamVid, and Pascal Context data sets, and the second best on Mapillary Vistas.",
+                    "authors": "Jefferson Fontinele, Gabriel Lefundes, Luciano Oliveira",
+                    "published": "2021-08-05",
+                    "updated": "2021-08-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2001.00335v1",
+                    "title": "Graph-FCN for image semantic segmentation",
+                    "abstract": "Semantic segmentation with deep learning has achieved great progress in\nclassifying the pixels in the image. However, the local location information is\nusually ignored in the high-level feature extraction by the deep learning,\nwhich is important for image semantic segmentation. To avoid this problem, we\npropose a graph model initialized by a fully convolutional network (FCN) named\nGraph-FCN for image semantic segmentation. Firstly, the image grid data is\nextended to graph structure data by a convolutional network, which transforms\nthe semantic segmentation problem into a graph node classification problem.\nThen we apply graph convolutional network to solve this graph node\nclassification problem. As far as we know, it is the first time that we apply\nthe graph convolutional network in image semantic segmentation. Our method\nachieves competitive performance in mean intersection over union (mIOU) on the\nVOC dataset(about 1.34% improvement), compared to the original FCN model.",
+                    "authors": "Yi Lu, Yaran Chen, Dongbin Zhao, Jianxin Chen",
+                    "published": "2020-01-02",
+                    "updated": "2020-01-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.02270v1",
+                    "title": "Weak-shot Semantic Segmentation via Dual Similarity Transfer",
+                    "abstract": "Semantic segmentation is an important and prevalent task, but severely\nsuffers from the high cost of pixel-level annotations when extending to more\nclasses in wider applications. To this end, we focus on the problem named\nweak-shot semantic segmentation, where the novel classes are learnt from\ncheaper image-level labels with the support of base classes having\noff-the-shelf pixel-level labels. To tackle this problem, we propose SimFormer,\nwhich performs dual similarity transfer upon MaskFormer. Specifically,\nMaskFormer disentangles the semantic segmentation task into two sub-tasks:\nproposal classification and proposal segmentation for each proposal. Proposal\nsegmentation allows proposal-pixel similarity transfer from base classes to\nnovel classes, which enables the mask learning of novel classes. We also learn\npixel-pixel similarity from base classes and distill such class-agnostic\nsemantic similarity to the semantic masks of novel classes, which regularizes\nthe segmentation model with pixel-level semantic relationship across images. In\naddition, we propose a complementary loss to facilitate the learning of novel\nclasses. Comprehensive experiments on the challenging COCO-Stuff-10K and ADE20K\ndatasets demonstrate the effectiveness of our method. Codes are available at\nhttps://github.com/bcmi/SimFormer-Weak-Shot-Semantic-Segmentation.",
+                    "authors": "Junjie Chen, Li Niu, Siyuan Zhou, Jianlou Si, Chen Qian, Liqing Zhang",
+                    "published": "2022-10-05",
+                    "updated": "2022-10-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2007.02361v1",
+                    "title": "Self-supervised Depth Estimation to Regularise Semantic Segmentation in Knee Arthroscopy",
+                    "abstract": "Intra-operative automatic semantic segmentation of knee joint structures can\nassist surgeons during knee arthroscopy in terms of situational awareness.\nHowever, due to poor imaging conditions (e.g., low texture, overexposure,\netc.), automatic semantic segmentation is a challenging scenario, which\njustifies the scarce literature on this topic. In this paper, we propose a\nnovel self-supervised monocular depth estimation to regularise the training of\nthe semantic segmentation in knee arthroscopy. To further regularise the depth\nestimation, we propose the use of clean training images captured by the stereo\narthroscope of routine objects (presenting none of the poor imaging conditions\nand with rich texture information) to pre-train the model. We fine-tune such\nmodel to produce both the semantic segmentation and self-supervised monocular\ndepth using stereo arthroscopic images taken from inside the knee. Using a data\nset containing 3868 arthroscopic images captured during cadaveric knee\narthroscopy with semantic segmentation annotations, 2000 stereo image pairs of\ncadaveric knee arthroscopy, and 2150 stereo image pairs of routine objects, we\nshow that our semantic segmentation regularised by self-supervised depth\nestimation produces a more accurate segmentation than a state-of-the-art\nsemantic segmentation approach modeled exclusively with semantic segmentation\nannotation.",
+                    "authors": "Fengbei Liu, Yaqub Jonmohamadi, Gabriel Maicas, Ajay K. Pandey, Gustavo Carneiro",
+                    "published": "2020-07-05",
+                    "updated": "2020-07-05",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1803.10464v2",
+                    "title": "Learning Pixel-level Semantic Affinity with Image-level Supervision for Weakly Supervised Semantic Segmentation",
+                    "abstract": "The deficiency of segmentation labels is one of the main obstacles to\nsemantic segmentation in the wild. To alleviate this issue, we present a novel\nframework that generates segmentation labels of images given their image-level\nclass labels. In this weakly supervised setting, trained models have been known\nto segment local discriminative parts rather than the entire object area. Our\nsolution is to propagate such local responses to nearby areas which belong to\nthe same semantic entity. To this end, we propose a Deep Neural Network (DNN)\ncalled AffinityNet that predicts semantic affinity between a pair of adjacent\nimage coordinates. The semantic propagation is then realized by random walk\nwith the affinities predicted by AffinityNet. More importantly, the supervision\nemployed to train AffinityNet is given by the initial discriminative part\nsegmentation, which is incomplete as a segmentation annotation but sufficient\nfor learning semantic affinities within small image areas. Thus the entire\nframework relies only on image-level class labels and does not require any\nextra data or annotations. On the PASCAL VOC 2012 dataset, a DNN learned with\nsegmentation labels generated by our method outperforms previous models trained\nwith the same level of supervision, and is even as competitive as those relying\non stronger supervision.",
+                    "authors": "Jiwoon Ahn, Suha Kwak",
+                    "published": "2018-03-28",
+                    "updated": "2018-04-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2111.11103v1",
+                    "title": "Improving Semantic Image Segmentation via Label Fusion in Semantically Textured Meshes",
+                    "abstract": "Models for semantic segmentation require a large amount of hand-labeled\ntraining data which is costly and time-consuming to produce. For this purpose,\nwe present a label fusion framework that is capable of improving semantic pixel\nlabels of video sequences in an unsupervised manner. We make use of a 3D mesh\nrepresentation of the environment and fuse the predictions of different frames\ninto a consistent representation using semantic mesh textures. Rendering the\nsemantic mesh using the original intrinsic and extrinsic camera parameters\nyields a set of improved semantic segmentation images. Due to our optimized\nCUDA implementation, we are able to exploit the entire $c$-dimensional\nprobability distribution of annotations over $c$ classes in an\nuncertainty-aware manner. We evaluate our method on the Scannet dataset where\nwe improve annotations produced by the state-of-the-art segmentation network\nESANet from $52.05 \\%$ to $58.25 \\%$ pixel accuracy. We publish the source code\nof our framework online to foster future research in this area\n(\\url{https://github.com/fferflo/semantic-meshes}). To the best of our\nknowledge, this is the first publicly available label fusion framework for\nsemantic image segmentation based on meshes with semantic textures.",
+                    "authors": "Florian Fervers, Timo Breuer, Gregor Stachowiak, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens",
+                    "published": "2021-11-22",
+                    "updated": "2021-11-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2006.07601v1",
+                    "title": "NoPeopleAllowed: The Three-Step Approach to Weakly Supervised Semantic Segmentation",
+                    "abstract": "We propose a novel approach to weakly supervised semantic segmentation, which\nconsists of three consecutive steps. The first two steps extract high-quality\npseudo masks from image-level annotated data, which are then used to train a\nsegmentation model on the third step. The presented approach also addresses two\nproblems in the data: class imbalance and missing labels. Using only\nimage-level annotations as supervision, our method is capable of segmenting\nvarious classes and complex objects. It achieves 37.34 mean IoU on the test\nset, placing 3rd at the LID Challenge in the task of weakly supervised semantic\nsegmentation.",
+                    "authors": "Mariia Dobko, Ostap Viniavskyi, Oles Dobosevych",
+                    "published": "2020-06-13",
+                    "updated": "2020-06-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2202.04754v2",
+                    "title": "Wireless Transmission of Images With The Assistance of Multi-level Semantic Information",
+                    "abstract": "Semantic-oriented communication has been considered as a promising to boost\nthe bandwidth efficiency by only transmitting the semantics of the data. In\nthis paper, we propose a multi-level semantic aware communication system for\nwireless image transmission, named MLSC-image, which is based on the deep\nlearning techniques and trained in an end to end manner. In particular, the\nproposed model includes a multilevel semantic feature extractor, that extracts\nboth the highlevel semantic information, such as the text semantics and the\nsegmentation semantics, and the low-level semantic information, such as local\nspatial details of the images. We employ a pretrained image caption to capture\nthe text semantics and a pretrained image segmentation model to obtain the\nsegmentation semantics. These high-level and low-level semantic features are\nthen combined and encoded by a joint semantic and channel encoder into symbols\nto transmit over the physical channel. The numerical results validate the\neffectiveness and efficiency of the proposed semantic communication system,\nespecially under the limited bandwidth condition, which indicates the\nadvantages of the high-level semantics in the compression of images.",
+                    "authors": "Zhenguo Zhang, Qianqian Yang, Shibo He, Mingyang Sun, Jiming Chen",
+                    "published": "2022-02-08",
+                    "updated": "2023-12-08",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2112.03241v1",
+                    "title": "Unsupervised Domain Adaptation for Semantic Image Segmentation: a Comprehensive Survey",
+                    "abstract": "Semantic segmentation plays a fundamental role in a broad variety of computer\nvision applications, providing key information for the global understanding of\nan image. Yet, the state-of-the-art models rely on large amount of annotated\nsamples, which are more expensive to obtain than in tasks such as image\nclassification. Since unlabelled data is instead significantly cheaper to\nobtain, it is not surprising that Unsupervised Domain Adaptation reached a\nbroad success within the semantic segmentation community.\n  This survey is an effort to summarize five years of this incredibly rapidly\ngrowing field, which embraces the importance of semantic segmentation itself\nand a critical need of adapting segmentation models to new environments. We\npresent the most important semantic segmentation methods; we provide a\ncomprehensive survey on domain adaptation techniques for semantic segmentation;\nwe unveil newer trends such as multi-domain learning, domain generalization,\ntest-time adaptation or source-free domain adaptation; we conclude this survey\nby describing datasets and benchmarks most widely used in semantic segmentation\nresearch. We hope that this survey will provide researchers across academia and\nindustry with a comprehensive reference guide and will help them in fostering\nnew research directions in the field.",
+                    "authors": "Gabriela Csurka, Riccardo Volpi, Boris Chidlovskii",
+                    "published": "2021-12-06",
+                    "updated": "2021-12-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "I.4.6; I.2"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.17083v1",
+                    "title": "SegRCDB: Semantic Segmentation via Formula-Driven Supervised Learning",
+                    "abstract": "Pre-training is a strong strategy for enhancing visual models to efficiently\ntrain them with a limited number of labeled images. In semantic segmentation,\ncreating annotation masks requires an intensive amount of labor and time, and\ntherefore, a large-scale pre-training dataset with semantic labels is quite\ndifficult to construct. Moreover, what matters in semantic segmentation\npre-training has not been fully investigated. In this paper, we propose the\nSegmentation Radial Contour DataBase (SegRCDB), which for the first time\napplies formula-driven supervised learning for semantic segmentation. SegRCDB\nenables pre-training for semantic segmentation without real images or any\nmanual semantic labels. SegRCDB is based on insights about what is important in\npre-training for semantic segmentation and allows efficient pre-training.\nPre-training with SegRCDB achieved higher mIoU than the pre-training with\nCOCO-Stuff for fine-tuning on ADE-20k and Cityscapes with the same number of\ntraining images. SegRCDB has a high potential to contribute to semantic\nsegmentation pre-training and investigation by enabling the creation of large\ndatasets without manual annotation. The SegRCDB dataset will be released under\na license that allows research and commercial use. Code is available at:\nhttps://github.com/dahlian00/SegRCDB",
+                    "authors": "Risa Shinoda, Ryo Hayamizu, Kodai Nakashima, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka",
+                    "published": "2023-09-29",
+                    "updated": "2023-09-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2401.07589v1",
+                    "title": "Semantic Scene Segmentation for Robotics",
+                    "abstract": "Comprehensive scene understanding is a critical enabler of robot autonomy.\nSemantic segmentation is one of the key scene understanding tasks which is\npivotal for several robotics applications including autonomous driving,\ndomestic service robotics, last mile delivery, amongst many others. Semantic\nsegmentation is a dense prediction task that aims to provide a scene\nrepresentation in which each pixel of an image is assigned a semantic class\nlabel. Therefore, semantic segmentation considers the full scene context,\nincorporating the object category, location, and shape of all the scene\nelements, including the background. Numerous algorithms have been proposed for\nsemantic segmentation over the years. However, the recent advances in deep\nlearning combined with the boost in the computational capacity and the\navailability of large-scale labeled datasets have led to significant advances\nin semantic segmentation. In this chapter, we introduce the task of semantic\nsegmentation and present the deep learning techniques that have been proposed\nto address this task over the years. We first define the task of semantic\nsegmentation and contrast it with other closely related scene understanding\nproblems. We detail different algorithms and architectures for semantic\nsegmentation and the commonly employed loss functions. Furthermore, we present\nan overview of datasets, benchmarks, and metrics that are used in semantic\nsegmentation. We conclude the chapter with a discussion of challenges and\nopportunities for further research in this area.",
+                    "authors": "Juana Valeria Hurtado, Abhinav Valada",
+                    "published": "2024-01-15",
+                    "updated": "2024-01-15",
+                    "primary_cat": "cs.RO",
+                    "cats": [
+                        "cs.RO"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1709.01956v1",
+                    "title": "Learning Dilation Factors for Semantic Segmentation of Street Scenes",
+                    "abstract": "Contextual information is crucial for semantic segmentation. However, finding\nthe optimal trade-off between keeping desired fine details and at the same time\nproviding sufficiently large receptive fields is non trivial. This is even more\nso, when objects or classes present in an image significantly vary in size.\nDilated convolutions have proven valuable for semantic segmentation, because\nthey allow to increase the size of the receptive field without sacrificing\nimage resolution. However, in current state-of-the-art methods, dilation\nparameters are hand-tuned and fixed. In this paper, we present an approach for\nlearning dilation parameters adaptively per channel, consistently improving\nsemantic segmentation results on street-scene datasets like Cityscapes and\nCamvid.",
+                    "authors": "Yang He, Margret Keuper, Bernt Schiele, Mario Fritz",
+                    "published": "2017-09-06",
+                    "updated": "2017-09-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.02525v1",
+                    "title": "Semantic Segmentation Alternative Technique: Segmentation Domain Generation",
+                    "abstract": "Detecting objects of interest in images was always a compelling task to\nautomate. In recent years this task was more and more explored using deep\nlearning techniques, mostly using region-based convolutional networks. In this\nproject we propose an alternative semantic segmentation technique making use of\nGenerative Adversarial Networks. We consider semantic segmentation to be a\ndomain transfer problem. Thus, we train a feed forward network (FFNN) to\nreceive as input a seed real image and generate as output its segmentation\nmask.",
+                    "authors": "Ana-Cristina Rogoz, Radu Muntean, Stefan Cobeli",
+                    "published": "2021-07-06",
+                    "updated": "2021-07-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.00679v3",
+                    "title": "Cooperative Semantic Segmentation and Image Restoration in Adverse Environmental Conditions",
+                    "abstract": "Most state-of-the-art semantic segmentation approaches only achieve high\naccuracy in good conditions. In practically-common but less-discussed adverse\nenvironmental conditions, their performance can decrease enormously. Existing\nstudies usually cast the handling of segmentation in adverse conditions as a\nseparate post-processing step after signal restoration, making the segmentation\nperformance largely depend on the quality of restoration. In this paper, we\npropose a novel deep-learning framework to tackle semantic segmentation and\nimage restoration in adverse environmental conditions in a holistic manner. The\nproposed approach contains two components: Semantically-Guided Adaptation,\nwhich exploits semantic information from degraded images to refine the\nsegmentation; and Exemplar-Guided Synthesis, which restores images from\nsemantic label maps given degraded exemplars as the guidance. Our method\ncooperatively leverages the complementarity and interdependence of low-level\nrestoration and high-level segmentation in adverse environmental conditions.\nExtensive experiments on various datasets demonstrate that our approach can not\nonly improve the accuracy of semantic segmentation with degradation cues, but\nalso boost the perceptual quality and structural similarity of image\nrestoration with semantic guidance.",
+                    "authors": "Weihao Xia, Zhanglin Cheng, Yujiu Yang, Jing-Hao Xue",
+                    "published": "2019-11-02",
+                    "updated": "2020-03-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2205.03186v1",
+                    "title": "Semantics-Guided Moving Object Segmentation with 3D LiDAR",
+                    "abstract": "Moving object segmentation (MOS) is a task to distinguish moving objects,\ne.g., moving vehicles and pedestrians, from the surrounding static environment.\nThe segmentation accuracy of MOS can have an influence on odometry, map\nconstruction, and planning tasks. In this paper, we propose a semantics-guided\nconvolutional neural network for moving object segmentation. The network takes\nsequential LiDAR range images as inputs. Instead of segmenting the moving\nobjects directly, the network conducts single-scan-based semantic segmentation\nand multiple-scan-based moving object segmentation in turn. The semantic\nsegmentation module provides semantic priors for the MOS module, where we\npropose an adjacent scan association (ASA) module to convert the semantic\nfeatures of adjacent scans into the same coordinate system to fully exploit the\ncross-scan semantic features. Finally, by analyzing the difference between the\ntransformed features, reliable MOS result can be obtained quickly. Experimental\nresults on the SemanticKITTI MOS dataset proves the effectiveness of our work.",
+                    "authors": "Shuo Gu, Suling Yao, Jian Yang, Hui Kong",
+                    "published": "2022-05-06",
+                    "updated": "2022-05-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.RO"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.12518v2",
+                    "title": "Segmentation in Style: Unsupervised Semantic Image Segmentation with Stylegan and CLIP",
+                    "abstract": "We introduce a method that allows to automatically segment images into\nsemantically meaningful regions without human supervision. Derived regions are\nconsistent across different images and coincide with human-defined semantic\nclasses on some datasets. In cases where semantic regions might be hard for\nhuman to define and consistently label, our method is still able to find\nmeaningful and consistent semantic classes. In our work, we use pretrained\nStyleGAN2 generative model: clustering in the feature space of the generative\nmodel allows to discover semantic classes. Once classes are discovered, a\nsynthetic dataset with generated images and corresponding segmentation masks\ncan be created. After that a segmentation model is trained on the synthetic\ndataset and is able to generalize to real images. Additionally, by using CLIP\nwe are able to use prompts defined in a natural language to discover some\ndesired semantic classes. We test our method on publicly available datasets and\nshow state-of-the-art results.",
+                    "authors": "Daniil Pakhomov, Sanchit Hira, Narayani Wagle, Kemar E. Green, Nassir Navab",
+                    "published": "2021-07-26",
+                    "updated": "2021-11-18",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1707.02432v2",
+                    "title": "Deep Semantic Segmentation for Automated Driving: Taxonomy, Roadmap and Challenges",
+                    "abstract": "Semantic segmentation was seen as a challenging computer vision problem few\nyears ago. Due to recent advancements in deep learning, relatively accurate\nsolutions are now possible for its use in automated driving. In this paper, the\nsemantic segmentation problem is explored from the perspective of automated\ndriving. Most of the current semantic segmentation algorithms are designed for\ngeneric images and do not incorporate prior structure and end goal for\nautomated driving. First, the paper begins with a generic taxonomic survey of\nsemantic segmentation algorithms and then discusses how it fits in the context\nof automated driving. Second, the particular challenges of deploying it into a\nsafety system which needs high level of accuracy and robustness are listed.\nThird, different alternatives instead of using an independent semantic\nsegmentation module are explored. Finally, an empirical evaluation of various\nsemantic segmentation architectures was performed on CamVid dataset in terms of\naccuracy and speed. This paper is a preliminary shorter version of a more\ndetailed survey which is work in progress.",
+                    "authors": "Mennatullah Siam, Sara Elkerdawy, Martin Jagersand, Senthil Yogamani",
+                    "published": "2017-07-08",
+                    "updated": "2017-08-03",
+                    "primary_cat": "stat.ML",
+                    "cats": [
+                        "stat.ML",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.10834v1",
+                    "title": "DANNet: A One-Stage Domain Adaptation Network for Unsupervised Nighttime Semantic Segmentation",
+                    "abstract": "Semantic segmentation of nighttime images plays an equally important role as\nthat of daytime images in autonomous driving, but the former is much more\nchallenging due to poor illuminations and arduous human annotations. In this\npaper, we propose a novel domain adaptation network (DANNet) for nighttime\nsemantic segmentation without using labeled nighttime image data. It employs an\nadversarial training with a labeled daytime dataset and an unlabeled dataset\nthat contains coarsely aligned day-night image pairs. Specifically, for the\nunlabeled day-night image pairs, we use the pixel-level predictions of static\nobject categories on a daytime image as a pseudo supervision to segment its\ncounterpart nighttime image. We further design a re-weighting strategy to\nhandle the inaccuracy caused by misalignment between day-night image pairs and\nwrong predictions of daytime images, as well as boost the prediction accuracy\nof small objects. The proposed DANNet is the first one stage adaptation\nframework for nighttime semantic segmentation, which does not train additional\nday-night image transfer models as a separate pre-processing stage. Extensive\nexperiments on Dark Zurich and Nighttime Driving datasets show that our method\nachieves state-of-the-art performance for nighttime semantic segmentation.",
+                    "authors": "Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, Song Wang",
+                    "published": "2021-04-22",
+                    "updated": "2021-04-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2009.12232v4",
+                    "title": "From Pixel to Patch: Synthesize Context-aware Features for Zero-shot Semantic Segmentation",
+                    "abstract": "Zero-shot learning has been actively studied for image classification task to\nrelieve the burden of annotating image labels. Interestingly, semantic\nsegmentation task requires more labor-intensive pixel-wise annotation, but\nzero-shot semantic segmentation has only attracted limited research interest.\nThus, we focus on zero-shot semantic segmentation, which aims to segment unseen\nobjects with only category-level semantic representations provided for unseen\ncategories. In this paper, we propose a novel Context-aware feature Generation\nNetwork (CaGNet), which can synthesize context-aware pixel-wise visual features\nfor unseen categories based on category-level semantic representations and\npixel-wise contextual information. The synthesized features are used to\nfinetune the classifier to enable segmenting unseen objects. Furthermore, we\nextend pixel-wise feature generation and finetuning to patch-wise feature\ngeneration and finetuning, which additionally considers inter-pixel\nrelationship. Experimental results on Pascal-VOC, Pascal-Context, and\nCOCO-stuff show that our method significantly outperforms the existing\nzero-shot semantic segmentation methods. Code is available at\nhttps://github.com/bcmi/CaGNetv2-Zero-Shot-Semantic-Segmentation.",
+                    "authors": "Zhangxuan Gu, Siyuan Zhou, Li Niu, Zihan Zhao, Liqing Zhang",
+                    "published": "2020-09-25",
+                    "updated": "2022-01-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.03273v1",
+                    "title": "Semantic Segmentation using Vision Transformers: A survey",
+                    "abstract": "Semantic segmentation has a broad range of applications in a variety of\ndomains including land coverage analysis, autonomous driving, and medical image\nanalysis. Convolutional neural networks (CNN) and Vision Transformers (ViTs)\nprovide the architecture models for semantic segmentation. Even though ViTs\nhave proven success in image classification, they cannot be directly applied to\ndense prediction tasks such as image segmentation and object detection since\nViT is not a general purpose backbone due to its patch partitioning scheme. In\nthis survey, we discuss some of the different ViT architectures that can be\nused for semantic segmentation and how their evolution managed the above-stated\nchallenge. The rise of ViT and its performance with a high success rate\nmotivated the community to slowly replace the traditional convolutional neural\nnetworks in various computer vision tasks. This survey aims to review and\ncompare the performances of ViT architectures designed for semantic\nsegmentation using benchmarking datasets. This will be worthwhile for the\ncommunity to yield knowledge regarding the implementations carried out in\nsemantic segmentation and to discover more efficient methodologies using ViTs.",
+                    "authors": "Hans Thisanke, Chamli Deshan, Kavindu Chamith, Sachith Seneviratne, Rajith Vidanaarachchi, Damayanthi Herath",
+                    "published": "2023-05-05",
+                    "updated": "2023-05-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1605.06885v1",
+                    "title": "Bridging Category-level and Instance-level Semantic Image Segmentation",
+                    "abstract": "We propose an approach to instance-level image segmentation that is built on\ntop of category-level segmentation. Specifically, for each pixel in a semantic\ncategory mask, its corresponding instance bounding box is predicted using a\ndeep fully convolutional regression network. Thus it follows a different\npipeline to the popular detect-then-segment approaches that first predict\ninstances' bounding boxes, which are the current state-of-the-art in instance\nsegmentation. We show that, by leveraging the strength of our state-of-the-art\nsemantic segmentation models, the proposed method can achieve comparable or\neven better results to detect-then-segment approaches. We make the following\ncontributions. (i) First, we propose a simple yet effective approach to\nsemantic instance segmentation. (ii) Second, we propose an online bootstrapping\nmethod during training, which is critically important for achieving good\nperformance for both semantic category segmentation and instance-level\nsegmentation. (iii) As the performance of semantic category segmentation has a\nsignificant impact on the instance-level segmentation, which is the second step\nof our approach, we train fully convolutional residual networks to achieve the\nbest semantic category segmentation accuracy. On the PASCAL VOC 2012 dataset,\nwe obtain the currently best mean intersection-over-union score of 79.1%. (iv)\nWe also achieve state-of-the-art results for instance-level segmentation.",
+                    "authors": "Zifeng Wu, Chunhua Shen, Anton van den Hengel",
+                    "published": "2016-05-23",
+                    "updated": "2016-05-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.08465v2",
+                    "title": "DeepAtlas: Joint Semi-Supervised Learning of Image Registration and Segmentation",
+                    "abstract": "Deep convolutional neural networks (CNNs) are state-of-the-art for semantic\nimage segmentation, but typically require many labeled training samples.\nObtaining 3D segmentations of medical images for supervised training is\ndifficult and labor intensive. Motivated by classical approaches for joint\nsegmentation and registration we therefore propose a deep learning framework\nthat jointly learns networks for image registration and image segmentation. In\ncontrast to previous work on deep unsupervised image registration, which showed\nthe benefit of weak supervision via image segmentations, our approach can use\nexisting segmentations when available and computes them via the segmentation\nnetwork otherwise, thereby providing the same registration benefit. Conversely,\nsegmentation network training benefits from the registration, which essentially\nprovides a realistic form of data augmentation. Experiments on knee and brain\n3D magnetic resonance (MR) images show that our approach achieves large\nsimultaneous improvements of segmentation and registration accuracy (over\nindependently trained networks) and allows training high-quality models with\nvery limited training data. Specifically, in a one-shot-scenario (with only one\nmanually labeled image) our approach increases Dice scores (%) over an\nunsupervised registration network by 2.7 and 1.8 on the knee and brain images\nrespectively.",
+                    "authors": "Zhenlin Xu, Marc Niethammer",
+                    "published": "2019-04-17",
+                    "updated": "2019-07-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.06753v1",
+                    "title": "3rd Place Solution for PVUW Challenge 2023: Video Panoptic Segmentation",
+                    "abstract": "In order to deal with the task of video panoptic segmentation in the wild, we\npropose a robust integrated video panoptic segmentation solution. In our\nsolution, we regard the video panoptic segmentation task as a segmentation\ntarget querying task, represent both semantic and instance targets as a set of\nqueries, and then combine these queries with video features extracted by neural\nnetworks to predict segmentation masks. In order to improve the learning\naccuracy and convergence speed of the solution, we add additional tasks of\nvideo semantic segmentation and video instance segmentation for joint training.\nIn addition, we also add an additional image semantic segmentation model to\nfurther improve the performance of semantic classes. In addition, we also add\nsome additional operations to improve the robustness of the model. Extensive\nexperiments on the VIPSeg dataset show that the proposed solution achieves\nstate-of-the-art performance with 50.04\\% VPQ on the VIPSeg test set, which is\n3rd place on the video panoptic segmentation track of the PVUW Challenge 2023.",
+                    "authors": "Jinming Su, Wangwang Yang, Junfeng Luo, Xiaolin Wei",
+                    "published": "2023-06-11",
+                    "updated": "2023-06-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.03138v1",
+                    "title": "Quantity beats quality for semantic segmentation of corrosion in images",
+                    "abstract": "Dataset creation is typically one of the first steps when applying Artificial\nIntelligence methods to a new task; and the real world performance of models\nhinges on the quality and quantity of data available. Producing an image\ndataset for semantic segmentation is resource intensive, particularly for\nspecialist subjects where class segmentation is not able to be effectively\nfarmed out. The benefit of producing a large, but poorly labelled, dataset\nversus a small, expertly segmented dataset for semantic segmentation is an open\nquestion. Here we show that a large, noisy dataset outperforms a small,\nexpertly segmented dataset for training a Fully Convolutional Network model for\nsemantic segmentation of corrosion in images. A large dataset of 250 images\nwith segmentations labelled by undergraduates and a second dataset of just 10\nimages, with segmentations labelled by subject matter experts were produced.\nThe mean Intersection over Union and micro F-score metrics were compared after\ntraining for 50,000 epochs. This work is illustrative for researchers setting\nout to develop deep learning models for detection and location of specialist\nfeatures.",
+                    "authors": "Will Nash, Tom Drummond, Nick Birbilis",
+                    "published": "2018-06-30",
+                    "updated": "2018-06-30",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.08455v3",
+                    "title": "Open-world Semantic Segmentation via Contrasting and Clustering Vision-Language Embedding",
+                    "abstract": "To bridge the gap between supervised semantic segmentation and real-world\napplications that acquires one model to recognize arbitrary new concepts,\nrecent zero-shot segmentation attracts a lot of attention by exploring the\nrelationships between unseen and seen object categories, yet requiring large\namounts of densely-annotated data with diverse base classes. In this paper, we\npropose a new open-world semantic segmentation pipeline that makes the first\nattempt to learn to segment semantic objects of various open-world categories\nwithout any efforts on dense annotations, by purely exploiting the\nimage-caption data that naturally exist on the Internet. Our method,\nVision-language-driven Semantic Segmentation (ViL-Seg), employs an image and a\ntext encoder to generate visual and text embeddings for the image-caption data,\nwith two core components that endow its segmentation ability: First, the image\nencoder is jointly trained with a vision-based contrasting and a cross-modal\ncontrasting, which encourage the visual embeddings to preserve both\nfine-grained semantics and high-level category information that are crucial for\nthe segmentation task. Furthermore, an online clustering head is devised over\nthe image encoder, which allows to dynamically segment the visual embeddings\ninto distinct semantic groups such that they can be classified by comparing\nwith various text embeddings to complete our segmentation pipeline. Experiments\nshow that without using any data with dense annotations, our method can\ndirectly segment objects of arbitrary categories, outperforming zero-shot\nsegmentation methods that require data labeling on three benchmark datasets.",
+                    "authors": "Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, Xiaodan Liang",
+                    "published": "2022-07-18",
+                    "updated": "2022-10-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1511.06988v1",
+                    "title": "Learning High-level Prior with Convolutional Neural Networks for Semantic Segmentation",
+                    "abstract": "This paper proposes a convolutional neural network that can fuse high-level\nprior for semantic image segmentation. Motivated by humans' vision recognition\nsystem, our key design is a three-layer generative structure consisting of\nhigh-level coding, middle-level segmentation and low-level image to introduce\nglobal prior for semantic segmentation. Based on this structure, we proposed a\ngenerative model called conditional variational auto-encoder (CVAE) that can\nbuild up the links behind these three layers. These important links include an\nimage encoder that extracts high level info from image, a segmentation encoder\nthat extracts high level info from segmentation, and a hybrid decoder that\noutputs semantic segmentation from the high level prior and input image. We\ntheoretically derive the semantic segmentation as an optimization problem\nparameterized by these links. Finally, the optimization problem enables us to\ntake advantage of state-of-the-art fully convolutional network structure for\nthe implementation of the above encoders and decoder. Experimental results on\nseveral representative datasets demonstrate our supreme performance for\nsemantic segmentation.",
+                    "authors": "Haitian Zheng, Yebin Liu, Mengqi Ji, Feng Wu, Lu Fang",
+                    "published": "2015-11-22",
+                    "updated": "2015-11-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.13118v1",
+                    "title": "Automatic Preprocessing and Ensemble Learning for Low Quality Cell Image Segmentation",
+                    "abstract": "We propose an automatic preprocessing and ensemble learning for segmentation\nof cell images with low quality. It is difficult to capture cells with strong\nlight. Therefore, the microscopic images of cells tend to have low image\nquality but these images are not good for semantic segmentation. Here we\npropose a method to translate an input image to the images that are easy to\nrecognize by deep learning. The proposed method consists of two deep neural\nnetworks. The first network is the usual training for semantic segmentation,\nand penultimate feature maps of the first network are used as filters to\ntranslate an input image to the images that emphasize each class. This is the\nautomatic preprocessing and translated cell images are easily classified. The\ninput cell image with low quality is translated by the feature maps in the\nfirst network, and the translated images are fed into the second network for\nsemantic segmentation. Since the outputs of the second network are multiple\nsegmentation results, we conduct the weighted ensemble of those segmentation\nimages. Two networks are trained by end-to-end manner, and we do not need to\nprepare images with high quality for the translation. We confirmed that our\nproposed method can translate cell images with low quality to the images that\nare easy to segment, and segmentation accuracy has improved using the weighted\nensemble learning.",
+                    "authors": "Sota Kato, Kazuhiro Hotta",
+                    "published": "2021-08-30",
+                    "updated": "2021-08-30",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.08988v1",
+                    "title": "Heterogeneous Feature Distillation Network for SAR Image Semantic Segmentation",
+                    "abstract": "Semantic segmentation for SAR (Synthetic Aperture Radar) images has attracted\nincreasing attention in the remote sensing community recently, due to SAR's\nall-time and all-weather imaging capability. However, SAR images are generally\nmore difficult to be segmented than their EO (Electro-Optical) counterparts,\nsince speckle noises and layovers are inevitably involved in SAR images. To\naddress this problem, we investigate how to introduce EO features to assist the\ntraining of a SAR-segmentation model, and propose a heterogeneous feature\ndistillation network for segmenting SAR images, called HFD-Net, where a\nSAR-segmentation student model gains knowledge from a pre-trained\nEO-segmentation teacher model. In the proposed HFD-Net, both the student and\nteacher models employ an identical architecture but different parameter\nconfigurations, and a heterogeneous feature distillation model is explored for\ntransferring latent EO features from the teacher model to the student model and\nthen enhancing the ability of the student model for SAR image segmentation. In\naddition, a heterogeneous feature alignment module is explored to aggregate\nmulti-scale features for segmentation in each of the student model and teacher\nmodel. Extensive experimental results on two public datasets demonstrate that\nthe proposed HFD-Net outperforms seven state-of-the-art SAR image semantic\nsegmentation methods.",
+                    "authors": "Gao Mengyu, Dong Qiulei",
+                    "published": "2022-10-17",
+                    "updated": "2022-10-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.17874v1",
+                    "title": "SmooSeg: Smoothness Prior for Unsupervised Semantic Segmentation",
+                    "abstract": "Unsupervised semantic segmentation is a challenging task that segments images\ninto semantic groups without manual annotation. Prior works have primarily\nfocused on leveraging prior knowledge of semantic consistency or priori\nconcepts from self-supervised learning methods, which often overlook the\ncoherence property of image segments. In this paper, we demonstrate that the\nsmoothness prior, asserting that close features in a metric space share the\nsame semantics, can significantly simplify segmentation by casting unsupervised\nsemantic segmentation as an energy minimization problem. Under this paradigm,\nwe propose a novel approach called SmooSeg that harnesses self-supervised\nlearning methods to model the closeness relationships among observations as\nsmoothness signals. To effectively discover coherent semantic segments, we\nintroduce a novel smoothness loss that promotes piecewise smoothness within\nsegments while preserving discontinuities across different segments.\nAdditionally, to further enhance segmentation quality, we design an asymmetric\nteacher-student style predictor that generates smoothly updated pseudo labels,\nfacilitating an optimal fit between observations and labeling outputs. Thanks\nto the rich supervision cues of the smoothness prior, our SmooSeg significantly\noutperforms STEGO in terms of pixel accuracy on three datasets: COCOStuff\n(+14.9%), Cityscapes (+13.0%), and Potsdam-3 (+5.7%).",
+                    "authors": "Mengcheng Lan, Xinjiang Wang, Yiping Ke, Jiaxing Xu, Litong Feng, Wayne Zhang",
+                    "published": "2023-10-27",
+                    "updated": "2023-10-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1809.10245v1",
+                    "title": "Cylindrical Transform: 3D Semantic Segmentation of Kidneys With Limited Annotated Images",
+                    "abstract": "In this paper, we propose a novel technique for sampling sequential images\nusing a cylindrical transform in a cylindrical coordinate system for kidney\nsemantic segmentation in abdominal computed tomography (CT). The images\ngenerated from a cylindrical transform augment a limited annotated set of\nimages in three dimensions. This approach enables us to train contemporary\nclassification deep convolutional neural networks (DCNNs) instead of fully\nconvolutional networks (FCNs) for semantic segmentation. Typical semantic\nsegmentation models segment a sequential set of images (e.g. CT or video) by\nsegmenting each image independently. However, the proposed method not only\nconsiders the spatial dependency in the x-y plane, but also the spatial\nsequential dependency along the z-axis. The results show that classification\nDCNNs, trained on cylindrical transformed images, can achieve a higher\nsegmentation performance value than FCNs using a limited number of annotated\nimages.",
+                    "authors": "Hojjat Salehinejad, Sumeya Naqvi, Errol Colak, Joseph Barfett, Shahrokh Valaee",
+                    "published": "2018-09-24",
+                    "updated": "2018-09-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.NE"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.12417v2",
+                    "title": "SLAMs: Semantic Learning based Activation Map for Weakly Supervised Semantic Segmentation",
+                    "abstract": "Recent mainstream weakly-supervised semantic segmentation (WSSS) approaches\nmainly relies on image-level classification learning, which has limited\nrepresentation capacity. In this paper, we propose a novel semantic learning\nbased framework, named SLAMs (Semantic Learning based Activation Map), for\nWSSS.",
+                    "authors": "Junliang Chen, Xiaodong Zhao, Minmin Liu, Linlin Shen",
+                    "published": "2022-10-22",
+                    "updated": "2022-11-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.02095v1",
+                    "title": "Content-aware Token Sharing for Efficient Semantic Segmentation with Vision Transformers",
+                    "abstract": "This paper introduces Content-aware Token Sharing (CTS), a token reduction\napproach that improves the computational efficiency of semantic segmentation\nnetworks that use Vision Transformers (ViTs). Existing works have proposed\ntoken reduction approaches to improve the efficiency of ViT-based image\nclassification networks, but these methods are not directly applicable to\nsemantic segmentation, which we address in this work. We observe that, for\nsemantic segmentation, multiple image patches can share a token if they contain\nthe same semantic class, as they contain redundant information. Our approach\nleverages this by employing an efficient, class-agnostic policy network that\npredicts if image patches contain the same semantic class, and lets them share\na token if they do. With experiments, we explore the critical design choices of\nCTS and show its effectiveness on the ADE20K, Pascal Context and Cityscapes\ndatasets, various ViT backbones, and different segmentation decoders. With\nContent-aware Token Sharing, we are able to reduce the number of processed\ntokens by up to 44%, without diminishing the segmentation quality.",
+                    "authors": "Chenyang Lu, Daan de Geus, Gijs Dubbelman",
+                    "published": "2023-06-03",
+                    "updated": "2023-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2205.13278v1",
+                    "title": "Semantic Segmentation for Thermal Images: A Comparative Survey",
+                    "abstract": "Semantic segmentation is a challenging task since it requires excessively\nmore low-level spatial information of the image compared to other computer\nvision problems. The accuracy of pixel-level classification can be affected by\nmany factors, such as imaging limitations and the ambiguity of object\nboundaries in an image. Conventional methods exploit three-channel RGB images\ncaptured in the visible spectrum with deep neural networks (DNN). Thermal\nimages can significantly contribute during the segmentation since thermal\nimaging cameras are capable of capturing details despite the weather and\nillumination conditions. Using infrared spectrum in semantic segmentation has\nmany real-world use cases, such as autonomous driving, medical imaging,\nagriculture, defense industry, etc. Due to this wide range of use cases,\ndesigning accurate semantic segmentation algorithms with the help of infrared\nspectrum is an important challenge. One approach is to use both visible and\ninfrared spectrum images as inputs. These methods can accomplish higher\naccuracy due to enriched input information, with the cost of extra effort for\nthe alignment and processing of multiple inputs. Another approach is to use\nonly thermal images, enabling less hardware cost for smaller use cases. Even\nthough there are multiple surveys on semantic segmentation methods, the\nliterature lacks a comprehensive survey centered explicitly around semantic\nsegmentation using infrared spectrum. This work aims to fill this gap by\npresenting algorithms in the literature and categorizing them by their input\nimages.",
+                    "authors": "Z\u00fclfiye K\u00fct\u00fck, G\u00f6rkem Algan",
+                    "published": "2022-05-26",
+                    "updated": "2022-05-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.12545v1",
+                    "title": "Unsupervised Domain Adaptation for Semantic Segmentation by Content Transfer",
+                    "abstract": "In this paper, we tackle the unsupervised domain adaptation (UDA) for\nsemantic segmentation, which aims to segment the unlabeled real data using\nlabeled synthetic data. The main problem of UDA for semantic segmentation\nrelies on reducing the domain gap between the real image and synthetic image.\nTo solve this problem, we focused on separating information in an image into\ncontent and style. Here, only the content has cues for semantic segmentation,\nand the style makes the domain gap. Thus, precise separation of content and\nstyle in an image leads to effect as supervision of real data even when\nlearning with synthetic data. To make the best of this effect, we propose a\nzero-style loss. Even though we perfectly extract content for semantic\nsegmentation in the real domain, another main challenge, the class imbalance\nproblem, still exists in UDA for semantic segmentation. We address this problem\nby transferring the contents of tail classes from synthetic to real domain.\nExperimental results show that the proposed method achieves the\nstate-of-the-art performance in semantic segmentation on the major two UDA\nsettings.",
+                    "authors": "Suhyeon Lee, Junhyuk Hyun, Hongje Seong, Euntai Kim",
+                    "published": "2020-12-23",
+                    "updated": "2020-12-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1502.04983v1",
+                    "title": "Context Tricks for Cheap Semantic Segmentation",
+                    "abstract": "Accurate semantic labeling of image pixels is difficult because intra-class\nvariability is often greater than inter-class variability. In turn, fast\nsemantic segmentation is hard because accurate models are usually too\ncomplicated to also run quickly at test-time. Our experience with building and\nrunning semantic segmentation systems has also shown a reasonably obvious\nbottleneck on model complexity, imposed by small training datasets. We\ntherefore propose two simple complementary strategies that leverage context to\ngive better semantic segmentation, while scaling up or down to train on\ndifferent-sized datasets.\n  As easy modifications for existing semantic segmentation algorithms, we\nintroduce Decorrelated Semantic Texton Forests, and the Context Sensitive Image\nLevel Prior. The proposed modifications are tested using a Semantic Texton\nForest (STF) system, and the modifications are validated on two standard\nbenchmark datasets, MSRC-21 and PascalVOC-2010. In Python based comparisons,\nour system is insignificantly slower than STF at test-time, yet produces\nsuperior semantic segmentations overall, with just push-button training.",
+                    "authors": "Thanapong Intharah, Gabriel J. Brostow",
+                    "published": "2015-02-17",
+                    "updated": "2015-02-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2007.13138v1",
+                    "title": "Virtual Multi-view Fusion for 3D Semantic Segmentation",
+                    "abstract": "Semantic segmentation of 3D meshes is an important problem for 3D scene\nunderstanding. In this paper we revisit the classic multiview representation of\n3D meshes and study several techniques that make them effective for 3D semantic\nsegmentation of meshes. Given a 3D mesh reconstructed from RGBD sensors, our\nmethod effectively chooses different virtual views of the 3D mesh and renders\nmultiple 2D channels for training an effective 2D semantic segmentation model.\nFeatures from multiple per view predictions are finally fused on 3D mesh\nvertices to predict mesh semantic segmentation labels. Using the large scale\nindoor 3D semantic segmentation benchmark of ScanNet, we show that our virtual\nviews enable more effective training of 2D semantic segmentation networks than\nprevious multiview approaches. When the 2D per pixel predictions are aggregated\non 3D surfaces, our virtual multiview fusion method is able to achieve\nsignificantly better 3D semantic segmentation results compared to all prior\nmultiview approaches and competitive with recent 3D convolution approaches.",
+                    "authors": "Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, Caroline Pantofaru",
+                    "published": "2020-07-26",
+                    "updated": "2020-07-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.00487v1",
+                    "title": "Linear Semantics in Generative Adversarial Networks",
+                    "abstract": "Generative Adversarial Networks (GANs) are able to generate high-quality\nimages, but it remains difficult to explicitly specify the semantics of\nsynthesized images. In this work, we aim to better understand the semantic\nrepresentation of GANs, and thereby enable semantic control in GAN's generation\nprocess. Interestingly, we find that a well-trained GAN encodes image semantics\nin its internal feature maps in a surprisingly simple way: a linear\ntransformation of feature maps suffices to extract the generated image\nsemantics. To verify this simplicity, we conduct extensive experiments on\nvarious GANs and datasets; and thanks to this simplicity, we are able to learn\na semantic segmentation model for a trained GAN from a small number (e.g., 8)\nof labeled images. Last but not least, leveraging our findings, we propose two\nfew-shot image editing approaches, namely Semantic-Conditional Sampling and\nSemantic Image Editing. Given a trained GAN and as few as eight semantic\nannotations, the user is able to generate diverse images subject to a\nuser-provided semantic layout, and control the synthesized image semantics. We\nhave made the code publicly available.",
+                    "authors": "Jianjin Xu, Changxi Zheng",
+                    "published": "2021-04-01",
+                    "updated": "2021-04-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.13697v1",
+                    "title": "Generalizable Semantic Vision Query Generation for Zero-shot Panoptic and Semantic Segmentation",
+                    "abstract": "Zero-shot Panoptic Segmentation (ZPS) aims to recognize foreground instances\nand background stuff without images containing unseen categories in training.\nDue to the visual data sparsity and the difficulty of generalizing from seen to\nunseen categories, this task remains challenging. To better generalize to\nunseen classes, we propose Conditional tOken aligNment and Cycle trAnsiTion\n(CONCAT), to produce generalizable semantic vision queries. First, a feature\nextractor is trained by CON to link the vision and semantics for providing\ntarget queries. Formally, CON is proposed to align the semantic queries with\nthe CLIP visual CLS token extracted from complete and masked images. To address\nthe lack of unseen categories, a generator is required. However, one of the\ngaps in synthesizing pseudo vision queries, ie, vision queries for unseen\ncategories, is describing fine-grained visual details through semantic\nembeddings. Therefore, we approach CAT to train the generator in\nsemantic-vision and vision-semantic manners. In semantic-vision, visual query\ncontrast is proposed to model the high granularity of vision by pulling the\npseudo vision queries with the corresponding targets containing segments while\npushing those without segments away. To ensure the generated queries retain\nsemantic information, in vision-semantic, the pseudo vision queries are mapped\nback to semantic and supervised by real semantic embeddings. Experiments on ZPS\nachieve a 5.2% hPQ increase surpassing SOTA. We also examine inductive ZPS and\nopen-vocabulary semantic segmentation and obtain comparative results while\nbeing 2 times faster in testing.",
+                    "authors": "Jialei Chen, Daisuke Deguchi, Chenkai Zhang, Hiroshi Murase",
+                    "published": "2024-02-21",
+                    "updated": "2024-02-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1610.01706v1",
+                    "title": "Exploiting Depth from Single Monocular Images for Object Detection and Semantic Segmentation",
+                    "abstract": "Augmenting RGB data with measured depth has been shown to improve the\nperformance of a range of tasks in computer vision including object detection\nand semantic segmentation. Although depth sensors such as the Microsoft Kinect\nhave facilitated easy acquisition of such depth information, the vast majority\nof images used in vision tasks do not contain depth information. In this paper,\nwe show that augmenting RGB images with estimated depth can also improve the\naccuracy of both object detection and semantic segmentation. Specifically, we\nfirst exploit the recent success of depth estimation from monocular images and\nlearn a deep depth estimation model. Then we learn deep depth features from the\nestimated depth and combine with RGB features for object detection and semantic\nsegmentation. Additionally, we propose an RGB-D semantic segmentation method\nwhich applies a multi-task training scheme: semantic label prediction and depth\nvalue regression. We test our methods on several datasets and demonstrate that\nincorporating information from estimated depth improves the performance of\nobject detection and semantic segmentation remarkably.",
+                    "authors": "Yuanzhouhan Cao, Chunhua Shen, Heng Tao Shen",
+                    "published": "2016-10-06",
+                    "updated": "2016-10-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2111.02444v2",
+                    "title": "Panoptic 3D Scene Reconstruction From a Single RGB Image",
+                    "abstract": "Understanding 3D scenes from a single image is fundamental to a wide variety\nof tasks, such as for robotics, motion planning, or augmented reality. Existing\nworks in 3D perception from a single RGB image tend to focus on geometric\nreconstruction only, or geometric reconstruction with semantic segmentation or\ninstance segmentation. Inspired by 2D panoptic segmentation, we propose to\nunify the tasks of geometric reconstruction, 3D semantic segmentation, and 3D\ninstance segmentation into the task of panoptic 3D scene reconstruction - from\na single RGB image, predicting the complete geometric reconstruction of the\nscene in the camera frustum of the image, along with semantic and instance\nsegmentations. We thus propose a new approach for holistic 3D scene\nunderstanding from a single RGB image which learns to lift and propagate 2D\nfeatures from an input image to a 3D volumetric scene representation. We\ndemonstrate that this holistic view of joint scene reconstruction, semantic,\nand instance segmentation is beneficial over treating the tasks independently,\nthus outperforming alternative approaches.",
+                    "authors": "Manuel Dahnert, Ji Hou, Matthias Nie\u00dfner, Angela Dai",
+                    "published": "2021-11-03",
+                    "updated": "2022-05-16",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                }
+            ],
+            [
+                {
+                    "url": "http://arxiv.org/abs/2404.01692v2",
+                    "title": "Beyond Image Super-Resolution for Image Recognition with Task-Driven Perceptual Loss",
+                    "abstract": "In real-world scenarios, image recognition tasks, such as semantic\nsegmentation and object detection, often pose greater challenges due to the\nlack of information available within low-resolution (LR) content. Image\nsuper-resolution (SR) is one of the promising solutions for addressing the\nchallenges. However, due to the ill-posed property of SR, it is challenging for\ntypical SR methods to restore task-relevant high-frequency contents, which may\ndilute the advantage of utilizing the SR method. Therefore, in this paper, we\npropose Super-Resolution for Image Recognition (SR4IR) that effectively guides\nthe generation of SR images beneficial to achieving satisfactory image\nrecognition performance when processing LR images. The critical component of\nour SR4IR is the task-driven perceptual (TDP) loss that enables the SR network\nto acquire task-specific knowledge from a network tailored for a specific task.\nMoreover, we propose a cross-quality patch mix and an alternate training\nframework that significantly enhances the efficacy of the TDP loss by\naddressing potential problems when employing the TDP loss. Through extensive\nexperiments, we demonstrate that our SR4IR achieves outstanding task\nperformance by generating SR images useful for a specific image recognition\ntask, including semantic segmentation, object detection, and image\nclassification. The implementation code is available at\nhttps://github.com/JaehaKim97/SR4IR.",
+                    "authors": "Jaeha Kim, Junghun Oh, Kyoung Mu Lee",
+                    "published": "2024-04-02",
+                    "updated": "2024-04-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Original Paper",
+                    "paper_cat": "Semantic AND Segmentation AND Image",
+                    "gt": "Single Image Super-Resolution. In the era of deep learning, deep neural network-based methods have risen to prominence in the field of SR. After the introduction of the simple deep-learning-based SR network SRCNN [12], VDSR [29] brought about innovation in network architecture design by stacking deeper convolutional layers with the inclusion of residual connections. EDSR [35] and SRResNet [32] brought the Residual Block concept into SR networks, and subsequent methods such as RDN [77], DBPN [18] further evolved the Dense Residual Block, resulting in high-performance network structures. SAN [11], HAN [45], and RCAN [76] enhance the SR performance by incorporating channel attention mechanisms, enabling the exploitation of useful channel-wise features. Recently, transformer-based methods such as SwinIR [33], Restormer [70], and Uformer [65] have demonstrated remarkable SR performance by harnessing the powerful capabilities of Vision Transformers [13]. On the other hand, perceptual super-resolution (SR) methods constitute an alternative approach dedicated to restoring photorealistic SR results. Johnson et al. [27] introduced perceptual loss, which minimizes the distance in the feature space and has shown its effectiveness in reconstructing high-frequency components in SR. Zhang et al. [74] illustrated that the perceptual quality of images is not directly correlated with the peak signal-to-noise ratio (PSNR), then introduced the perceptual metric known as LPIPS. Subsequent to the development of adversarial loss [17], GANbased SR methods [32, 34, 46, 52, 56, 57, 62, 63] have been successful in achieving realistic SR results. More recently, diffusion-based SR methods [8, 28, 43, 51, 64, 66, 68] have further pushed the boundaries by leveraging a diffusionbased model [21] to produce even more realistic SR results. 2 Figure 2. Overview of the proposed SR4IR framework. Our SR4IR framework consists of two training phases, where SR and task networks are alternately trained. During the first phase, SR4IR updates the SR network using the TDP loss, which is introduced in Section 3.1, while the task network is temporarily frozen. In the second phase, SR4IR updates the task network using the proposed data augmentation strategy called CQMix, which is introduced in Section 3.2, while the SR network is temporarily frozen. Image restoration for high-level vision tasks. There exist pioneering recent works [9, 10, 38, 41, 54, 78] focused on designing image restoration techniques that offer advantages in high-level computer vision tasks. Zhou et al. [79] introduced a weight map that uses pixel loss primarily in high-frequency regions, showing its effectiveness in image classification. TDSR [19] presented an object detectiondriven SR method that updates the SR network using a combination of reconstruction and detection losses and provides a comprehensive study on balancing the joint loss. Wang et al. [61] proposed adding an SR branch to the segmentation branch, enabling the effective utilization of HR representation. Bai et al. [2, 3] have demonstrated that incorporating adversarial loss through a multitask generative adversarial network can further enhance the performance of low-quality object detection. In contrast to previous approaches, our method successfully restores task-specific high-frequency components by leveraging the concept of perceptual loss. Moreover, we are the first to introduce a generalizable taskcombined SR framework that achieves remarkable performance across a range of image recognition tasks.",
+                    "pre_questions": [],
+                    "main_content": "Introduction After the pioneering AlexNet [31], numerous deep neural network-based methods have achieved superior performance in various image recognition tasks, such as image classification [13, 20, 23, 24, 40, 55, 58], object detection [37, 39, 48, 49], and semantic segmentation [6, 42, 67, 75]. In real-world applications, such as medical image recognition and surveillance systems, a network for a task, which we refer to as a task network, may encounter input images containing low-resolution (LR) contents due to (a) Ground-truth (b) Bilinear-up (c) SwinIR [33] (d) SwinIR[33] + SR4IR(Ours) Figure 1. Visualizations and comparisons of the results on semantic segmentation task with PASCAL VOC dataset [14]. (a) The ground-truth image and the class label map. (b) The bilinear-upsampled image and the predicted result when trained on bilinear-upsampled images. (c) the SR result of the SwinIR [33] and the predicted result when trained on the SR images. (d) the SR result of the SwinIR and the predicted result when trained by the proposed SR4IR framework. We use DeepLabV3 [7] as a task network and the downsampling scale factor is x4. various factors, including limitations of cameras or storage capabilities and the presence of tiny objects. In this circumstance, achieving satisfactory task performance is limited by the lack of high-frequency components crucial for a task. In this paper, we address these challenges by employing single image super-resolution (SR) methods, as SR can serve as a prominent pre-processing tool to restore missing high-frequency details in LR images. Conventionally, SR aims to restore high-resolution (HR) images mainly via pixel-wise loss [18, 29, 33, 35, 73, 77] or perceptual loss [5, 15, 27]. However, due to the ill-posed nature of SR, these approaches may not ensure the reconstruction of features crucial for achieving better performance for a specific task, thereby leading to marginal performance improvement from SR. Thus, in this paper, our primary goal is to develop a comprehensive framework, named Super-Resolution for 1 arXiv:2404.01692v2  [cs.CV]  4 Apr 2024 Image Recognition (SR4IR), wherein the SR and task network interplay harmoniously to generate SR images, especially focusing on incorporating task-relevant features, which are exploited to enhance task performance. To attain our goal, we focus on leveraging the concept of perceptual loss, as it is devised to encourage SR images to mimic the HR counterparts within a feature space. Typically, the perceptual loss is computed using VGGNet [55] pre-trained on ImageNet dataset [50]. However, we note that employing perceptual loss in the conventional manner results in modest improvements in task performance. Instead, we propose a Task-Driven Perceptual (TDP) loss that enables the emulation of HR images within the feature space of a task network. By doing so, our TDP loss facilitates the restoration of high-frequency details, especially beneficial for enhancing task performance. The effectiveness of our TDP loss is significantly influenced by the knowledge that the task network acquires. For high-level vision tasks, it is widely acknowledged that a network can fall into learning non-intrinsic features, leading to biased feature representations [16]. A network possessing biased knowledge on a task can diminish the efficacy of the TDP loss, as it may fail to deliver meaningful signals to the SR network for the restoration of other valuable features. To tackle the problem, we introduce a novel data augmentation strategy for training the task network, called Cross-Quality patch Mix (CQMix), which blends SR and HR images at the patch level to randomly eliminate high-frequency components. Consequently, CQMix can prevent a task network from learning shortcut features, thereby allowing us to fully harness the benefits of TDP loss. In addition to calculating perceptual loss using a task network, we introduce further variations to the perceptual loss by applying it to a task network undergoing training instead of a pre-trained one. This strategy substantially improves the effectiveness of our TDP loss, since a pre-trained (then fixed) task network cannot learn to produce meaningful features from SR images. To this end, we constitute the SR4IR as an alternate training framework that trains the SR and task network alternately throughout the training procedure. In the first phase, we train the SR network using the TDP loss with the temporarily frozen task network. Subsequently, we train only the task network with task-specific loss using data augmented by the CQMix. To evaluate the proposed methods, we have examined the integration of SR with three representative image recognition tasks: semantic segmentation, object detection, and image classification. For each task, we evaluate our method on two representative SR architectures, EDSR [35] and SwinIR [33], with scale factors of \u00d74 and \u00d78. Our extensive experiments clearly demonstrate that our SR4IR significantly improves task performance while achieving perceptually appealing SR results compared to baseline methods, as shown in Figure 1. These results indicate that our framework can be generally adopted in various applications. We summarize our contributions as follows: \u2022 To our knowledge, we are the first to introduce a comprehensive SR framework that addresses challenges posed by LR contents across various image recognition tasks. \u2022 We propose the task-driven perceptual (TDP) loss that facilitates learning to restore task-related features acquired by a task network, enhancing task performance. \u2022 We propose the cross-quality patch mix and the alternate training framework to address the potential problems of TDP loss, further enhancing the efficacy of TDP loss. \u2022 The extensive experiments demonstrate that our SR4IR is generally applicable across a wide range of image recognition tasks to enhance task performance while generating perceptually satisfying SR images. Overview. Figure 2 shows the overview of the proposed Super-Resolution for Image Recognition (SR4IR) framework. The SR4IR comprises two distinct training phases that are alternately executed to train either the superresolution (SR) or the task network. In the first phase, we train the SR network using the Task-Driven Perceptual (TDP) loss, which is designed to guide the SR network to restore high-frequency details relevant to the task (Section 3.1). In the second phase, we train the task network using the novel Cross-Quality patch Mix (CQMix) augmentation method, which prevents the task network from learning biased features, further enhancing the effectiveness of the TDP loss (Section 3.2). Lastly, we present the effect of alternate training with respect to TDP loss and summarize our final training objectives in each training phase (Section 3.3). Notation. Let IHR \u2208IHR and ILR \u2208ILR denote the highresolution (HR) and low-resolution (LR) images, paired with their corresponding task labels y \u2208Y (e.g., class labels in image classification task). S\u03b8SR : ILR \u2192ISR denotes the SR network parameterized by \u03b8SR where ISR is the super-resolved image set. Given the SR images ISR, we predict the corresponding labels using the task network: H\u03b8head \u25e6F\u03b8feat : ISR \u2192Y, where F\u03b8feat and H\u03b8head indicate the feature extractor and the task-specific head module (e.g., the last fully-connected layer for image classification tasks) parameterized by \u03b8feat and \u03b8head, respectively. Problem Definition. The goal of SR4IR is to produce photo-realistic SR images from ILR that can benefit the subsequent image recognition task, which is formally given by: \\l abel {eq:problem_definition} \\beg i n {aligned} \\ min \\l i mits _{\\bo lds t and second terms are used to evalua (1) where the first and second terms are used to evaluate the quality of the super-resolved results (e.g., LPIPS [74]) and the performance for the target high-level vision task (e.g., classification accuracy), respectively. 3.1. Task-Driven Perceptual Loss High-resolution (HR) images contain high-frequency information that is crucial to solving high-level vision tasks. However, due to the ill-posed nature of the SR, such taskrelevant features are hard to restore using traditional pixelwise reconstruction loss. Even with the use of perceptual loss [5, 15, 27], designed to produce realistic SR results, there is no guarantee of successfully reconstructing useful features for a specific task. To address this challenge, we propose task-driven perceptual (TDP) loss that effectively guides the SR network to produce images containing highfrequency details relevant to the task. Specifically, the TDP loss aims to maximize the similarity between HR and SR 3 images within the latent space of the feature extractor of the task network F\u03b8feat, whose outputs are substantially relevant to the task. We define the TDP loss as follows:  \\la b el {eq:perce p } \\mathcal {L}_{\\text {TDP}} = \\lVert F_{\\boldsymbol {\\theta }_\\text {feat}}(\\boldsymbol {I}_\\text {SR}) F_{\\boldsymbol {\\theta }_\\text {feat}}(\\boldsymbol {I}_\\text {HR}) \\rVert _1.  (2) 3.2. Cross-Quality Patch Mix The effectiveness of TDP loss is inherently related to the training procedure of the task network. For high-level vision tasks, it is widely known that a network is susceptible to shortcut learning [16], which means that a network can rely on biased and non-intrinsic features in training data, leading to poor generalization ability. In particular, learning biased features can pose challenges in employing the TDP loss, as biased features learned by the task network can hinder the SR network from acquiring comprehensive features present in HR images through the TDP loss. For instance, if a task network exclusively learns \u2018beak\u2019 features to classify bird species, the TDP loss may struggle to effectively guide SR images to restore other key high-frequency contents, such as \u2018wings\u2019 or \u2018feet\u2019. This limitation could hinder the overall task performance boosting that could be attained through the SR method. To address this issue, we introduce a novel data augmentation strategy called Cross-Quality patch Mix (CQMix) when training the task network in the second phase. CQMix is designed to enhance the effectiveness of the TDP loss by encouraging the task network to use diverse high-frequency content as clues. Specifically, CQMix augments an input image by randomly choosing either HR or SR patches for each grid region, as shown in Figure 3, where the number of patches is hyper-parameter. By doing so, CQMix creates an image in which high-frequency components are randomly erased at a patch level, effectively preventing the task network from learning specific high-frequency features as shortcuts. Moreover, it is noteworthy that our CQMix is generally applicable to various vision tasks, unlike other data augmentation strategies, such as Mix-Up [72] or CutMix [69], as long as HR and LR pairs are available. 3.3. Alternate Training Framework Conventional perceptual loss calculates the similarity between SR and HR images within the feature spaces of a fixed network, such as fixed VGG [55] pre-trained on ImageNet dataset [50]. However, this approach can be problematic for TDP loss, which aims to restore task-relevant features from HR images. To effectively utilize the TDP loss to enhance task performance, the task network must be capable of extracting task-relevant features from SR images. However, since the fixed network is not originally trained on ISR, which keeps changing during training, it cannot effectively Figure 3. Concept of CQMix. The black and white region in the figure represents 0 and 1 region in the binary mask M. extract task-relevant features from ISR, thus reducing the effectiveness of the TDP loss. To resolve this domain gap issue, we propose an alternative training framework, as illustrated in Figure 2, which enables us to employ the feature space of the on-training feature extractor of the task network F\u03b8feat, for calculating the TDP loss. By doing so, we can further boost the efficacy of our TDP loss by providing a more meaningful guide to the SR network to acquire task-related knowledge. Specifically, in the first phase, we train SR network S\u03b8SR with pixel-wise reconstruction loss Lpixel and the proposed TDP loss, while keeping the feature extractor F\u03b8feat frozen. In the second phase, we train the task network H\u03b8head \u25e6F\u03b8feat with a task-specific loss Ltask, e.g., cross-entropy loss for image classification, while freezing the SR network. We use three types of images in this phase: HR, SR, and augmented images by our CQMix. In summary, the objectives of our SR4IR in each phase are defined as follows:  \\l abe l {eq:training_lo ss} \\ begin  { align ed } \\ min \\limits  _{\\boldsymb o l {\\theta }_{ \\te xt  {SR} }} \\:\\: & \\mathcal {L}_{\\text {pixel}}(S_{\\boldsymbol {\\theta }_{\\text {SR}}} (\\boldsymbol {I}_{\\text {LR}}), \\boldsymbol {I}_{\\text {HR}}) + \\mathcal {L}_{\\text {TDP}}, & \\text {\\textit {in}\\,~phase 1,}\\\\ \\min \\limits _{\\boldsymbol {\\theta }_{\\text {feat}},\\boldsymbol {\\theta }_{\\text {head}}} \\:\\: & \\mathcal {L}_{\\text {task}}(H_{\\boldsymbol {\\theta }_{\\text {head}}} \\circ F_{\\boldsymbol {\\theta }_{\\text {feat}}} (\\boldsymbol {I}_{\\text {cat}}), \\boldsymbol {y}), & \\text {\\textit {in}\\,~phase 2,}\\\\ \\end {aligned} (3) where Icat is the concatenation of ISR, IHR, and Iaug, where the Iaug are the images augmented by the CQMix. 4. Experiments Datasets. We select representative datasets and corresponding evaluation metrics for various image recognition tasks, including semantic segmentation, object detection, and image classification. For semantic segmentation and object detection, we use the PASCAL VOC2012 dataset [14]. For the image classification task, we use the Stanford Cars [30] and CUB-200-2011 [60] datasets. Following the conventional SR datasets [1, 4, 71], we construct the corresponding LR datasets applying bicubic downsampling with scale factors x4 and x8. Evaluation. We denote the task network H\u03b8head \u25e6F\u03b8feat and the SR network S\u03b8SR as T and S, respectively, for conve4 nience. Then, the available training scenarios when training both T and S concurrently are as follows: \u2022 IHR \u2192T : Train the T with Ltask using (IHR, y). This can be regarded as an oracle in our problem. \u2022 ILR \u2192T : Train the T with Ltask using (B (ILR) , y), where B is bilinear upsampling operator. \u2022 S \u2192T : Train the S with Lpixel using (ILR, IHR) first. Then train T with Ltask using (S (ILR) , y), while freezing S, as in previous work [79]. \u2022 T \u2192S : Train T with Ltask using (IHR, y) first. Then, train S with Lpixel+Ltask using (ILR, IHR, y), while freezing T, as in previous work [19, 38, 41]. \u2022 S + T : Jointly train both S and T with Lpixel + Ltask using (ILR, IHR, y). We evaluate the proposed method, named Super-Resolution for Image Recognition (SR4IR), by comparing it with the above methods. For evaluation metrics, we use mean Intersection over Union (mIoU) for semantic segmentation, mean Average Precision (mAP) for object detection, and Top-1 accuracy for image classification. The mAP score is calculated via the COCO [36] mAP evaluator, which uses 10 IoU thresholds from 0.5 to 0.95 with a step size of 0.05. Furthermore, we use LPIPS [74] and PSNR to measure the quality of the SR images. Network architecture. For the SR network, we employ two architectures: EDSR-baseline [35] and SwinIR [33]. We initialize the SR network from network weights pretrained on DIV2K [1], widely used for SR training. Regarding the task network, we select representative network architectures tailored to each specific task. Specifically, we adopt DeepLabV3 [7] for semantic segmentation, FasterRCNN [49] for object detection, and ResNet18 [20] for image classification. The DeepLabV3 and Faster-RCNN are based on MobileNetV3 [22] backbone. The backbone networks are initialized using ImageNet [50] pre-trained weights provided by PyTorch [47] for easier optimization. Implementations. We utilize AdamW [26] and SGD optimizer for the SR and task network. The learning rates are set to 10\u22124 for the SR network, 2 \u00d7 10\u22122 for DeepLabV3 and Faster-RCNN, and 3 \u00d7 10\u22122 for ResNet18, as these values have been shown to produce the best results for their respective tasks. The batch sizes are set to 16 for both semantic segmentation and object detection and 128 for image classification. We apply cosine annealing [25] scheduling to all optimizers, and all experiments are conducted under PyTorch [47] with eight NVIDIA RTX A6000 GPUs. For the CQMix, the number of patches is set to 64 for both semantic segmentation and object detection and 16 for image classification, regarding the image resolution for each dataset. 4.1. Quantitative results Semantic segmentation. Table 1 shows the quantitative results on the semantic segmentation task. On the x4 SR scale, our SR4IR significantly boosts the mIoU scores compared to the case trained on the bilinear upsampled images (ILR \u2192T), by +2.1 and +3.6, for EDSR-baseline and SwinIR, respectively. Remarkably, our method with SwinIR can achieve a mIoU score of 62.2, which is comparable to the 63.3 mIoU score achieved by the oracle (IHR \u2192T), demonstrating the effectiveness of combining the SR method when dealing with LR images. Furthermore, SR4IR surpasses the mIoU scores of all SR combined baselines, S \u2192T, T \u2192S, and S + T, with a gain of +0.5 and +2.7 compared to the second-best baselines, in the results of the EDSR-baseline with SR scales of x4 and x8, respectively. This validates that our method successfully restores the high-frequency details that are beneficial for the subsequent task. In particular, compared to ILR \u2192T, the performance improvement of the proposed SR4IR becomes even more significant as the SR scale increases from x4 to x8, the performance gain increases from +2.1 to +5.7, for the EDSR-baseline case. This indicates that our method is even more effective when the LR images lose a substantial amount of essential high-frequency details relevant to the task, demonstrating the effectiveness of SR4IR. Furthermore, the proposed SR4IR achieves the best LPIPS score in all cases, with a substantial margin of +12% gain over the second-best baseline for SwinIR on a scale of x4. These results demonstrate that our SR results are not only beneficial from a task perspective but are also capable of producing visually appealing results. Noticeably, separate and sequential training of SR and task networks (S \u2192T) achieves the best PSNR values but a much lower mIoU score than our SR4IR. The results validate that restoring high-frequency information relevant to the task is much more critical to improving task performance than minimizing the pixel-wise error from ground-truth images. Moreover, it is noteworthy that training jointly both SR and the task network (S + T) performs even worse than sequential training of the SR and task networks (S \u2192T). This indicates that simultaneously updating both the SR and task networks diminishes the restoring power of the SR due to joint optimization, highlighting the advantages of our alternate training framework. Object detection. Table 2 presents the numerical results of the object detection. The proposed SR4IR consistently achieves the best mAP score compared to all SR combined baselines. Specifically, for the case of x8 SR scale with the SwinIR model, our SR4IR outperforms the task network trained on bilinear upsampled images (ILR \u2192T) and the second-best SR combined baselines (S \u2192T) by substantial margins of +8.9 and +4.1, respectively. 5 Methods SR Method x4 x8 mIoU\u2191 LPIPS\u2193 PSNR\u2191 mIoU\u2191 LPIPS\u2193 PSNR\u2191 IHR \u2192T Oracle (IHR) 63.3 0.000 + inf 63.3 0.000 + inf ILR \u2192T Bilinear up 58.6 0.300 26.16 49.3 0.476 22.54 S \u2192T 60.2 0.241 28.94 52.3 0.407 24.41 T \u2192S EDSR56.6 0.352 26.83 49.2 0.461 23.14 S + T baseline [35] 58.0 0.318 27.64 50.0 0.480 23.58 SR4IR (Ours) 60.7 0.220 28.21 55.0 0.380 23.91 S \u2192T SwinIR [33] 61.4 0.221 29.57 53.9 0.387 24.75 T \u2192S 53.2 0.376 26.04 48.7 0.466 23.06 S + T 57.5 0.341 27.55 50.0 0.480 23.63 SR4IR (Ours) 62.2 0.194 28.83 56.5 0.355 24.23 Table 1. Performance on the semantic segmentation. We adopt DeepLabV3 [7] with the MobileNetV3 [22] backbone as the task network T and PASCAL VOC2012 [14] as an evaluation dataset. The bold and underlined mean the best and second best, respectively. Methods SR Method x4 x8 mAP\u2191 LPIPS\u2193 PSNR\u2191 mAP\u2191 LPIPS\u2193 PSNR\u2191 IHR \u2192T Oracle (IHR) 36.7 0.000 + inf 36.7 0.000 + inf ILR \u2192T Bilinear up 27.5 0.411 22.77 18.9 0.559 20.29 S \u2192T 30.3 0.336 24.62 21.9 0.494 21.62 T \u2192S EDSR27.2 0.317 23.37 15.5 0.476 20.48 S + T baseline [35] 28.3 0.354 23.71 20.3 0.506 20.98 SR4IR (Ours) 32.4 0.275 23.94 25.5 0.416 20.86 S \u2192T SwinIR [33] 31.5 0.309 25.13 23.7 0.474 21.85 T \u2192S 29.7 0.284 23.59 18.9 0.449 20.59 S + T 31.4 0.331 23.94 21.9 0.489 21.11 SR4IR (Ours) 33.9 0.239 24.38 27.8 0.382 21.07 Table 2. Performance on the object detection. We adopt Faster-RCNN [49] with the MobileNetV3 backbone [22] as the task network T and PASCAL VOC2012 [14] as an evaluation dataset. The bold and underlined mean the best and second best, respectively. Methods SR Method StanfordCars [30] CUB-200-2011 [60] x4 x8 x4 x8 Top-1 Acc.\u2191(%) LPIPS\u2193 Top-1 Acc.\u2191(%) LPIPS\u2193 Top-1 Acc.\u2191(%) LPIPS\u2193 Top-1 Acc.\u2191(%) LPIPS\u2193 IHR \u2192T Oracle (IHR) 86.4 0.000 86.4 0.000 78.3 0.000 78.3 0.000 ILR \u2192T Bilinear up 78.7 0.304 61.0 0.478 70.8 0.298 58.2 0.465 S \u2192T 82.7 0.176 68.3 0.328 71.1 0.216 59.8 0.365 T \u2192S EDSR80.1 0.190 56.9 0.448 70.5 0.208 56.5 0.401 S + T baseline [35] 80.8 0.196 65.2 0.378 70.8 0.226 58.6 0.380 SR4IR (Ours) 83.3 0.158 71.4 0.326 72.8 0.179 62.6 0.326 S \u2192T SwinIR [33] 84.0 0.159 71.4 0.308 72.8 0.201 61.2 0.349 T \u2192S 82.0 0.162 61.0 0.397 72.1 0.190 58.7 0.377 S + T 83.4 0.179 66.6 0.355 72.9 0.212 61.2 0.364 SR4IR (Ours) 85.1 0.136 73.9 0.292 74.7 0.157 65.3 0.301 Table 3. Performance on the image classification. We adopt ResNet18 [20] with a fully-connected layer as the task network T. The bold and underlined mean the best and second best, respectively. Image classification. Table 3 shows the quantitative results of the image classification task. Our SR4IR significantly improves image classification performance, remarkably achieving +12.9% improvement on the x8 scale StanfordCars dataset and a +7.1% enhancement on the x8 scale CUB-200-2011 dataset when combined with the SwinIR model, compared to the task network trained on bilinear upsampled images (ILR \u2192T). Furthermore, we note that our method consistently improves performance across various datasets, demonstrating its strong generalization capability. 4.2. Qualitative results In this section, we provide visual results of the SR with their corresponding task performance. Specifically, we compare our SR4IR with two baselines, ILR \u2192T and S \u2192T, where the latter achieved the second-best quantitative results. For a more extensive visual comparison with all baselines, please refer to our supplementary materials. In the context of semantic segmentation, as shown in Figure 4, the proposed SR4IR successfully generates a segmentation map that closely resembles the ground truth. In contrast, the baselines ILR \u2192T and S \u2192T produce incorrect estimations on \u2018husky\u2019. Regarding object detection, illustrated in Figure 5, SR4IR enables the object detection network to accurately identify two people who are closely situated, while ILR \u2192T and S \u2192T fail to distinguish and predict them as a single person. In image classification, as shown in Figure 6, the proposed SR4IR empowers the sub6 (a) ILR \u2192T (b) S \u2192T (c) SR4IR (Ours) (d) Ground-truth Figure 4. Visualization of images and semantic segmentation results on PASCAL VOC dataset. We present the restored images and the corresponding predicted segmentation maps, respectively. For (b) and (c), we use the SwinIR model with an SR scale factor of x4. (a) ILR \u2192T (b) S \u2192T (c) SR4IR (Ours) (d) Ground-truth Figure 5. Visualization of object detection results on PASCAL VOC dataset. The red box with orange annotation means the predicted object bounding box with the corresponding prediction. For (b) and (c), we use the SwinIR model with an SR scale factor of x4. (a) ILR \u2192T (b) S \u2192T (c) SR4IR (Ours) (d) Ground-truth Figure 6. Visualization of images and image classification results on CUB-200-2011 dataset. We present the restored images and the corresponding caption. The caption below the image represents the predicted image classification results, which is indicated by a checkmark if the prediction is correct. For (b) and (c), we use the SwinIR model with an SR scale factor of x4. sequent task network to make the correct label predictions, while the other methods fail to do so. Furthermore, it should be noted that the visualized images produced by our SR4IR achieve the most visually pleasing results. 4.3. Ablation studies In the ablation studies, we analyze the effectiveness of each component of our method and further report the comparison with previous related works. We report the results of semantic segmentation on the x8 scale PASCAL VOC dataset and image classification on the x8 scale StanfordCars dataset, using the EDSR-baseline network. Please refer to our supplementary materials for more diverse ablation studies. Effectiveness of the TDP loss. Table 4 shows the task performance of our SR4IR, comparing the cases with and without the proposed TDP loss. Without the TDP loss, the performance of SR4IR decreases significantly, with a decrease of -1.7 mIoU score in semantic segmentation and a decrease in precision of -1.5% in image classification. The Methods Segmentation Classification mIoU\u2191 LPIPS\u2193 Top-1 Acc.\u2191(%) LPIPS\u2193 SR4IR without TDP loss 53.3 0.407 69.9 0.327 SR4IR (Ours) 55.0 0.380 71.4 0.326 Table 4. Effectiveness of the proposed TDP loss. We evaluate the impact of TDP loss by including or excluding LTDP in phase 1 of equation (3) on the performance of SR4IR. results demonstrate that the TDP loss is crucial for a significant boost in task performance in our SR4IR, showing that restoring the task-relevant high-frequency details is essential in enhancing the performance of the subsequent task network. Furthermore, the absence of TDP loss results in a significant deterioration in the LPIPS score with a -7% decrease in segmentation. These results validate that our TDP loss not only enhances task performance but also contributes to generating perceptually realistic SR results. Effectiveness of the CQMix. To verify the effectiveness of the proposed CQMix, we conduct experiments on training the task network with various types of images, as shown in Table 5. When we train the task network with only ISR, 7 the performance is poor, indicating that the effectiveness of the TDP loss decreases when the task network is not trained to properly utilize the high-frequency information present in HR images. However, training the task network solely on IHR also yielded subpar results, as the task network is not trained to handle ISR. Combining both ISR and IHR performs better compared to individual methods. Finally, adding augmented images through our CQMix further enhances the performance, with gains of +0.3 in segmentation and +1.1% in classification, respectively, highlighting the substantial benefits of preventing shortcut learning. Effectiveness of the alternate training framework. To validate the effectiveness of the proposed alternate training framework, we compare our methods with three perceptualloss-related approaches, denoted as (A), (B) and (C). In (A), we substitute the proposed TDP loss to the conventional perceptual loss, which uses a fixed VGG network pre-trained on ImageNet. In (B), rather than using the ontraining network proposed in our SR4IR, we employ TDP loss using the fixed task network pre-trained on the corresponding target task, i.e., IHR \u2192T. In (C), rather than adopting alternate training, we apply the TDP loss with a trainable on-training task network within a joint training, where both the SR and task networks are trained jointly. Table 6 presents performance results between (A), (B), (C), and our method. Compared to the performance of SR4IR without TDP loss, which achieves a 53.3 mIoU score and 69.9% accuracy as presented in Table 4, the method (A) achieves similar or even worse performance, demonstrating that adopting conventional perceptual loss does not help to boost the task performance. The method (B) also results in marginal performance gains, as the fixed and pretrained task network is not trained to handle the SR results. Notably, the method (C) fails to learn because optimizing the TDP loss leads to undesirable local optima when the task network is trainable, which results in forcing the task network to produce a constant output regardless of varied inputs. In contrast, our SR4IR achieves significant performance improvement compared to the above approaches, highlighting the importance of adopting an alternate training framework to fully leverage the power of the TDP loss. Comparison with previous methods. To highlight the advantages of our SR4IR, we include a quantitative comparison with two related previous methods, TDSR [19] and SOD-MTGAN [3], which are task-driven restoration methods. The results are presented in Table 7. Due to the absence of a publicly available training code, we report the results of our re-implemented version, which is elaborated in the supplementary material. In comparison to TDSR and SODMTGAN, our SR4IR achieves significantly superior performance, underlining the effectiveness of leveraging perceptual loss when compared to the case of training without perceptual loss (TDSR) and adversarial loss (SOD-MTGAN). Training image types Segmentation Classification (mIoU\u2191) (Top-1 Acc.%\u2191) ISR 51.7 66.7 IHR 53.3 59.3 ISR + IHR 54.7 70.3 ISR + IHR + Iaug (Ours) 55.0 71.4 Table 5. Effectiveness of the proposed CQMix. We compare the performance of SR4IR by varying the training images Icat in phase 2 of equation (3). Methods Feature extractor for perceptual loss Segmentation Classification (mIoU\u2191) (Top-1 Acc.%\u2191) (A) Pre-trained on ImageNet 53.6 69.4 (B) Pre-trained on corresponding task 53.6 70.3 (C) On-training (joint) 3.5 3.7 Ours On-training (alternate) 55.0 71.4 Table 6. Effectiveness of the proposed alternate training framework. We compare our SR4IR with three perceptual-loss-related approaches, by varying the feature extractor F\u03b8feat used for the TDP loss in equation (2). Methods Segmentation Classification (mIoU\u2191) (Top-1 Acc. %\u2191) TDSR\u2020 [19] 49.8 56.4 SOD-MTGAN\u2020 [3] 51.6 68.1 SR4IR (Ours) 55.0 71.4 Table 7. Comparison with related previous methods. The symbol \u2020 denotes the results of our re-implementation version. 5. Conclusion In this paper, we tackle the practical scenarios for image recognition tasks where low-resolution (LR) content prevails when processing the images. We tackle the challenge by employing image super-resolution (SR) techniques. Considering the ill-posed nature of SR, we focus on developing methods that encourage an SR network to reconstruct task-related features beneficial for a specific task. In particular, we propose task-driven perceptual (TDP) loss that effectively guides the restoration of task-related high-frequency contents. Furthermore, our cross-quality patch mix (CQMix) and alternate training framework help us fully harness the advantages of our TDP loss by preventing a task network from learning biased features and eliminating the domain gap problem. Extensive experimental results demonstrate that our Super-Resolution for Image Recognition (SR4IR) can be generally and effectively applicable to various image recognition tasks, such as semantic segmentation, object detection, and image classification. Acknowledgments. This work was supported in part by the IITP grants [No. 2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No.2021-0-02068, and No.2023-0-00156], the NRF grant [No.2021M3A9E4080782] funded by the Korean government (MSIT). 8"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1803.11316v1",
+                    "title": "Task-Driven Super Resolution: Object Detection in Low-resolution Images",
+                    "abstract": "We consider how image super resolution (SR) can contribute to an object\ndetection task in low-resolution images. Intuitively, SR gives a positive\nimpact on the object detection task. While several previous works demonstrated\nthat this intuition is correct, SR and detector are optimized independently in\nthese works. This paper proposes a novel framework to train a deep neural\nnetwork where the SR sub-network explicitly incorporates a detection loss in\nits training objective, via a tradeoff with a traditional detection loss. This\nend-to-end training procedure allows us to train SR preprocessing for any\ndifferentiable detector. We demonstrate that our task-driven SR consistently\nand significantly improves accuracy of an object detector on low-resolution\nimages for a variety of conditions and scaling factors.",
+                    "authors": "Muhammad Haris, Greg Shakhnarovich, Norimichi Ukita",
+                    "published": "2018-03-30",
+                    "updated": "2018-03-30",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.10257v1",
+                    "title": "SwinIR: Image Restoration Using Swin Transformer",
+                    "abstract": "Image restoration is a long-standing low-level vision problem that aims to\nrestore high-quality images from low-quality images (e.g., downscaled, noisy\nand compressed images). While state-of-the-art image restoration methods are\nbased on convolutional neural networks, few attempts have been made with\nTransformers which show impressive performance on high-level vision tasks. In\nthis paper, we propose a strong baseline model SwinIR for image restoration\nbased on the Swin Transformer. SwinIR consists of three parts: shallow feature\nextraction, deep feature extraction and high-quality image reconstruction. In\nparticular, the deep feature extraction module is composed of several residual\nSwin Transformer blocks (RSTB), each of which has several Swin Transformer\nlayers together with a residual connection. We conduct experiments on three\nrepresentative tasks: image super-resolution (including classical, lightweight\nand real-world image super-resolution), image denoising (including grayscale\nand color image denoising) and JPEG compression artifact reduction.\nExperimental results demonstrate that SwinIR outperforms state-of-the-art\nmethods on different tasks by $\\textbf{up to 0.14$\\sim$0.45dB}$, while the\ntotal number of parameters can be reduced by $\\textbf{up to 67%}$.",
+                    "authors": "Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, Radu Timofte",
+                    "published": "2021-08-23",
+                    "updated": "2021-08-23",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.10833v2",
+                    "title": "Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data",
+                    "abstract": "Though many attempts have been made in blind super-resolution to restore\nlow-resolution images with unknown and complex degradations, they are still far\nfrom addressing general real-world degraded images. In this work, we extend the\npowerful ESRGAN to a practical restoration application (namely, Real-ESRGAN),\nwhich is trained with pure synthetic data. Specifically, a high-order\ndegradation modeling process is introduced to better simulate complex\nreal-world degradations. We also consider the common ringing and overshoot\nartifacts in the synthesis process. In addition, we employ a U-Net\ndiscriminator with spectral normalization to increase discriminator capability\nand stabilize the training dynamics. Extensive comparisons have shown its\nsuperior visual performance than prior works on various real datasets. We also\nprovide efficient implementations to synthesize training pairs on the fly.",
+                    "authors": "Xintao Wang, Liangbin Xie, Chao Dong, Ying Shan",
+                    "published": "2021-07-22",
+                    "updated": "2021-08-17",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.09472v3",
+                    "title": "DiffIR: Efficient Diffusion Model for Image Restoration",
+                    "abstract": "Diffusion model (DM) has achieved SOTA performance by modeling the image\nsynthesis process into a sequential application of a denoising network.\nHowever, different from image synthesis, image restoration (IR) has a strong\nconstraint to generate results in accordance with ground-truth. Thus, for IR,\ntraditional DMs running massive iterations on a large model to estimate whole\nimages or feature maps is inefficient. To address this issue, we propose an\nefficient DM for IR (DiffIR), which consists of a compact IR prior extraction\nnetwork (CPEN), dynamic IR transformer (DIRformer), and denoising network.\nSpecifically, DiffIR has two training stages: pretraining and training DM. In\npretraining, we input ground-truth images into CPEN$_{S1}$ to capture a compact\nIR prior representation (IPR) to guide DIRformer. In the second stage, we train\nthe DM to directly estimate the same IRP as pretrained CPEN$_{S1}$ only using\nLQ images. We observe that since the IPR is only a compact vector, DiffIR can\nuse fewer iterations than traditional DM to obtain accurate estimations and\ngenerate more stable and realistic results. Since the iterations are few, our\nDiffIR can adopt a joint optimization of CPEN$_{S2}$, DIRformer, and denoising\nnetwork, which can further reduce the estimation error influence. We conduct\nextensive experiments on several IR tasks and achieve SOTA performance while\nconsuming less computational costs. Code is available at\n\\url{https://github.com/Zj-BinXia/DiffIR}.",
+                    "authors": "Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, Luc Van Gool",
+                    "published": "2023-03-16",
+                    "updated": "2023-08-16",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1801.03924v2",
+                    "title": "The Unreasonable Effectiveness of Deep Features as a Perceptual Metric",
+                    "abstract": "While it is nearly effortless for humans to quickly assess the perceptual\nsimilarity between two images, the underlying processes are thought to be quite\ncomplex. Despite this, the most widely used perceptual metrics today, such as\nPSNR and SSIM, are simple, shallow functions, and fail to account for many\nnuances of human perception. Recently, the deep learning community has found\nthat features of the VGG network trained on ImageNet classification has been\nremarkably useful as a training loss for image synthesis. But how perceptual\nare these so-called \"perceptual losses\"? What elements are critical for their\nsuccess? To answer these questions, we introduce a new dataset of human\nperceptual similarity judgments. We systematically evaluate deep features\nacross different architectures and tasks and compare them with classic metrics.\nWe find that deep features outperform all previous metrics by large margins on\nour dataset. More surprisingly, this result is not restricted to\nImageNet-trained VGG features, but holds across different deep architectures\nand levels of supervision (supervised, self-supervised, or even unsupervised).\nOur results suggest that perceptual similarity is an emergent property shared\nacross deep visual representations.",
+                    "authors": "Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, Oliver Wang",
+                    "published": "2018-01-11",
+                    "updated": "2018-04-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.GR"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1910.09185v4",
+                    "title": "Exploring Simple and Transferable Recognition-Aware Image Processing",
+                    "abstract": "Recent progress in image recognition has stimulated the deployment of vision\nsystems at an unprecedented scale. As a result, visual data are now often\nconsumed not only by humans but also by machines. Existing image processing\nmethods only optimize for better human perception, yet the resulting images may\nnot be accurately recognized by machines. This can be undesirable, e.g., the\nimages can be improperly handled by search engines or recommendation systems.\nIn this work, we examine simple approaches to improve machine recognition of\nprocessed images: optimizing the recognition loss directly on the image\nprocessing network or through an intermediate input transformation model.\nInterestingly, the processing model's ability to enhance recognition quality\ncan transfer when evaluated on models of different architectures, recognized\ncategories, tasks and training datasets. This makes the methods applicable even\nwhen we do not have the knowledge of future recognition models, e.g., when\nuploading processed images to the Internet. We conduct experiments on multiple\nimage processing tasks paired with ImageNet classification and PASCAL VOC\ndetection as recognition tasks. With these simple yet effective methods,\nsubstantial accuracy gain can be achieved with strong transferability and\nminimal image quality loss. Through a user study we further show that the\naccuracy gain can transfer to a black-box cloud model. Finally, we try to\nexplain this transferability phenomenon by demonstrating the similarities of\ndifferent models' decision boundaries. Code is available at\nhttps://github.com/liuzhuang13/Transferable_RA .",
+                    "authors": "Zhuang Liu, Hung-Ju Wang, Tinghui Zhou, Zhiqiang Shen, Bingyi Kang, Evan Shelhamer, Trevor Darrell",
+                    "published": "2019-10-21",
+                    "updated": "2022-09-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1609.04802v5",
+                    "title": "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network",
+                    "abstract": "Despite the breakthroughs in accuracy and speed of single image\nsuper-resolution using faster and deeper convolutional neural networks, one\ncentral problem remains largely unsolved: how do we recover the finer texture\ndetails when we super-resolve at large upscaling factors? The behavior of\noptimization-based super-resolution methods is principally driven by the choice\nof the objective function. Recent work has largely focused on minimizing the\nmean squared reconstruction error. The resulting estimates have high peak\nsignal-to-noise ratios, but they are often lacking high-frequency details and\nare perceptually unsatisfying in the sense that they fail to match the fidelity\nexpected at the higher resolution. In this paper, we present SRGAN, a\ngenerative adversarial network (GAN) for image super-resolution (SR). To our\nknowledge, it is the first framework capable of inferring photo-realistic\nnatural images for 4x upscaling factors. To achieve this, we propose a\nperceptual loss function which consists of an adversarial loss and a content\nloss. The adversarial loss pushes our solution to the natural image manifold\nusing a discriminator network that is trained to differentiate between the\nsuper-resolved images and original photo-realistic images. In addition, we use\na content loss motivated by perceptual similarity instead of similarity in\npixel space. Our deep residual network is able to recover photo-realistic\ntextures from heavily downsampled images on public benchmarks. An extensive\nmean-opinion-score (MOS) test shows hugely significant gains in perceptual\nquality using SRGAN. The MOS scores obtained with SRGAN are closer to those of\nthe original high-resolution images than to those obtained with any\nstate-of-the-art method.",
+                    "authors": "Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, Wenzhe Shi",
+                    "published": "2016-09-15",
+                    "updated": "2017-05-25",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.07636v2",
+                    "title": "Image Super-Resolution via Iterative Refinement",
+                    "abstract": "We present SR3, an approach to image Super-Resolution via Repeated\nRefinement. SR3 adapts denoising diffusion probabilistic models to conditional\nimage generation and performs super-resolution through a stochastic denoising\nprocess. Inference starts with pure Gaussian noise and iteratively refines the\nnoisy output using a U-Net model trained on denoising at various noise levels.\nSR3 exhibits strong performance on super-resolution tasks at different\nmagnification factors, on faces and natural images. We conduct human evaluation\non a standard 8X face super-resolution task on CelebA-HQ, comparing with SOTA\nGAN methods. SR3 achieves a fool rate close to 50%, suggesting photo-realistic\noutputs, while GANs do not exceed a fool rate of 34%. We further show the\neffectiveness of SR3 in cascaded image generation, where generative models are\nchained with super-resolution models, yielding a competitive FID score of 11.3\non ImageNet.",
+                    "authors": "Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J. Fleet, Mohammad Norouzi",
+                    "published": "2021-04-15",
+                    "updated": "2021-06-30",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2008.08767v1",
+                    "title": "Single Image Super-Resolution via a Holistic Attention Network",
+                    "abstract": "Informative features play a crucial role in the single image super-resolution\ntask. Channel attention has been demonstrated to be effective for preserving\ninformation-rich features in each layer. However, channel attention treats each\nconvolution layer as a separate process that misses the correlation among\ndifferent layers. To address this problem, we propose a new holistic attention\nnetwork (HAN), which consists of a layer attention module (LAM) and a\nchannel-spatial attention module (CSAM), to model the holistic\ninterdependencies among layers, channels, and positions. Specifically, the\nproposed LAM adaptively emphasizes hierarchical features by considering\ncorrelations among layers. Meanwhile, CSAM learns the confidence at all the\npositions of each channel to selectively capture more informative features.\nExtensive experiments demonstrate that the proposed HAN performs favorably\nagainst the state-of-the-art single image super-resolution approaches.",
+                    "authors": "Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, Haifeng Shen",
+                    "published": "2020-08-20",
+                    "updated": "2020-08-20",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2401.13627v2",
+                    "title": "Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild",
+                    "abstract": "We introduce SUPIR (Scaling-UP Image Restoration), a groundbreaking image\nrestoration method that harnesses generative prior and the power of model\nscaling up. Leveraging multi-modal techniques and advanced generative prior,\nSUPIR marks a significant advance in intelligent and realistic image\nrestoration. As a pivotal catalyst within SUPIR, model scaling dramatically\nenhances its capabilities and demonstrates new potential for image restoration.\nWe collect a dataset comprising 20 million high-resolution, high-quality images\nfor model training, each enriched with descriptive text annotations. SUPIR\nprovides the capability to restore images guided by textual prompts, broadening\nits application scope and potential. Moreover, we introduce negative-quality\nprompts to further improve perceptual quality. We also develop a\nrestoration-guided sampling method to suppress the fidelity issue encountered\nin generative-based restoration. Experiments demonstrate SUPIR's exceptional\nrestoration effects and its novel capacity to manipulate restoration through\ntextual prompts.",
+                    "authors": "Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, Chao Dong",
+                    "published": "2024-01-24",
+                    "updated": "2024-04-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1612.07919v2",
+                    "title": "EnhanceNet: Single Image Super-Resolution Through Automated Texture Synthesis",
+                    "abstract": "Single image super-resolution is the task of inferring a high-resolution\nimage from a single low-resolution input. Traditionally, the performance of\nalgorithms for this task is measured using pixel-wise reconstruction measures\nsuch as peak signal-to-noise ratio (PSNR) which have been shown to correlate\npoorly with the human perception of image quality. As a result, algorithms\nminimizing these metrics tend to produce over-smoothed images that lack\nhigh-frequency textures and do not look natural despite yielding high PSNR\nvalues.\n  We propose a novel application of automated texture synthesis in combination\nwith a perceptual loss focusing on creating realistic textures rather than\noptimizing for a pixel-accurate reproduction of ground truth images during\ntraining. By using feed-forward fully convolutional neural networks in an\nadversarial training setting, we achieve a significant boost in image quality\nat high magnification ratios. Extensive experiments on a number of datasets\nshow the effectiveness of our approach, yielding state-of-the-art results in\nboth quantitative and qualitative benchmarks.",
+                    "authors": "Mehdi S. M. Sajjadi, Bernhard Sch\u00f6lkopf, Michael Hirsch",
+                    "published": "2016-12-23",
+                    "updated": "2017-07-30",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1603.08155v1",
+                    "title": "Perceptual Losses for Real-Time Style Transfer and Super-Resolution",
+                    "abstract": "We consider image transformation problems, where an input image is\ntransformed into an output image. Recent methods for such problems typically\ntrain feed-forward convolutional neural networks using a \\emph{per-pixel} loss\nbetween the output and ground-truth images. Parallel work has shown that\nhigh-quality images can be generated by defining and optimizing\n\\emph{perceptual} loss functions based on high-level features extracted from\npretrained networks. We combine the benefits of both approaches, and propose\nthe use of perceptual loss functions for training feed-forward networks for\nimage transformation tasks. We show results on image style transfer, where a\nfeed-forward network is trained to solve the optimization problem proposed by\nGatys et al in real-time. Compared to the optimization-based method, our\nnetwork gives similar qualitative results but is three orders of magnitude\nfaster. We also experiment with single-image super-resolution, where replacing\na per-pixel loss with a perceptual loss gives visually pleasing results.",
+                    "authors": "Justin Johnson, Alexandre Alahi, Li Fei-Fei",
+                    "published": "2016-03-27",
+                    "updated": "2016-03-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1809.00219v2",
+                    "title": "ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks",
+                    "abstract": "The Super-Resolution Generative Adversarial Network (SRGAN) is a seminal work\nthat is capable of generating realistic textures during single image\nsuper-resolution. However, the hallucinated details are often accompanied with\nunpleasant artifacts. To further enhance the visual quality, we thoroughly\nstudy three key components of SRGAN - network architecture, adversarial loss\nand perceptual loss, and improve each of them to derive an Enhanced SRGAN\n(ESRGAN). In particular, we introduce the Residual-in-Residual Dense Block\n(RRDB) without batch normalization as the basic network building unit.\nMoreover, we borrow the idea from relativistic GAN to let the discriminator\npredict relative realness instead of the absolute value. Finally, we improve\nthe perceptual loss by using the features before activation, which could\nprovide stronger supervision for brightness consistency and texture recovery.\nBenefiting from these improvements, the proposed ESRGAN achieves consistently\nbetter visual quality with more realistic and natural textures than SRGAN and\nwon the first place in the PIRM2018-SR Challenge. The code is available at\nhttps://github.com/xinntao/ESRGAN .",
+                    "authors": "Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Chen Change Loy, Yu Qiao, Xiaoou Tang",
+                    "published": "2018-09-01",
+                    "updated": "2018-09-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.02938v2",
+                    "title": "ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models",
+                    "abstract": "Denoising diffusion probabilistic models (DDPM) have shown remarkable\nperformance in unconditional image generation. However, due to the\nstochasticity of the generative process in DDPM, it is challenging to generate\nimages with the desired semantics. In this work, we propose Iterative Latent\nVariable Refinement (ILVR), a method to guide the generative process in DDPM to\ngenerate high-quality images based on a given reference image. Here, the\nrefinement of the generative process in DDPM enables a single DDPM to sample\nimages from various sets directed by the reference image. The proposed ILVR\nmethod generates high-quality images while controlling the generation. The\ncontrollability of our method allows adaptation of a single DDPM without any\nadditional learning in various image generation tasks, such as generation from\nvarious downsampling factors, multi-domain image translation, paint-to-image,\nand editing with scribbles.",
+                    "authors": "Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, Sungroh Yoon",
+                    "published": "2021-08-06",
+                    "updated": "2021-09-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1511.04587v2",
+                    "title": "Accurate Image Super-Resolution Using Very Deep Convolutional Networks",
+                    "abstract": "We present a highly accurate single-image super-resolution (SR) method. Our\nmethod uses a very deep convolutional network inspired by VGG-net used for\nImageNet classification \\cite{simonyan2015very}. We find increasing our network\ndepth shows a significant improvement in accuracy. Our final model uses 20\nweight layers. By cascading small filters many times in a deep network\nstructure, contextual information over large image regions is exploited in an\nefficient way. With very deep networks, however, convergence speed becomes a\ncritical issue during training. We propose a simple yet effective training\nprocedure. We learn residuals only and use extremely high learning rates\n($10^4$ times higher than SRCNN \\cite{dong2015image}) enabled by adjustable\ngradient clipping. Our proposed method performs better than existing methods in\naccuracy and visual improvements in our results are easily noticeable.",
+                    "authors": "Jiwon Kim, Jung Kwon Lee, Kyoung Mu Lee",
+                    "published": "2015-11-14",
+                    "updated": "2016-11-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.02758v2",
+                    "title": "Image Super-Resolution Using Very Deep Residual Channel Attention Networks",
+                    "abstract": "Convolutional neural network (CNN) depth is of crucial importance for image\nsuper-resolution (SR). However, we observe that deeper networks for image SR\nare more difficult to train. The low-resolution inputs and features contain\nabundant low-frequency information, which is treated equally across channels,\nhence hindering the representational ability of CNNs. To solve these problems,\nwe propose the very deep residual channel attention networks (RCAN).\nSpecifically, we propose a residual in residual (RIR) structure to form very\ndeep network, which consists of several residual groups with long skip\nconnections. Each residual group contains some residual blocks with short skip\nconnections. Meanwhile, RIR allows abundant low-frequency information to be\nbypassed through multiple skip connections, making the main network focus on\nlearning high-frequency information. Furthermore, we propose a channel\nattention mechanism to adaptively rescale channel-wise features by considering\ninterdependencies among channels. Extensive experiments show that our RCAN\nachieves better accuracy and visual improvements against state-of-the-art\nmethods.",
+                    "authors": "Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, Yun Fu",
+                    "published": "2018-07-08",
+                    "updated": "2018-07-12",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2006.11239v2",
+                    "title": "Denoising Diffusion Probabilistic Models",
+                    "abstract": "We present high quality image synthesis results using diffusion probabilistic\nmodels, a class of latent variable models inspired by considerations from\nnonequilibrium thermodynamics. Our best results are obtained by training on a\nweighted variational bound designed according to a novel connection between\ndiffusion probabilistic models and denoising score matching with Langevin\ndynamics, and our models naturally admit a progressive lossy decompression\nscheme that can be interpreted as a generalization of autoregressive decoding.\nOn the unconditional CIFAR10 dataset, we obtain an Inception score of 9.46 and\na state-of-the-art FID score of 3.17. On 256x256 LSUN, we obtain sample quality\nsimilar to ProgressiveGAN. Our implementation is available at\nhttps://github.com/hojonathanho/diffusion",
+                    "authors": "Jonathan Ho, Ajay Jain, Pieter Abbeel",
+                    "published": "2020-06-19",
+                    "updated": "2020-12-16",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "stat.ML"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1707.02921v1",
+                    "title": "Enhanced Deep Residual Networks for Single Image Super-Resolution",
+                    "abstract": "Recent research on super-resolution has progressed with the development of\ndeep convolutional neural networks (DCNN). In particular, residual learning\ntechniques exhibit improved performance. In this paper, we develop an enhanced\ndeep super-resolution network (EDSR) with performance exceeding those of\ncurrent state-of-the-art SR methods. The significant performance improvement of\nour model is due to optimization by removing unnecessary modules in\nconventional residual networks. The performance is further improved by\nexpanding the model size while we stabilize the training procedure. We also\npropose a new multi-scale deep super-resolution system (MDSR) and training\nmethod, which can reconstruct high-resolution images of different upscaling\nfactors in a single model. The proposed methods show superior performance over\nthe state-of-the-art methods on benchmark datasets and prove its excellence by\nwinning the NTIRE2017 Super-Resolution Challenge.",
+                    "authors": "Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, Kyoung Mu Lee",
+                    "published": "2017-07-10",
+                    "updated": "2017-07-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1501.00092v3",
+                    "title": "Image Super-Resolution Using Deep Convolutional Networks",
+                    "abstract": "We propose a deep learning method for single image super-resolution (SR). Our\nmethod directly learns an end-to-end mapping between the low/high-resolution\nimages. The mapping is represented as a deep convolutional neural network (CNN)\nthat takes the low-resolution image as the input and outputs the\nhigh-resolution one. We further show that traditional sparse-coding-based SR\nmethods can also be viewed as a deep convolutional network. But unlike\ntraditional methods that handle each component separately, our method jointly\noptimizes all layers. Our deep CNN has a lightweight structure, yet\ndemonstrates state-of-the-art restoration quality, and achieves fast speed for\npractical on-line usage. We explore different network structures and parameter\nsettings to achieve trade-offs between performance and speed. Moreover, we\nextend our network to cope with three color channels simultaneously, and show\nbetter overall reconstruction quality.",
+                    "authors": "Chao Dong, Chen Change Loy, Kaiming He, Xiaoou Tang",
+                    "published": "2014-12-31",
+                    "updated": "2015-07-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.NE",
+                        "I.4.5; I.2.6"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2301.11699v3",
+                    "title": "Image Restoration with Mean-Reverting Stochastic Differential Equations",
+                    "abstract": "This paper presents a stochastic differential equation (SDE) approach for\ngeneral-purpose image restoration. The key construction consists in a\nmean-reverting SDE that transforms a high-quality image into a degraded\ncounterpart as a mean state with fixed Gaussian noise. Then, by simulating the\ncorresponding reverse-time SDE, we are able to restore the origin of the\nlow-quality image without relying on any task-specific prior knowledge.\nCrucially, the proposed mean-reverting SDE has a closed-form solution, allowing\nus to compute the ground truth time-dependent score and learn it with a neural\nnetwork. Moreover, we propose a maximum likelihood objective to learn an\noptimal reverse trajectory that stabilizes the training and improves the\nrestoration results. The experiments show that our proposed method achieves\nhighly competitive performance in quantitative comparisons on image deraining,\ndeblurring, and denoising, setting a new state-of-the-art on two deraining\ndatasets. Finally, the general applicability of our approach is further\ndemonstrated via qualitative results on image super-resolution, inpainting, and\ndehazing. Code is available at\nhttps://github.com/Algolzw/image-restoration-sde.",
+                    "authors": "Ziwei Luo, Fredrik K. Gustafsson, Zheng Zhao, Jens Sj\u00f6lund, Thomas B. Sch\u00f6n",
+                    "published": "2023-01-27",
+                    "updated": "2023-05-31",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2109.03444v1",
+                    "title": "Toward Real-World Super-Resolution via Adaptive Downsampling Models",
+                    "abstract": "Most image super-resolution (SR) methods are developed on synthetic\nlow-resolution (LR) and high-resolution (HR) image pairs that are constructed\nby a predetermined operation, e.g., bicubic downsampling. As existing methods\ntypically learn an inverse mapping of the specific function, they produce\nblurry results when applied to real-world images whose exact formulation is\ndifferent and unknown. Therefore, several methods attempt to synthesize much\nmore diverse LR samples or learn a realistic downsampling model. However, due\nto restrictive assumptions on the downsampling process, they are still biased\nand less generalizable. This study proposes a novel method to simulate an\nunknown downsampling process without imposing restrictive prior knowledge. We\npropose a generalizable low-frequency loss (LFL) in the adversarial training\nframework to imitate the distribution of target LR images without using any\npaired examples. Furthermore, we design an adaptive data loss (ADL) for the\ndownsampler, which can be adaptively learned and updated from the data during\nthe training loops. Extensive experiments validate that our downsampling model\ncan facilitate existing SR methods to perform more accurate reconstructions on\nvarious synthetic and real-world examples than the conventional approaches.",
+                    "authors": "Sanghyun Son, Jaeha Kim, Wei-Sheng Lai, Ming-Husan Yang, Kyoung Mu Lee",
+                    "published": "2021-09-08",
+                    "updated": "2021-09-08",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2201.11793v3",
+                    "title": "Denoising Diffusion Restoration Models",
+                    "abstract": "Many interesting tasks in image restoration can be cast as linear inverse\nproblems. A recent family of approaches for solving these problems uses\nstochastic algorithms that sample from the posterior distribution of natural\nimages given the measurements. However, efficient solutions often require\nproblem-specific supervised training to model the posterior, whereas\nunsupervised methods that are not problem-specific typically rely on\ninefficient iterative methods. This work addresses these issues by introducing\nDenoising Diffusion Restoration Models (DDRM), an efficient, unsupervised\nposterior sampling method. Motivated by variational inference, DDRM takes\nadvantage of a pre-trained denoising diffusion generative model for solving any\nlinear inverse problem. We demonstrate DDRM's versatility on several image\ndatasets for super-resolution, deblurring, inpainting, and colorization under\nvarious amounts of measurement noise. DDRM outperforms the current leading\nunsupervised methods on the diverse ImageNet dataset in reconstruction quality,\nperceptual quality, and runtime, being 5x faster than the nearest competitor.\nDDRM also generalizes well for natural images out of the distribution of the\nobserved ImageNet training set.",
+                    "authors": "Bahjat Kawar, Michael Elad, Stefano Ermon, Jiaming Song",
+                    "published": "2022-01-27",
+                    "updated": "2022-10-12",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2106.03106v2",
+                    "title": "Uformer: A General U-Shaped Transformer for Image Restoration",
+                    "abstract": "In this paper, we present Uformer, an effective and efficient\nTransformer-based architecture for image restoration, in which we build a\nhierarchical encoder-decoder network using the Transformer block. In Uformer,\nthere are two core designs. First, we introduce a novel locally-enhanced window\n(LeWin) Transformer block, which performs nonoverlapping window-based\nself-attention instead of global self-attention. It significantly reduces the\ncomputational complexity on high resolution feature map while capturing local\ncontext. Second, we propose a learnable multi-scale restoration modulator in\nthe form of a multi-scale spatial bias to adjust features in multiple layers of\nthe Uformer decoder. Our modulator demonstrates superior capability for\nrestoring details for various image restoration tasks while introducing\nmarginal extra parameters and computational cost. Powered by these two designs,\nUformer enjoys a high capability for capturing both local and global\ndependencies for image restoration. To evaluate our approach, extensive\nexperiments are conducted on several image restoration tasks, including image\ndenoising, motion deblurring, defocus deblurring and deraining. Without bells\nand whistles, our Uformer achieves superior or comparable performance compared\nwith the state-of-the-art algorithms. The code and models are available at\nhttps://github.com/ZhendongWang6/Uformer.",
+                    "authors": "Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, Houqiang Li",
+                    "published": "2021-06-06",
+                    "updated": "2021-11-25",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1803.02735v1",
+                    "title": "Deep Back-Projection Networks For Super-Resolution",
+                    "abstract": "The feed-forward architectures of recently proposed deep super-resolution\nnetworks learn representations of low-resolution inputs, and the non-linear\nmapping from those to high-resolution output. However, this approach does not\nfully address the mutual dependencies of low- and high-resolution images. We\npropose Deep Back-Projection Networks (DBPN), that exploit iterative up- and\ndown-sampling layers, providing an error feedback mechanism for projection\nerrors at each stage. We construct mutually-connected up- and down-sampling\nstages each of which represents different types of image degradation and\nhigh-resolution components. We show that extending this idea to allow\nconcatenation of features across up- and down-sampling stages (Dense DBPN)\nallows us to reconstruct further improve super-resolution, yielding superior\nresults and in particular establishing new state of the art results for large\nscaling factors such as 8x across multiple data sets.",
+                    "authors": "Muhammad Haris, Greg Shakhnarovich, Norimichi Ukita",
+                    "published": "2018-03-07",
+                    "updated": "2018-03-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2111.09881v2",
+                    "title": "Restormer: Efficient Transformer for High-Resolution Image Restoration",
+                    "abstract": "Since convolutional neural networks (CNNs) perform well at learning\ngeneralizable image priors from large-scale data, these models have been\nextensively applied to image restoration and related tasks. Recently, another\nclass of neural architectures, Transformers, have shown significant performance\ngains on natural language and high-level vision tasks. While the Transformer\nmodel mitigates the shortcomings of CNNs (i.e., limited receptive field and\ninadaptability to input content), its computational complexity grows\nquadratically with the spatial resolution, therefore making it infeasible to\napply to most image restoration tasks involving high-resolution images. In this\nwork, we propose an efficient Transformer model by making several key designs\nin the building blocks (multi-head attention and feed-forward network) such\nthat it can capture long-range pixel interactions, while still remaining\napplicable to large images. Our model, named Restoration Transformer\n(Restormer), achieves state-of-the-art results on several image restoration\ntasks, including image deraining, single-image motion deblurring, defocus\ndeblurring (single-image and dual-pixel data), and image denoising (Gaussian\ngrayscale/color denoising, and real image denoising). The source code and\npre-trained models are available at https://github.com/swz30/Restormer.",
+                    "authors": "Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang",
+                    "published": "2021-11-18",
+                    "updated": "2022-03-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1802.08797v2",
+                    "title": "Residual Dense Network for Image Super-Resolution",
+                    "abstract": "A very deep convolutional neural network (CNN) has recently achieved great\nsuccess for image super-resolution (SR) and offered hierarchical features as\nwell. However, most deep CNN based SR models do not make full use of the\nhierarchical features from the original low-resolution (LR) images, thereby\nachieving relatively-low performance. In this paper, we propose a novel\nresidual dense network (RDN) to address this problem in image SR. We fully\nexploit the hierarchical features from all the convolutional layers.\nSpecifically, we propose residual dense block (RDB) to extract abundant local\nfeatures via dense connected convolutional layers. RDB further allows direct\nconnections from the state of preceding RDB to all the layers of current RDB,\nleading to a contiguous memory (CM) mechanism. Local feature fusion in RDB is\nthen used to adaptively learn more effective features from preceding and\ncurrent local features and stabilizes the training of wider network. After\nfully obtaining dense local features, we use global feature fusion to jointly\nand adaptively learn global hierarchical features in a holistic way. Extensive\nexperiments on benchmark datasets with different degradation models show that\nour RDN achieves favorable performance against state-of-the-art methods.",
+                    "authors": "Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, Yun Fu",
+                    "published": "2018-02-24",
+                    "updated": "2018-03-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2203.09195v1",
+                    "title": "Details or Artifacts: A Locally Discriminative Learning Approach to Realistic Image Super-Resolution",
+                    "abstract": "Single image super-resolution (SISR) with generative adversarial networks\n(GAN) has recently attracted increasing attention due to its potentials to\ngenerate rich details. However, the training of GAN is unstable, and it often\nintroduces many perceptually unpleasant artifacts along with the generated\ndetails. In this paper, we demonstrate that it is possible to train a GAN-based\nSISR model which can stably generate perceptually realistic details while\ninhibiting visual artifacts. Based on the observation that the local statistics\n(e.g., residual variance) of artifact areas are often different from the areas\nof perceptually friendly details, we develop a framework to discriminate\nbetween GAN-generated artifacts and realistic details, and consequently\ngenerate an artifact map to regularize and stabilize the model training\nprocess. Our proposed locally discriminative learning (LDL) method is simple\nyet effective, which can be easily plugged in off-the-shelf SISR methods and\nboost their performance. Experiments demonstrate that LDL outperforms the\nstate-of-the-art GAN based SISR methods, achieving not only higher\nreconstruction accuracy but also superior perceptual quality on both synthetic\nand real-world datasets. Codes and models are available at\nhttps://github.com/csjliang/LDL.",
+                    "authors": "Jie Liang, Hui Zeng, Lei Zhang",
+                    "published": "2022-03-17",
+                    "updated": "2022-03-17",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2208.03062v1",
+                    "title": "Exploring Resolution and Degradation Clues as Self-supervised Signal for Low Quality Object Detection",
+                    "abstract": "Image restoration algorithms such as super resolution (SR) are indispensable\npre-processing modules for object detection in low quality images. Most of\nthese algorithms assume the degradation is fixed and known a priori. However,\nin practical, either the real degradation or optimal up-sampling ratio rate is\nunknown or differs from assumption, leading to a deteriorating performance for\nboth the pre-processing module and the consequent high-level task such as\nobject detection. Here, we propose a novel self-supervised framework to detect\nobjects in degraded low resolution images. We utilizes the downsampling\ndegradation as a kind of transformation for self-supervised signals to explore\nthe equivariant representation against various resolutions and other\ndegradation conditions. The Auto Encoding Resolution in Self-supervision\n(AERIS) framework could further take the advantage of advanced SR architectures\nwith an arbitrary resolution restoring decoder to reconstruct the original\ncorrespondence from the degraded input image. Both the representation learning\nand object detection are optimized jointly in an end-to-end training fashion.\nThe generic AERIS framework could be implemented on various mainstream object\ndetection architectures with different backbones. The extensive experiments\nshow that our methods has achieved superior performance compared with existing\nmethods when facing variant degradation situations. Code would be released at\nhttps://github.com/cuiziteng/ECCV_AERIS.",
+                    "authors": "Ziteng Cui, Yingying Zhu, Lin Gu, Guo-Jun Qi, Xiaoxiao Li, Renrui Zhang, Zenghui Zhang, Tatsuya Harada",
+                    "published": "2022-08-05",
+                    "updated": "2022-08-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2212.00490v2",
+                    "title": "Zero-Shot Image Restoration Using Denoising Diffusion Null-Space Model",
+                    "abstract": "Most existing Image Restoration (IR) models are task-specific, which can not\nbe generalized to different degradation operators. In this work, we propose the\nDenoising Diffusion Null-Space Model (DDNM), a novel zero-shot framework for\narbitrary linear IR problems, including but not limited to image\nsuper-resolution, colorization, inpainting, compressed sensing, and deblurring.\nDDNM only needs a pre-trained off-the-shelf diffusion model as the generative\nprior, without any extra training or network modifications. By refining only\nthe null-space contents during the reverse diffusion process, we can yield\ndiverse results satisfying both data consistency and realness. We further\npropose an enhanced and robust version, dubbed DDNM+, to support noisy\nrestoration and improve restoration quality for hard tasks. Our experiments on\nseveral IR tasks reveal that DDNM outperforms other state-of-the-art zero-shot\nIR methods. We also demonstrate that DDNM+ can solve complex real-world\napplications, e.g., old photo restoration.",
+                    "authors": "Yinhuai Wang, Jiwen Yu, Jian Zhang",
+                    "published": "2022-12-01",
+                    "updated": "2022-12-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2010.11929v2",
+                    "title": "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale",
+                    "abstract": "While the Transformer architecture has become the de-facto standard for\nnatural language processing tasks, its applications to computer vision remain\nlimited. In vision, attention is either applied in conjunction with\nconvolutional networks, or used to replace certain components of convolutional\nnetworks while keeping their overall structure in place. We show that this\nreliance on CNNs is not necessary and a pure transformer applied directly to\nsequences of image patches can perform very well on image classification tasks.\nWhen pre-trained on large amounts of data and transferred to multiple mid-sized\nor small image recognition benchmarks (ImageNet, CIFAR-100, VTAB, etc.), Vision\nTransformer (ViT) attains excellent results compared to state-of-the-art\nconvolutional networks while requiring substantially fewer computational\nresources to train.",
+                    "authors": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby",
+                    "published": "2020-10-22",
+                    "updated": "2021-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1706.04284v3",
+                    "title": "When Image Denoising Meets High-Level Vision Tasks: A Deep Learning Approach",
+                    "abstract": "Conventionally, image denoising and high-level vision tasks are handled\nseparately in computer vision. In this paper, we cope with the two jointly and\nexplore the mutual influence between them. First we propose a convolutional\nneural network for image denoising which achieves the state-of-the-art\nperformance. Second we propose a deep neural network solution that cascades two\nmodules for image denoising and various high-level tasks, respectively, and use\nthe joint loss for updating only the denoising network via back-propagation. We\ndemonstrate that on one hand, the proposed denoiser has the generality to\novercome the performance degradation of different high-level vision tasks. On\nthe other hand, with the guidance of high-level vision information, the\ndenoising network can generate more visually appealing results. To the best of\nour knowledge, this is the first work investigating the benefit of exploiting\nimage semantics simultaneously for image denoising and high-level vision tasks\nvia deep learning. The code is available online\nhttps://github.com/Ding-Liu/DeepDenoising.",
+                    "authors": "Ding Liu, Bihan Wen, Xianming Liu, Zhangyang Wang, Thomas S. Huang",
+                    "published": "2017-06-14",
+                    "updated": "2018-04-16",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1812.04098v3",
+                    "title": "The Effects of Super-Resolution on Object Detection Performance in Satellite Imagery",
+                    "abstract": "We explore the application of super-resolution techniques to satellite\nimagery, and the effects of these techniques on object detection algorithm\nperformance. Specifically, we enhance satellite imagery beyond its native\nresolution, and test if we can identify various types of vehicles, planes, and\nboats with greater accuracy than native resolution. Using the Very Deep\nSuper-Resolution (VDSR) framework and a custom Random Forest Super-Resolution\n(RFSR) framework we generate enhancement levels of 2x, 4x, and 8x over five\ndistinct resolutions ranging from 30 cm to 4.8 meters. Using both native and\nsuper-resolved data, we then train several custom detection models using the\nSIMRDWN object detection framework. SIMRDWN combines a number of popular object\ndetection algorithms (e.g. SSD, YOLO) into a unified framework that is designed\nto rapidly detect objects in large satellite images. This approach allows us to\nquantify the effects of super-resolution techniques on object detection\nperformance across multiple classes and resolutions. We also quantify the\nperformance of object detection as a function of native resolution and object\npixel size. For our test set we note that performance degrades from mean\naverage precision (mAP) = 0.53 at 30 cm resolution, down to mAP = 0.11 at 4.8 m\nresolution. Super-resolving native 30 cm imagery to 15 cm yields the greatest\nbenefit; a 13-36% improvement in mAP. Super-resolution is less beneficial at\ncoarser resolutions, though still provides a small improvement in performance.",
+                    "authors": "Jacob Shermeyer, Adam Van Etten",
+                    "published": "2018-12-10",
+                    "updated": "2019-04-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.03624v1",
+                    "title": "Natural and Realistic Single Image Super-Resolution with Explicit Natural Manifold Discrimination",
+                    "abstract": "Recently, many convolutional neural networks for single image\nsuper-resolution (SISR) have been proposed, which focus on reconstructing the\nhigh-resolution images in terms of objective distortion measures. However, the\nnetworks trained with objective loss functions generally fail to reconstruct\nthe realistic fine textures and details that are essential for better\nperceptual quality. Recovering the realistic details remains a challenging\nproblem, and only a few works have been proposed which aim at increasing the\nperceptual quality by generating enhanced textures. However, the generated fake\ndetails often make undesirable artifacts and the overall image looks somewhat\nunnatural. Therefore, in this paper, we present a new approach to\nreconstructing realistic super-resolved images with high perceptual quality,\nwhile maintaining the naturalness of the result. In particular, we focus on the\ndomain prior properties of SISR problem. Specifically, we define the\nnaturalness prior in the low-level domain and constrain the output image in the\nnatural manifold, which eventually generates more natural and realistic images.\nOur results show better naturalness compared to the recent super-resolution\nalgorithms including perception-oriented ones.",
+                    "authors": "Jae Woong Soh, Gu Yong Park, Junho Jo, Nam Ik Cho",
+                    "published": "2019-11-09",
+                    "updated": "2019-11-09",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "label": "Related Work"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2004.10126v3",
+                    "title": "Generative Synthetic Augmentation using Label-to-Image Translation for Nuclei Image Segmentation",
+                    "abstract": "In medical image diagnosis, pathology image analysis using semantic\nsegmentation becomes important for efficient screening as a field of digital\npathology. The spatial augmentation is ordinary used for semantic segmentation.\nTumor images under malignant are rare and to annotate the labels of nuclei\nregion takes much time-consuming. We require an effective use of dataset to\nmaximize the segmentation accuracy. It is expected that some augmentation to\ntransform generalized images influence the segmentation performance. We propose\na synthetic augmentation using label-to-image translation, mapping from a\nsemantic label with the edge structure to a real image. Exactly this paper deal\nwith stain slides of nuclei in tumor. Actually, we demonstrate several\nsegmentation algorithms applied to the initial dataset that contains real\nimages and labels using synthetic augmentation in order to add their\ngeneralized images. We computes and reports that a proposed synthetic\naugmentation procedure improve their accuracy.",
+                    "authors": "Takato Yasuno",
+                    "published": "2020-04-21",
+                    "updated": "2021-03-02",
+                    "primary_cat": "cs.LG",
+                    "cats": [
+                        "cs.LG",
+                        "eess.IV",
+                        "stat.ML",
+                        "I.4.6; I.2.6"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1606.01481v1",
+                    "title": "Better Image Segmentation by Exploiting Dense Semantic Predictions",
+                    "abstract": "It is well accepted that image segmentation can benefit from utilizing\nmultilevel cues. The paper focuses on utilizing the FCNN-based dense semantic\npredictions in the bottom-up image segmentation, arguing to take semantic cues\ninto account from the very beginning. By this we can avoid merging regions of\nsimilar appearance but distinct semantic categories as possible. The semantic\ninefficiency problem is handled. We also propose a straightforward way to use\nthe contour cues to suppress the noise in multilevel cues, thus to improve the\nsegmentation robustness. The evaluation on the BSDS500 shows that we obtain the\ncompetitive region and boundary performance. Furthermore, since all individual\nregions can be assigned with appropriate semantic labels during the\ncomputation, we are capable of extracting the adjusted semantic segmentations.\nThe experiment on Pascal VOC 2012 shows our improvement to the original\nsemantic segmentations which derives directly from the dense predictions.",
+                    "authors": "Qiyang Zhao, Lewis D Griffin",
+                    "published": "2016-06-05",
+                    "updated": "2016-06-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1509.02441v1",
+                    "title": "Semantic Video Segmentation : Exploring Inference Efficiency",
+                    "abstract": "We explore the efficiency of the CRF inference beyond image level semantic\nsegmentation and perform joint inference in video frames. The key idea is to\ncombine best of two worlds: semantic co-labeling and more expressive models.\nOur formulation enables us to perform inference over ten thousand images within\nseconds and makes the system amenable to perform video semantic segmentation\nmost effectively. On CamVid dataset, with TextonBoost unaries, our proposed\nmethod achieves up to 8% improvement in accuracy over individual semantic image\nsegmentation without additional time overhead. The source code is available at\nhttps://github.com/subtri/video_inference",
+                    "authors": "Subarna Tripathi, Serge Belongie, Youngbae Hwang, Truong Nguyen",
+                    "published": "2015-09-04",
+                    "updated": "2015-09-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2401.07589v1",
+                    "title": "Semantic Scene Segmentation for Robotics",
+                    "abstract": "Comprehensive scene understanding is a critical enabler of robot autonomy.\nSemantic segmentation is one of the key scene understanding tasks which is\npivotal for several robotics applications including autonomous driving,\ndomestic service robotics, last mile delivery, amongst many others. Semantic\nsegmentation is a dense prediction task that aims to provide a scene\nrepresentation in which each pixel of an image is assigned a semantic class\nlabel. Therefore, semantic segmentation considers the full scene context,\nincorporating the object category, location, and shape of all the scene\nelements, including the background. Numerous algorithms have been proposed for\nsemantic segmentation over the years. However, the recent advances in deep\nlearning combined with the boost in the computational capacity and the\navailability of large-scale labeled datasets have led to significant advances\nin semantic segmentation. In this chapter, we introduce the task of semantic\nsegmentation and present the deep learning techniques that have been proposed\nto address this task over the years. We first define the task of semantic\nsegmentation and contrast it with other closely related scene understanding\nproblems. We detail different algorithms and architectures for semantic\nsegmentation and the commonly employed loss functions. Furthermore, we present\nan overview of datasets, benchmarks, and metrics that are used in semantic\nsegmentation. We conclude the chapter with a discussion of challenges and\nopportunities for further research in this area.",
+                    "authors": "Juana Valeria Hurtado, Abhinav Valada",
+                    "published": "2024-01-15",
+                    "updated": "2024-01-15",
+                    "primary_cat": "cs.RO",
+                    "cats": [
+                        "cs.RO"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2201.05869v2",
+                    "title": "Prototype Guided Network for Anomaly Segmentation",
+                    "abstract": "Semantic segmentation methods can not directly identify abnormal objects in\nimages. Anomaly Segmentation algorithm from this realistic setting can\ndistinguish between in-distribution objects and Out-Of-Distribution (OOD)\nobjects and output the anomaly probability for pixels. In this paper, a\nPrototype Guided Anomaly segmentation Network (PGAN) is proposed to extract\nsemantic prototypes for in-distribution training data from limited annotated\nimages. In the model, prototypes are used to model the hierarchical category\nsemantic information and distinguish OOD pixels. The proposed PGAN model\nincludes a semantic segmentation network and a prototype extraction network.\nSimilarity measures are adopted to optimize the prototypes. The learned\nsemantic prototypes are used as category semantics to compare the similarity\nwith features extracted from test images and then to generate semantic\nsegmentation prediction. The proposed prototype extraction network can also be\nintegrated into most semantic segmentation networks and recognize OOD pixels.\nOn the StreetHazards dataset, the proposed PGAN model produced mIoU of 53.4%\nfor anomaly segmentation. The experimental results demonstrate PGAN may achieve\nthe SOTA performance in the anomaly segmentation tasks.",
+                    "authors": "Yiqing Hao, Yi Jin, Gaoyun An",
+                    "published": "2022-01-15",
+                    "updated": "2022-03-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2009.12232v4",
+                    "title": "From Pixel to Patch: Synthesize Context-aware Features for Zero-shot Semantic Segmentation",
+                    "abstract": "Zero-shot learning has been actively studied for image classification task to\nrelieve the burden of annotating image labels. Interestingly, semantic\nsegmentation task requires more labor-intensive pixel-wise annotation, but\nzero-shot semantic segmentation has only attracted limited research interest.\nThus, we focus on zero-shot semantic segmentation, which aims to segment unseen\nobjects with only category-level semantic representations provided for unseen\ncategories. In this paper, we propose a novel Context-aware feature Generation\nNetwork (CaGNet), which can synthesize context-aware pixel-wise visual features\nfor unseen categories based on category-level semantic representations and\npixel-wise contextual information. The synthesized features are used to\nfinetune the classifier to enable segmenting unseen objects. Furthermore, we\nextend pixel-wise feature generation and finetuning to patch-wise feature\ngeneration and finetuning, which additionally considers inter-pixel\nrelationship. Experimental results on Pascal-VOC, Pascal-Context, and\nCOCO-stuff show that our method significantly outperforms the existing\nzero-shot semantic segmentation methods. Code is available at\nhttps://github.com/bcmi/CaGNetv2-Zero-Shot-Semantic-Segmentation.",
+                    "authors": "Zhangxuan Gu, Siyuan Zhou, Li Niu, Zihan Zhao, Liqing Zhang",
+                    "published": "2020-09-25",
+                    "updated": "2022-01-21",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1605.06885v1",
+                    "title": "Bridging Category-level and Instance-level Semantic Image Segmentation",
+                    "abstract": "We propose an approach to instance-level image segmentation that is built on\ntop of category-level segmentation. Specifically, for each pixel in a semantic\ncategory mask, its corresponding instance bounding box is predicted using a\ndeep fully convolutional regression network. Thus it follows a different\npipeline to the popular detect-then-segment approaches that first predict\ninstances' bounding boxes, which are the current state-of-the-art in instance\nsegmentation. We show that, by leveraging the strength of our state-of-the-art\nsemantic segmentation models, the proposed method can achieve comparable or\neven better results to detect-then-segment approaches. We make the following\ncontributions. (i) First, we propose a simple yet effective approach to\nsemantic instance segmentation. (ii) Second, we propose an online bootstrapping\nmethod during training, which is critically important for achieving good\nperformance for both semantic category segmentation and instance-level\nsegmentation. (iii) As the performance of semantic category segmentation has a\nsignificant impact on the instance-level segmentation, which is the second step\nof our approach, we train fully convolutional residual networks to achieve the\nbest semantic category segmentation accuracy. On the PASCAL VOC 2012 dataset,\nwe obtain the currently best mean intersection-over-union score of 79.1%. (iv)\nWe also achieve state-of-the-art results for instance-level segmentation.",
+                    "authors": "Zifeng Wu, Chunhua Shen, Anton van den Hengel",
+                    "published": "2016-05-23",
+                    "updated": "2016-05-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.06753v1",
+                    "title": "3rd Place Solution for PVUW Challenge 2023: Video Panoptic Segmentation",
+                    "abstract": "In order to deal with the task of video panoptic segmentation in the wild, we\npropose a robust integrated video panoptic segmentation solution. In our\nsolution, we regard the video panoptic segmentation task as a segmentation\ntarget querying task, represent both semantic and instance targets as a set of\nqueries, and then combine these queries with video features extracted by neural\nnetworks to predict segmentation masks. In order to improve the learning\naccuracy and convergence speed of the solution, we add additional tasks of\nvideo semantic segmentation and video instance segmentation for joint training.\nIn addition, we also add an additional image semantic segmentation model to\nfurther improve the performance of semantic classes. In addition, we also add\nsome additional operations to improve the robustness of the model. Extensive\nexperiments on the VIPSeg dataset show that the proposed solution achieves\nstate-of-the-art performance with 50.04\\% VPQ on the VIPSeg test set, which is\n3rd place on the video panoptic segmentation track of the PVUW Challenge 2023.",
+                    "authors": "Jinming Su, Wangwang Yang, Junfeng Luo, Xiaolin Wei",
+                    "published": "2023-06-11",
+                    "updated": "2023-06-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2205.13278v1",
+                    "title": "Semantic Segmentation for Thermal Images: A Comparative Survey",
+                    "abstract": "Semantic segmentation is a challenging task since it requires excessively\nmore low-level spatial information of the image compared to other computer\nvision problems. The accuracy of pixel-level classification can be affected by\nmany factors, such as imaging limitations and the ambiguity of object\nboundaries in an image. Conventional methods exploit three-channel RGB images\ncaptured in the visible spectrum with deep neural networks (DNN). Thermal\nimages can significantly contribute during the segmentation since thermal\nimaging cameras are capable of capturing details despite the weather and\nillumination conditions. Using infrared spectrum in semantic segmentation has\nmany real-world use cases, such as autonomous driving, medical imaging,\nagriculture, defense industry, etc. Due to this wide range of use cases,\ndesigning accurate semantic segmentation algorithms with the help of infrared\nspectrum is an important challenge. One approach is to use both visible and\ninfrared spectrum images as inputs. These methods can accomplish higher\naccuracy due to enriched input information, with the cost of extra effort for\nthe alignment and processing of multiple inputs. Another approach is to use\nonly thermal images, enabling less hardware cost for smaller use cases. Even\nthough there are multiple surveys on semantic segmentation methods, the\nliterature lacks a comprehensive survey centered explicitly around semantic\nsegmentation using infrared spectrum. This work aims to fill this gap by\npresenting algorithms in the literature and categorizing them by their input\nimages.",
+                    "authors": "Z\u00fclfiye K\u00fct\u00fck, G\u00f6rkem Algan",
+                    "published": "2022-05-26",
+                    "updated": "2022-05-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2401.11739v1",
+                    "title": "EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models",
+                    "abstract": "Diffusion models have recently received increasing research attention for\ntheir remarkable transfer abilities in semantic segmentation tasks. However,\ngenerating fine-grained segmentation masks with diffusion models often requires\nadditional training on annotated datasets, leaving it unclear to what extent\npre-trained diffusion models alone understand the semantic relations of their\ngenerated images. To address this question, we leverage the semantic knowledge\nextracted from Stable Diffusion (SD) and aim to develop an image segmentor\ncapable of generating fine-grained segmentation maps without any additional\ntraining. The primary difficulty stems from the fact that semantically\nmeaningful feature maps typically exist only in the spatially lower-dimensional\nlayers, which poses a challenge in directly extracting pixel-level semantic\nrelations from these feature maps. To overcome this issue, our framework\nidentifies semantic correspondences between image pixels and spatial locations\nof low-dimensional feature maps by exploiting SD's generation process and\nutilizes them for constructing image-resolution segmentation maps. In extensive\nexperiments, the produced segmentation maps are demonstrated to be well\ndelineated and capture detailed parts of the images, indicating the existence\nof highly accurate pixel-level semantic knowledge in diffusion models.",
+                    "authors": "Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, Seung Wook Kim",
+                    "published": "2024-01-22",
+                    "updated": "2024-01-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.12903v1",
+                    "title": "Land Cover Change Detection via Semantic Segmentation",
+                    "abstract": "This paper presents a change detection method that identifies land cover\nchanges from aerial imagery, using semantic segmentation, a machine learning\napproach. We present a land cover classification training pipeline with Deeplab\nv3+, state-of-the-art semantic segmentation technology, including data\npreparation, model training for seven land cover types, and model exporting\nmodules. In the land cover change detection system, the inputs are images\nretrieved from Google Earth at the same location but from different times. The\nsystem then predicts semantic segmentation results on these images using the\ntrained model and calculates the land cover class percentage for each input\nimage. We see an improvement in the accuracy of the land cover semantic\nsegmentation model, with a mean IoU of 0.756 compared to 0.433, as reported in\nthe DeepGlobe land cover classification challenge. The land cover change\ndetection system that leverages the state-of-the-art semantic segmentation\ntechnology is proposed and can be used for deforestation analysis, land\nmanagement, and urban planning.",
+                    "authors": "Renee Su, Rong Chen",
+                    "published": "2019-11-28",
+                    "updated": "2019-11-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1803.10464v2",
+                    "title": "Learning Pixel-level Semantic Affinity with Image-level Supervision for Weakly Supervised Semantic Segmentation",
+                    "abstract": "The deficiency of segmentation labels is one of the main obstacles to\nsemantic segmentation in the wild. To alleviate this issue, we present a novel\nframework that generates segmentation labels of images given their image-level\nclass labels. In this weakly supervised setting, trained models have been known\nto segment local discriminative parts rather than the entire object area. Our\nsolution is to propagate such local responses to nearby areas which belong to\nthe same semantic entity. To this end, we propose a Deep Neural Network (DNN)\ncalled AffinityNet that predicts semantic affinity between a pair of adjacent\nimage coordinates. The semantic propagation is then realized by random walk\nwith the affinities predicted by AffinityNet. More importantly, the supervision\nemployed to train AffinityNet is given by the initial discriminative part\nsegmentation, which is incomplete as a segmentation annotation but sufficient\nfor learning semantic affinities within small image areas. Thus the entire\nframework relies only on image-level class labels and does not require any\nextra data or annotations. On the PASCAL VOC 2012 dataset, a DNN learned with\nsegmentation labels generated by our method outperforms previous models trained\nwith the same level of supervision, and is even as competitive as those relying\non stronger supervision.",
+                    "authors": "Jiwoon Ahn, Suha Kwak",
+                    "published": "2018-03-28",
+                    "updated": "2018-04-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.08465v2",
+                    "title": "DeepAtlas: Joint Semi-Supervised Learning of Image Registration and Segmentation",
+                    "abstract": "Deep convolutional neural networks (CNNs) are state-of-the-art for semantic\nimage segmentation, but typically require many labeled training samples.\nObtaining 3D segmentations of medical images for supervised training is\ndifficult and labor intensive. Motivated by classical approaches for joint\nsegmentation and registration we therefore propose a deep learning framework\nthat jointly learns networks for image registration and image segmentation. In\ncontrast to previous work on deep unsupervised image registration, which showed\nthe benefit of weak supervision via image segmentations, our approach can use\nexisting segmentations when available and computes them via the segmentation\nnetwork otherwise, thereby providing the same registration benefit. Conversely,\nsegmentation network training benefits from the registration, which essentially\nprovides a realistic form of data augmentation. Experiments on knee and brain\n3D magnetic resonance (MR) images show that our approach achieves large\nsimultaneous improvements of segmentation and registration accuracy (over\nindependently trained networks) and allows training high-quality models with\nvery limited training data. Specifically, in a one-shot-scenario (with only one\nmanually labeled image) our approach increases Dice scores (%) over an\nunsupervised registration network by 2.7 and 1.8 on the knee and brain images\nrespectively.",
+                    "authors": "Zhenlin Xu, Marc Niethammer",
+                    "published": "2019-04-17",
+                    "updated": "2019-07-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.04829v2",
+                    "title": "MixReorg: Cross-Modal Mixed Patch Reorganization is a Good Mask Learner for Open-World Semantic Segmentation",
+                    "abstract": "Recently, semantic segmentation models trained with image-level text\nsupervision have shown promising results in challenging open-world scenarios.\nHowever, these models still face difficulties in learning fine-grained semantic\nalignment at the pixel level and predicting accurate object masks. To address\nthis issue, we propose MixReorg, a novel and straightforward pre-training\nparadigm for semantic segmentation that enhances a model's ability to\nreorganize patches mixed across images, exploring both local visual relevance\nand global semantic coherence. Our approach involves generating fine-grained\npatch-text pairs data by mixing image patches while preserving the\ncorrespondence between patches and text. The model is then trained to minimize\nthe segmentation loss of the mixed images and the two contrastive losses of the\noriginal and restored features. With MixReorg as a mask learner, conventional\ntext-supervised semantic segmentation models can achieve highly generalizable\npixel-semantic alignment ability, which is crucial for open-world segmentation.\nAfter training with large-scale image-text data, MixReorg models can be applied\ndirectly to segment visual objects of arbitrary categories, without the need\nfor further fine-tuning. Our proposed framework demonstrates strong performance\non popular zero-shot semantic segmentation benchmarks, outperforming GroupViT\nby significant margins of 5.0%, 6.2%, 2.5%, and 3.4% mIoU on PASCAL VOC2012,\nPASCAL Context, MS COCO, and ADE20K, respectively.",
+                    "authors": "Kaixin Cai, Pengzhen Ren, Yi Zhu, Hang Xu, Jianzhuang Liu, Changlin Li, Guangrun Wang, Xiaodan Liang",
+                    "published": "2023-08-09",
+                    "updated": "2024-03-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.11316v2",
+                    "title": "Generative Semantic Segmentation",
+                    "abstract": "We present Generative Semantic Segmentation (GSS), a generative learning\napproach for semantic segmentation. Uniquely, we cast semantic segmentation as\nan image-conditioned mask generation problem. This is achieved by replacing the\nconventional per-pixel discriminative learning with a latent prior learning\nprocess. Specifically, we model the variational posterior distribution of\nlatent variables given the segmentation mask. To that end, the segmentation\nmask is expressed with a special type of image (dubbed as maskige). This\nposterior distribution allows to generate segmentation masks unconditionally.\nTo achieve semantic segmentation on a given image, we further introduce a\nconditioning network. It is optimized by minimizing the divergence between the\nposterior distribution of maskige (i.e., segmentation masks) and the latent\nprior distribution of input training images. Extensive experiments on standard\nbenchmarks show that our GSS can perform competitively to prior art\nalternatives in the standard semantic segmentation setting, whilst achieving a\nnew state of the art in the more challenging cross-domain setting.",
+                    "authors": "Jiaqi Chen, Jiachen Lu, Xiatian Zhu, Li Zhang",
+                    "published": "2023-03-20",
+                    "updated": "2023-08-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2303.07898v5",
+                    "title": "ISLE: A Framework for Image Level Semantic Segmentation Ensemble",
+                    "abstract": "One key bottleneck of employing state-of-the-art semantic segmentation\nnetworks in the real world is the availability of training labels. Conventional\nsemantic segmentation networks require massive pixel-wise annotated labels to\nreach state-of-the-art prediction quality. Hence, several works focus on\nsemantic segmentation networks trained with only image-level annotations.\nHowever, when scrutinizing the results of state-of-the-art in more detail, we\nnotice that they are remarkably close to each other on average prediction\nquality, different approaches perform better in different classes while\nproviding low quality in others. To address this problem, we propose a novel\nframework, ISLE, which employs an ensemble of the \"pseudo-labels\" for a given\nset of different semantic segmentation techniques on a class-wise level.\nPseudo-labels are the pixel-wise predictions of the image-level semantic\nsegmentation frameworks used to train the final segmentation model. Our\npseudo-labels seamlessly combine the strong points of multiple segmentation\ntechniques approaches to reach superior prediction quality. We reach up to 2.4%\nimprovement over ISLE's individual components. An exhaustive analysis was\nperformed to demonstrate ISLE's effectiveness over state-of-the-art frameworks\nfor image-level semantic segmentation.",
+                    "authors": "Erik Ostrowski, Muhammad Shafique",
+                    "published": "2023-03-14",
+                    "updated": "2023-09-20",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2007.13138v1",
+                    "title": "Virtual Multi-view Fusion for 3D Semantic Segmentation",
+                    "abstract": "Semantic segmentation of 3D meshes is an important problem for 3D scene\nunderstanding. In this paper we revisit the classic multiview representation of\n3D meshes and study several techniques that make them effective for 3D semantic\nsegmentation of meshes. Given a 3D mesh reconstructed from RGBD sensors, our\nmethod effectively chooses different virtual views of the 3D mesh and renders\nmultiple 2D channels for training an effective 2D semantic segmentation model.\nFeatures from multiple per view predictions are finally fused on 3D mesh\nvertices to predict mesh semantic segmentation labels. Using the large scale\nindoor 3D semantic segmentation benchmark of ScanNet, we show that our virtual\nviews enable more effective training of 2D semantic segmentation networks than\nprevious multiview approaches. When the 2D per pixel predictions are aggregated\non 3D surfaces, our virtual multiview fusion method is able to achieve\nsignificantly better 3D semantic segmentation results compared to all prior\nmultiview approaches and competitive with recent 3D convolution approaches.",
+                    "authors": "Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, Caroline Pantofaru",
+                    "published": "2020-07-26",
+                    "updated": "2020-07-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2205.03186v1",
+                    "title": "Semantics-Guided Moving Object Segmentation with 3D LiDAR",
+                    "abstract": "Moving object segmentation (MOS) is a task to distinguish moving objects,\ne.g., moving vehicles and pedestrians, from the surrounding static environment.\nThe segmentation accuracy of MOS can have an influence on odometry, map\nconstruction, and planning tasks. In this paper, we propose a semantics-guided\nconvolutional neural network for moving object segmentation. The network takes\nsequential LiDAR range images as inputs. Instead of segmenting the moving\nobjects directly, the network conducts single-scan-based semantic segmentation\nand multiple-scan-based moving object segmentation in turn. The semantic\nsegmentation module provides semantic priors for the MOS module, where we\npropose an adjacent scan association (ASA) module to convert the semantic\nfeatures of adjacent scans into the same coordinate system to fully exploit the\ncross-scan semantic features. Finally, by analyzing the difference between the\ntransformed features, reliable MOS result can be obtained quickly. Experimental\nresults on the SemanticKITTI MOS dataset proves the effectiveness of our work.",
+                    "authors": "Shuo Gu, Suling Yao, Jian Yang, Hui Kong",
+                    "published": "2022-05-06",
+                    "updated": "2022-05-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.RO"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2310.13026v1",
+                    "title": "Weakly-Supervised Semantic Segmentation with Image-Level Labels: from Traditional Models to Foundation Models",
+                    "abstract": "The rapid development of deep learning has driven significant progress in the\nfield of image semantic segmentation - a fundamental task in computer vision.\nSemantic segmentation algorithms often depend on the availability of\npixel-level labels (i.e., masks of objects), which are expensive,\ntime-consuming, and labor-intensive. Weakly-supervised semantic segmentation\n(WSSS) is an effective solution to avoid such labeling. It utilizes only\npartial or incomplete annotations and provides a cost-effective alternative to\nfully-supervised semantic segmentation. In this paper, we focus on the WSSS\nwith image-level labels, which is the most challenging form of WSSS. Our work\nhas two parts. First, we conduct a comprehensive survey on traditional methods,\nprimarily focusing on those presented at premier research conferences. We\ncategorize them into four groups based on where their methods operate:\npixel-wise, image-wise, cross-image, and external data. Second, we investigate\nthe applicability of visual foundation models, such as the Segment Anything\nModel (SAM), in the context of WSSS. We scrutinize SAM in two intriguing\nscenarios: text prompting and zero-shot learning. We provide insights into the\npotential and challenges associated with deploying visual foundational models\nfor WSSS, facilitating future developments in this exciting research area.",
+                    "authors": "Zhaozheng Chen, Qianru Sun",
+                    "published": "2023-10-19",
+                    "updated": "2023-10-19",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2404.04231v1",
+                    "title": "Image-Text Co-Decomposition for Text-Supervised Semantic Segmentation",
+                    "abstract": "This paper addresses text-supervised semantic segmentation, aiming to learn a\nmodel capable of segmenting arbitrary visual concepts within images by using\nonly image-text pairs without dense annotations. Existing methods have\ndemonstrated that contrastive learning on image-text pairs effectively aligns\nvisual segments with the meanings of texts. We notice that there is a\ndiscrepancy between text alignment and semantic segmentation: A text often\nconsists of multiple semantic concepts, whereas semantic segmentation strives\nto create semantically homogeneous segments. To address this issue, we propose\na novel framework, Image-Text Co-Decomposition (CoDe), where the paired image\nand text are jointly decomposed into a set of image regions and a set of word\nsegments, respectively, and contrastive learning is developed to enforce\nregion-word alignment. To work with a vision-language model, we present a\nprompt learning mechanism that derives an extra representation to highlight an\nimage segment or a word segment of interest, with which more effective features\ncan be extracted from that segment. Comprehensive experimental results\ndemonstrate that our method performs favorably against existing text-supervised\nsemantic segmentation methods on six benchmark datasets.",
+                    "authors": "Ji-Jia Wu, Andy Chia-Hao Chang, Chieh-Yu Chuang, Chun-Pei Chen, Yu-Lun Liu, Min-Hung Chen, Hou-Ning Hu, Yung-Yu Chuang, Yen-Yu Lin",
+                    "published": "2024-04-05",
+                    "updated": "2024-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2112.03241v1",
+                    "title": "Unsupervised Domain Adaptation for Semantic Image Segmentation: a Comprehensive Survey",
+                    "abstract": "Semantic segmentation plays a fundamental role in a broad variety of computer\nvision applications, providing key information for the global understanding of\nan image. Yet, the state-of-the-art models rely on large amount of annotated\nsamples, which are more expensive to obtain than in tasks such as image\nclassification. Since unlabelled data is instead significantly cheaper to\nobtain, it is not surprising that Unsupervised Domain Adaptation reached a\nbroad success within the semantic segmentation community.\n  This survey is an effort to summarize five years of this incredibly rapidly\ngrowing field, which embraces the importance of semantic segmentation itself\nand a critical need of adapting segmentation models to new environments. We\npresent the most important semantic segmentation methods; we provide a\ncomprehensive survey on domain adaptation techniques for semantic segmentation;\nwe unveil newer trends such as multi-domain learning, domain generalization,\ntest-time adaptation or source-free domain adaptation; we conclude this survey\nby describing datasets and benchmarks most widely used in semantic segmentation\nresearch. We hope that this survey will provide researchers across academia and\nindustry with a comprehensive reference guide and will help them in fostering\nnew research directions in the field.",
+                    "authors": "Gabriela Csurka, Riccardo Volpi, Boris Chidlovskii",
+                    "published": "2021-12-06",
+                    "updated": "2021-12-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "I.4.6; I.2"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.13978v3",
+                    "title": "Personalized Image Semantic Segmentation",
+                    "abstract": "Semantic segmentation models trained on public datasets have achieved great\nsuccess in recent years. However, these models didn't consider the\npersonalization issue of segmentation though it is important in practice. In\nthis paper, we address the problem of personalized image segmentation. The\nobjective is to generate more accurate segmentation results on unlabeled\npersonalized images by investigating the data's personalized traits. To open up\nfuture research in this area, we collect a large dataset containing various\nusers' personalized images called PIS (Personalized Image Semantic\nSegmentation). We also survey some recent researches related to this problem\nand report their performance on our dataset. Furthermore, by observing the\ncorrelation among a user's personalized images, we propose a baseline method\nthat incorporates the inter-image context when segmenting certain images.\nExtensive experiments show that our method outperforms the existing methods on\nthe proposed dataset. The code and the PIS dataset will be made publicly\navailable.",
+                    "authors": "Yu Zhang, Chang-Bin Zhang, Peng-Tao Jiang, Ming-Ming Cheng, Feng Mao",
+                    "published": "2021-07-24",
+                    "updated": "2021-09-04",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.02525v1",
+                    "title": "Semantic Segmentation Alternative Technique: Segmentation Domain Generation",
+                    "abstract": "Detecting objects of interest in images was always a compelling task to\nautomate. In recent years this task was more and more explored using deep\nlearning techniques, mostly using region-based convolutional networks. In this\nproject we propose an alternative semantic segmentation technique making use of\nGenerative Adversarial Networks. We consider semantic segmentation to be a\ndomain transfer problem. Thus, we train a feed forward network (FFNN) to\nreceive as input a seed real image and generate as output its segmentation\nmask.",
+                    "authors": "Ana-Cristina Rogoz, Radu Muntean, Stefan Cobeli",
+                    "published": "2021-07-06",
+                    "updated": "2021-07-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2404.04807v1",
+                    "title": "D2SL: Decouple Defogging and Semantic Learning for Foggy Domain-Adaptive Segmentation",
+                    "abstract": "We investigated domain adaptive semantic segmentation in foggy weather\nscenarios, which aims to enhance the utilization of unlabeled foggy data and\nimprove the model's adaptability to foggy conditions. Current methods rely on\nclear images as references, jointly learning defogging and segmentation for\nfoggy images. Despite making some progress, there are still two main drawbacks:\n(1) the coupling of segmentation and defogging feature representations,\nresulting in a decrease in semantic representation capability, and (2) the\nfailure to leverage real fog priors in unlabeled foggy data, leading to\ninsufficient model generalization ability. To address these issues, we propose\na novel training framework, Decouple Defogging and Semantic learning, called\nD2SL, aiming to alleviate the adverse impact of defogging tasks on the final\nsegmentation task. In this framework, we introduce a domain-consistent transfer\nstrategy to establish a connection between defogging and segmentation tasks.\nFurthermore, we design a real fog transfer strategy to improve defogging\neffects by fully leveraging the fog priors from real foggy images. Our approach\nenhances the semantic representations required for segmentation during the\ndefogging learning process and maximizes the representation capability of fog\ninvariance by effectively utilizing real fog data. Comprehensive experiments\nvalidate the effectiveness of the proposed method.",
+                    "authors": "Xuan Sun, Zhanfu An, Yuyu Liu",
+                    "published": "2024-04-07",
+                    "updated": "2024-04-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.MM"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2006.07601v1",
+                    "title": "NoPeopleAllowed: The Three-Step Approach to Weakly Supervised Semantic Segmentation",
+                    "abstract": "We propose a novel approach to weakly supervised semantic segmentation, which\nconsists of three consecutive steps. The first two steps extract high-quality\npseudo masks from image-level annotated data, which are then used to train a\nsegmentation model on the third step. The presented approach also addresses two\nproblems in the data: class imbalance and missing labels. Using only\nimage-level annotations as supervision, our method is capable of segmenting\nvarious classes and complex objects. It achieves 37.34 mean IoU on the test\nset, placing 3rd at the LID Challenge in the task of weakly supervised semantic\nsegmentation.",
+                    "authors": "Mariia Dobko, Ostap Viniavskyi, Oles Dobosevych",
+                    "published": "2020-06-13",
+                    "updated": "2020-06-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2111.11103v1",
+                    "title": "Improving Semantic Image Segmentation via Label Fusion in Semantically Textured Meshes",
+                    "abstract": "Models for semantic segmentation require a large amount of hand-labeled\ntraining data which is costly and time-consuming to produce. For this purpose,\nwe present a label fusion framework that is capable of improving semantic pixel\nlabels of video sequences in an unsupervised manner. We make use of a 3D mesh\nrepresentation of the environment and fuse the predictions of different frames\ninto a consistent representation using semantic mesh textures. Rendering the\nsemantic mesh using the original intrinsic and extrinsic camera parameters\nyields a set of improved semantic segmentation images. Due to our optimized\nCUDA implementation, we are able to exploit the entire $c$-dimensional\nprobability distribution of annotations over $c$ classes in an\nuncertainty-aware manner. We evaluate our method on the Scannet dataset where\nwe improve annotations produced by the state-of-the-art segmentation network\nESANet from $52.05 \\%$ to $58.25 \\%$ pixel accuracy. We publish the source code\nof our framework online to foster future research in this area\n(\\url{https://github.com/fferflo/semantic-meshes}). To the best of our\nknowledge, this is the first publicly available label fusion framework for\nsemantic image segmentation based on meshes with semantic textures.",
+                    "authors": "Florian Fervers, Timo Breuer, Gregor Stachowiak, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens",
+                    "published": "2021-11-22",
+                    "updated": "2021-11-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.12417v2",
+                    "title": "SLAMs: Semantic Learning based Activation Map for Weakly Supervised Semantic Segmentation",
+                    "abstract": "Recent mainstream weakly-supervised semantic segmentation (WSSS) approaches\nmainly relies on image-level classification learning, which has limited\nrepresentation capacity. In this paper, we propose a novel semantic learning\nbased framework, named SLAMs (Semantic Learning based Activation Map), for\nWSSS.",
+                    "authors": "Junliang Chen, Xiaodong Zhao, Minmin Liu, Linlin Shen",
+                    "published": "2022-10-22",
+                    "updated": "2022-11-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.02840v1",
+                    "title": "Attention-based fusion of semantic boundary and non-boundary information to improve semantic segmentation",
+                    "abstract": "This paper introduces a method for image semantic segmentation grounded on a\nnovel fusion scheme, which takes place inside a deep convolutional neural\nnetwork. The main goal of our proposal is to explore object boundary\ninformation to improve the overall segmentation performance. Unlike previous\nworks that combine boundary and segmentation features, or those that use\nboundary information to regularize semantic segmentation, we instead propose a\nnovel approach that embodies boundary information onto segmentation. For that,\nour semantic segmentation method uses two streams, which are combined through\nan attention gate, forming an end-to-end Y-model. To the best of our knowledge,\nours is the first work to show that boundary detection can improve semantic\nsegmentation when fused through a semantic fusion gate (attention model). We\nperformed an extensive evaluation of our method over public data sets. We found\ncompetitive results on all data sets after comparing our proposed model with\nother twelve state-of-the-art segmenters, considering the same training\nconditions. Our proposed model achieved the best mIoU on the CityScapes,\nCamVid, and Pascal Context data sets, and the second best on Mapillary Vistas.",
+                    "authors": "Jefferson Fontinele, Gabriel Lefundes, Luciano Oliveira",
+                    "published": "2021-08-05",
+                    "updated": "2021-08-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2011.00674v1",
+                    "title": "Highway Driving Dataset for Semantic Video Segmentation",
+                    "abstract": "Scene understanding is an essential technique in semantic segmentation.\nAlthough there exist several datasets that can be used for semantic\nsegmentation, they are mainly focused on semantic image segmentation with large\ndeep neural networks. Therefore, these networks are not useful for real time\napplications, especially in autonomous driving systems. In order to solve this\nproblem, we make two contributions to semantic segmentation task. The first\ncontribution is that we introduce the semantic video dataset, the Highway\nDriving dataset, which is a densely annotated benchmark for a semantic video\nsegmentation task. The Highway Driving dataset consists of 20 video sequences\nhaving a 30Hz frame rate, and every frame is densely annotated. Secondly, we\npropose a baseline algorithm that utilizes a temporal correlation. Together\nwith our attempt to analyze the temporal correlation, we expect the Highway\nDriving dataset to encourage research on semantic video segmentation.",
+                    "authors": "Byungju Kim, Junho Yim, Junmo Kim",
+                    "published": "2020-11-02",
+                    "updated": "2020-11-02",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2312.17243v1",
+                    "title": "Unsupervised Universal Image Segmentation",
+                    "abstract": "Several unsupervised image segmentation approaches have been proposed which\neliminate the need for dense manually-annotated segmentation masks; current\nmodels separately handle either semantic segmentation (e.g., STEGO) or\nclass-agnostic instance segmentation (e.g., CutLER), but not both (i.e.,\npanoptic segmentation). We propose an Unsupervised Universal Segmentation model\n(U2Seg) adept at performing various image segmentation tasks -- instance,\nsemantic and panoptic -- using a novel unified framework. U2Seg generates\npseudo semantic labels for these segmentation tasks via leveraging\nself-supervised models followed by clustering; each cluster represents\ndifferent semantic and/or instance membership of pixels. We then self-train the\nmodel on these pseudo semantic labels, yielding substantial performance gains\nover specialized methods tailored to each task: a +2.6 AP$^{\\text{box}}$ boost\nvs. CutLER in unsupervised instance segmentation on COCO and a +7.0 PixelAcc\nincrease (vs. STEGO) in unsupervised semantic segmentation on COCOStuff.\nMoreover, our method sets up a new baseline for unsupervised panoptic\nsegmentation, which has not been previously explored. U2Seg is also a strong\npretrained model for few-shot segmentation, surpassing CutLER by +5.0\nAP$^{\\text{mask}}$ when trained on a low-data regime, e.g., only 1% COCO\nlabels. We hope our simple yet effective method can inspire more research on\nunsupervised universal image segmentation.",
+                    "authors": "Dantong Niu, Xudong Wang, Xinyang Han, Long Lian, Roei Herzig, Trevor Darrell",
+                    "published": "2023-12-28",
+                    "updated": "2023-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2103.00286v1",
+                    "title": "A Novel Adaptive Deep Network for Building Footprint Segmentation",
+                    "abstract": "Building footprint segmentations for high resolution images are increasingly\ndemanded for many remote sensing applications. By the emerging deep learning\napproaches, segmentation networks have made significant advances in the\nsemantic segmentation of objects. However, these advances and the increased\naccess to satellite images require the generation of accurate object boundaries\nin satellite images. In the current paper, we propose a novel network-based on\nPix2Pix methodology to solve the problem of inaccurate boundaries obtained by\nconverting satellite images into maps using segmentation networks in order to\nsegment building footprints. To define the new network named G2G, our framework\nincludes two generators where the first generator extracts localization\nfeatures in order to merge them with the boundary features extracted from the\nsecond generator to segment all detailed building edges. Moreover, different\nstrategies are implemented to enhance the quality of the proposed networks'\nresults, implying that the proposed network outperforms state-of-the-art\nnetworks in segmentation accuracy with a large margin for all evaluation\nmetrics. The implementation is available at\nhttps://github.com/A2Amir/A-Novel-Adaptive-Deep-Network-for-Building-Footprint-Segmentation.",
+                    "authors": "A. Ziaee, R. Dehbozorgi, M. D\u00f6ller",
+                    "published": "2021-02-27",
+                    "updated": "2021-02-27",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG",
+                        "cs.NA",
+                        "eess.IV",
+                        "math.NA"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1707.02432v2",
+                    "title": "Deep Semantic Segmentation for Automated Driving: Taxonomy, Roadmap and Challenges",
+                    "abstract": "Semantic segmentation was seen as a challenging computer vision problem few\nyears ago. Due to recent advancements in deep learning, relatively accurate\nsolutions are now possible for its use in automated driving. In this paper, the\nsemantic segmentation problem is explored from the perspective of automated\ndriving. Most of the current semantic segmentation algorithms are designed for\ngeneric images and do not incorporate prior structure and end goal for\nautomated driving. First, the paper begins with a generic taxonomic survey of\nsemantic segmentation algorithms and then discusses how it fits in the context\nof automated driving. Second, the particular challenges of deploying it into a\nsafety system which needs high level of accuracy and robustness are listed.\nThird, different alternatives instead of using an independent semantic\nsegmentation module are explored. Finally, an empirical evaluation of various\nsemantic segmentation architectures was performed on CamVid dataset in terms of\naccuracy and speed. This paper is a preliminary shorter version of a more\ndetailed survey which is work in progress.",
+                    "authors": "Mennatullah Siam, Sara Elkerdawy, Martin Jagersand, Senthil Yogamani",
+                    "published": "2017-07-08",
+                    "updated": "2017-08-03",
+                    "primary_cat": "stat.ML",
+                    "cats": [
+                        "stat.ML",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.05490v1",
+                    "title": "Learning Semantic Segmentation with Query Points Supervision on Aerial Images",
+                    "abstract": "Semantic segmentation is crucial in remote sensing, where high-resolution\nsatellite images are segmented into meaningful regions. Recent advancements in\ndeep learning have significantly improved satellite image segmentation.\nHowever, most of these methods are typically trained in fully supervised\nsettings that require high-quality pixel-level annotations, which are expensive\nand time-consuming to obtain. In this work, we present a weakly supervised\nlearning algorithm to train semantic segmentation algorithms that only rely on\nquery point annotations instead of full mask labels. Our proposed approach\nperforms accurate semantic segmentation and improves efficiency by\nsignificantly reducing the cost and time required for manual annotation.\nSpecifically, we generate superpixels and extend the query point labels into\nthose superpixels that group similar meaningful semantics. Then, we train\nsemantic segmentation models, supervised with images partially labeled with the\nsuperpixels pseudo-labels. We benchmark our weakly supervised training approach\non an aerial image dataset and different semantic segmentation architectures,\nshowing that we can reach competitive performance compared to fully supervised\ntraining while reducing the annotation effort.",
+                    "authors": "Santiago Rivier, Carlos Hinojosa, Silvio Giancola, Bernard Ghanem",
+                    "published": "2023-09-11",
+                    "updated": "2023-09-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI",
+                        "cs.LG"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1610.01706v1",
+                    "title": "Exploiting Depth from Single Monocular Images for Object Detection and Semantic Segmentation",
+                    "abstract": "Augmenting RGB data with measured depth has been shown to improve the\nperformance of a range of tasks in computer vision including object detection\nand semantic segmentation. Although depth sensors such as the Microsoft Kinect\nhave facilitated easy acquisition of such depth information, the vast majority\nof images used in vision tasks do not contain depth information. In this paper,\nwe show that augmenting RGB images with estimated depth can also improve the\naccuracy of both object detection and semantic segmentation. Specifically, we\nfirst exploit the recent success of depth estimation from monocular images and\nlearn a deep depth estimation model. Then we learn deep depth features from the\nestimated depth and combine with RGB features for object detection and semantic\nsegmentation. Additionally, we propose an RGB-D semantic segmentation method\nwhich applies a multi-task training scheme: semantic label prediction and depth\nvalue regression. We test our methods on several datasets and demonstrate that\nincorporating information from estimated depth improves the performance of\nobject detection and semantic segmentation remarkably.",
+                    "authors": "Yuanzhouhan Cao, Chunhua Shen, Heng Tao Shen",
+                    "published": "2016-10-06",
+                    "updated": "2016-10-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2107.12518v2",
+                    "title": "Segmentation in Style: Unsupervised Semantic Image Segmentation with Stylegan and CLIP",
+                    "abstract": "We introduce a method that allows to automatically segment images into\nsemantically meaningful regions without human supervision. Derived regions are\nconsistent across different images and coincide with human-defined semantic\nclasses on some datasets. In cases where semantic regions might be hard for\nhuman to define and consistently label, our method is still able to find\nmeaningful and consistent semantic classes. In our work, we use pretrained\nStyleGAN2 generative model: clustering in the feature space of the generative\nmodel allows to discover semantic classes. Once classes are discovered, a\nsynthetic dataset with generated images and corresponding segmentation masks\ncan be created. After that a segmentation model is trained on the synthetic\ndataset and is able to generalize to real images. Additionally, by using CLIP\nwe are able to use prompts defined in a natural language to discover some\ndesired semantic classes. We test our method on publicly available datasets and\nshow state-of-the-art results.",
+                    "authors": "Daniil Pakhomov, Sanchit Hira, Narayani Wagle, Kemar E. Green, Nassir Navab",
+                    "published": "2021-07-26",
+                    "updated": "2021-11-18",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.04297v1",
+                    "title": "SHDM-NET: Heat Map Detail Guidance with Image Matting for Industrial Weld Semantic Segmentation Network",
+                    "abstract": "In actual industrial production, the assessment of the steel plate welding\neffect is an important task, and the segmentation of the weld section is the\nbasis of the assessment. This paper proposes an industrial weld segmentation\nnetwork based on a deep learning semantic segmentation algorithm fused with\nheatmap detail guidance and Image Matting to solve the automatic segmentation\nproblem of weld regions. In the existing semantic segmentation networks, the\nboundary information can be preserved by fusing the features of both high-level\nand low-level layers. However, this method can lead to insufficient expression\nof the spatial information in the low-level layer, resulting in inaccurate\nsegmentation boundary positioning. We propose a detailed guidance module based\non heatmaps to fully express the segmented region boundary information in the\nlow-level network to address this problem. Specifically, the expression of\nboundary information can be enhanced by adding a detailed branch to predict\nsegmented boundary and then matching it with the boundary heat map generated by\nmask labels to calculate the mean square error loss. In addition, although deep\nlearning has achieved great success in the field of semantic segmentation, the\nprecision of the segmentation boundary region is not high due to the loss of\ndetailed information caused by the classical segmentation network in the\nprocess of encoding and decoding process. This paper introduces a matting\nalgorithm to calibrate the boundary of the segmentation region of the semantic\nsegmentation network to solve this problem. Through many experiments on\nindustrial weld data sets, the effectiveness of our method is demonstrated, and\nthe MIOU reaches 97.93%. It is worth noting that this performance is comparable\nto human manual segmentation ( MIOU 97.96%).",
+                    "authors": "Qi Wang, Jingwu Mei",
+                    "published": "2022-07-09",
+                    "updated": "2022-07-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.03138v1",
+                    "title": "Quantity beats quality for semantic segmentation of corrosion in images",
+                    "abstract": "Dataset creation is typically one of the first steps when applying Artificial\nIntelligence methods to a new task; and the real world performance of models\nhinges on the quality and quantity of data available. Producing an image\ndataset for semantic segmentation is resource intensive, particularly for\nspecialist subjects where class segmentation is not able to be effectively\nfarmed out. The benefit of producing a large, but poorly labelled, dataset\nversus a small, expertly segmented dataset for semantic segmentation is an open\nquestion. Here we show that a large, noisy dataset outperforms a small,\nexpertly segmented dataset for training a Fully Convolutional Network model for\nsemantic segmentation of corrosion in images. A large dataset of 250 images\nwith segmentations labelled by undergraduates and a second dataset of just 10\nimages, with segmentations labelled by subject matter experts were produced.\nThe mean Intersection over Union and micro F-score metrics were compared after\ntraining for 50,000 epochs. This work is illustrative for researchers setting\nout to develop deep learning models for detection and location of specialist\nfeatures.",
+                    "authors": "Will Nash, Tom Drummond, Nick Birbilis",
+                    "published": "2018-06-30",
+                    "updated": "2018-06-30",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2405.04913v1",
+                    "title": "Weakly-supervised Semantic Segmentation via Dual-stream Contrastive Learning of Cross-image Contextual Information",
+                    "abstract": "Weakly supervised semantic segmentation (WSSS) aims at learning a semantic\nsegmentation model with only image-level tags. Despite intensive research on\ndeep learning approaches over a decade, there is still a significant\nperformance gap between WSSS and full semantic segmentation. Most current WSSS\nmethods always focus on a limited single image (pixel-wise) information while\nignoring the valuable inter-image (semantic-wise) information. From this\nperspective, a novel end-to-end WSSS framework called DSCNet is developed along\nwith two innovations: i) pixel-wise group contrast and semantic-wise graph\ncontrast are proposed and introduced into the WSSS framework; ii) a novel\ndual-stream contrastive learning (DSCL) mechanism is designed to jointly handle\npixel-wise and semantic-wise context information for better WSSS performance.\nSpecifically, the pixel-wise group contrast learning (PGCL) and semantic-wise\ngraph contrast learning (SGCL) tasks form a more comprehensive solution.\nExtensive experiments on PASCAL VOC and MS COCO benchmarks verify the\nsuperiority of DSCNet over SOTA approaches and baseline models.",
+                    "authors": "Qi Lai, Chi-Man Vong",
+                    "published": "2024-05-08",
+                    "updated": "2024-05-08",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.13395v3",
+                    "title": "ACDC: The Adverse Conditions Dataset with Correspondences for Semantic Driving Scene Understanding",
+                    "abstract": "Level 5 autonomy for self-driving cars requires a robust visual perception\nsystem that can parse input images under any visual condition. However,\nexisting semantic segmentation datasets are either dominated by images captured\nunder normal conditions or are small in scale. To address this, we introduce\nACDC, the Adverse Conditions Dataset with Correspondences for training and\ntesting semantic segmentation methods on adverse visual conditions. ACDC\nconsists of a large set of 4006 images which are equally distributed between\nfour common adverse conditions: fog, nighttime, rain, and snow. Each\nadverse-condition image comes with a high-quality fine pixel-level semantic\nannotation, a corresponding image of the same scene taken under normal\nconditions, and a binary mask that distinguishes between intra-image regions of\nclear and uncertain semantic content. Thus, ACDC supports both standard\nsemantic segmentation and the newly introduced uncertainty-aware semantic\nsegmentation. A detailed empirical study demonstrates the challenges that the\nadverse domains of ACDC pose to state-of-the-art supervised and unsupervised\napproaches and indicates the value of our dataset in steering future progress\nin the field. Our dataset and benchmark are publicly available.",
+                    "authors": "Christos Sakaridis, Dengxin Dai, Luc Van Gool",
+                    "published": "2021-04-27",
+                    "updated": "2021-09-01",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.12545v1",
+                    "title": "Unsupervised Domain Adaptation for Semantic Segmentation by Content Transfer",
+                    "abstract": "In this paper, we tackle the unsupervised domain adaptation (UDA) for\nsemantic segmentation, which aims to segment the unlabeled real data using\nlabeled synthetic data. The main problem of UDA for semantic segmentation\nrelies on reducing the domain gap between the real image and synthetic image.\nTo solve this problem, we focused on separating information in an image into\ncontent and style. Here, only the content has cues for semantic segmentation,\nand the style makes the domain gap. Thus, precise separation of content and\nstyle in an image leads to effect as supervision of real data even when\nlearning with synthetic data. To make the best of this effect, we propose a\nzero-style loss. Even though we perfectly extract content for semantic\nsegmentation in the real domain, another main challenge, the class imbalance\nproblem, still exists in UDA for semantic segmentation. We address this problem\nby transferring the contents of tail classes from synthetic to real domain.\nExperimental results show that the proposed method achieves the\nstate-of-the-art performance in semantic segmentation on the major two UDA\nsettings.",
+                    "authors": "Suhyeon Lee, Junhyuk Hyun, Hongje Seong, Euntai Kim",
+                    "published": "2020-12-23",
+                    "updated": "2020-12-23",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1905.08748v3",
+                    "title": "RIU-Net: Embarrassingly simple semantic segmentation of 3D LiDAR point cloud",
+                    "abstract": "This paper proposes RIU-Net (for Range-Image U-Net), the adaptation of a\npopular semantic segmentation network for the semantic segmentation of a 3D\nLiDAR point cloud. The point cloud is turned into a 2D range-image by\nexploiting the topology of the sensor. This image is then used as input to a\nU-net. This architecture has already proved its efficiency for the task of\nsemantic segmentation of medical images. We demonstrate how it can also be used\nfor the accurate semantic segmentation of a 3D LiDAR point cloud and how it\nrepresents a valid bridge between image processing and 3D point cloud\nprocessing. Our model is trained on range-images built from KITTI 3D object\ndetection dataset. Experiments show that RIU-Net, despite being very simple,\noffers results that are comparable to the state-of-the-art of range-image based\nmethods. Finally, we demonstrate that this architecture is able to operate at\n90fps on a single GPU, which enables deployment for real-time segmentation.",
+                    "authors": "Pierre Biasutti, Aur\u00e9lie Bugeau, Jean-Fran\u00e7ois Aujol, Mathieu Br\u00e9dif",
+                    "published": "2019-05-21",
+                    "updated": "2019-06-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2207.08455v3",
+                    "title": "Open-world Semantic Segmentation via Contrasting and Clustering Vision-Language Embedding",
+                    "abstract": "To bridge the gap between supervised semantic segmentation and real-world\napplications that acquires one model to recognize arbitrary new concepts,\nrecent zero-shot segmentation attracts a lot of attention by exploring the\nrelationships between unseen and seen object categories, yet requiring large\namounts of densely-annotated data with diverse base classes. In this paper, we\npropose a new open-world semantic segmentation pipeline that makes the first\nattempt to learn to segment semantic objects of various open-world categories\nwithout any efforts on dense annotations, by purely exploiting the\nimage-caption data that naturally exist on the Internet. Our method,\nVision-language-driven Semantic Segmentation (ViL-Seg), employs an image and a\ntext encoder to generate visual and text embeddings for the image-caption data,\nwith two core components that endow its segmentation ability: First, the image\nencoder is jointly trained with a vision-based contrasting and a cross-modal\ncontrasting, which encourage the visual embeddings to preserve both\nfine-grained semantics and high-level category information that are crucial for\nthe segmentation task. Furthermore, an online clustering head is devised over\nthe image encoder, which allows to dynamically segment the visual embeddings\ninto distinct semantic groups such that they can be classified by comparing\nwith various text embeddings to complete our segmentation pipeline. Experiments\nshow that without using any data with dense annotations, our method can\ndirectly segment objects of arbitrary categories, outperforming zero-shot\nsegmentation methods that require data labeling on three benchmark datasets.",
+                    "authors": "Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, Xiaodan Liang",
+                    "published": "2022-07-18",
+                    "updated": "2022-10-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2211.08352v1",
+                    "title": "Visual Semantic Segmentation Based on Few/Zero-Shot Learning: An Overview",
+                    "abstract": "Visual semantic segmentation aims at separating a visual sample into diverse\nblocks with specific semantic attributes and identifying the category for each\nblock, and it plays a crucial role in environmental perception. Conventional\nlearning-based visual semantic segmentation approaches count heavily on\nlarge-scale training data with dense annotations and consistently fail to\nestimate accurate semantic labels for unseen categories. This obstruction spurs\na craze for studying visual semantic segmentation with the assistance of\nfew/zero-shot learning. The emergence and rapid progress of few/zero-shot\nvisual semantic segmentation make it possible to learn unseen-category from a\nfew labeled or zero-labeled samples, which advances the extension to practical\napplications. Therefore, this paper focuses on the recently published\nfew/zero-shot visual semantic segmentation methods varying from 2D to 3D space\nand explores the commonalities and discrepancies of technical settlements under\ndifferent segmentation circumstances. Specifically, the preliminaries on\nfew/zero-shot visual semantic segmentation, including the problem definitions,\ntypical datasets, and technical remedies, are briefly reviewed and discussed.\nMoreover, three typical instantiations are involved to uncover the interactions\nof few/zero-shot learning with visual semantic segmentation, including image\nsemantic segmentation, video object segmentation, and 3D segmentation. Finally,\nthe future challenges of few/zero-shot visual semantic segmentation are\ndiscussed.",
+                    "authors": "Wenqi Ren, Yang Tang, Qiyu Sun, Chaoqiang Zhao, Qing-Long Han",
+                    "published": "2022-11-13",
+                    "updated": "2022-11-13",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2101.08418v2",
+                    "title": "Rethinking Semantic Segmentation Evaluation for Explainability and Model Selection",
+                    "abstract": "Semantic segmentation aims to robustly predict coherent class labels for\nentire regions of an image. It is a scene understanding task that powers\nreal-world applications (e.g., autonomous navigation). One important\napplication, the use of imagery for automated semantic understanding of\npedestrian environments, provides remote mapping of accessibility features in\nstreet environments. This application (and others like it) require detailed\ngeometric information of geographical objects. Semantic segmentation is a\nprerequisite for this task since it maps contiguous regions of the same class\nas single entities. Importantly, semantic segmentation uses like ours are not\npixel-wise outcomes; however, most of their quantitative evaluation metrics\n(e.g., mean Intersection Over Union) are based on pixel-wise similarities to a\nground-truth, which fails to emphasize over- and under-segmentation properties\nof a segmentation model. Here, we introduce a new metric to assess region-based\nover- and under-segmentation. We analyze and compare it to other metrics,\ndemonstrating that the use of our metric lends greater explainability to\nsemantic segmentation model performance in real-world applications.",
+                    "authors": "Yuxiang Zhang, Sachin Mehta, Anat Caspi",
+                    "published": "2021-01-21",
+                    "updated": "2023-02-15",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2403.01482v4",
+                    "title": "EAGLE: Eigen Aggregation Learning for Object-Centric Unsupervised Semantic Segmentation",
+                    "abstract": "Semantic segmentation has innately relied on extensive pixel-level annotated\ndata, leading to the emergence of unsupervised methodologies. Among them,\nleveraging self-supervised Vision Transformers for unsupervised semantic\nsegmentation (USS) has been making steady progress with expressive deep\nfeatures. Yet, for semantically segmenting images with complex objects, a\npredominant challenge remains: the lack of explicit object-level semantic\nencoding in patch-level features. This technical limitation often leads to\ninadequate segmentation of complex objects with diverse structures. To address\nthis gap, we present a novel approach, EAGLE, which emphasizes object-centric\nrepresentation learning for unsupervised semantic segmentation. Specifically,\nwe introduce EiCue, a spectral technique providing semantic and structural cues\nthrough an eigenbasis derived from the semantic similarity matrix of deep image\nfeatures and color affinity from an image. Further, by incorporating our\nobject-centric contrastive loss with EiCue, we guide our model to learn\nobject-level representations with intra- and inter-image object-feature\nconsistency, thereby enhancing semantic accuracy. Extensive experiments on\nCOCO-Stuff, Cityscapes, and Potsdam-3 datasets demonstrate the state-of-the-art\nUSS results of EAGLE with accurate and consistent semantic segmentation across\ncomplex scenes.",
+                    "authors": "Chanyoung Kim, Woojung Han, Dayun Ju, Seong Jae Hwang",
+                    "published": "2024-03-03",
+                    "updated": "2024-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1502.04983v1",
+                    "title": "Context Tricks for Cheap Semantic Segmentation",
+                    "abstract": "Accurate semantic labeling of image pixels is difficult because intra-class\nvariability is often greater than inter-class variability. In turn, fast\nsemantic segmentation is hard because accurate models are usually too\ncomplicated to also run quickly at test-time. Our experience with building and\nrunning semantic segmentation systems has also shown a reasonably obvious\nbottleneck on model complexity, imposed by small training datasets. We\ntherefore propose two simple complementary strategies that leverage context to\ngive better semantic segmentation, while scaling up or down to train on\ndifferent-sized datasets.\n  As easy modifications for existing semantic segmentation algorithms, we\nintroduce Decorrelated Semantic Texton Forests, and the Context Sensitive Image\nLevel Prior. The proposed modifications are tested using a Semantic Texton\nForest (STF) system, and the modifications are validated on two standard\nbenchmark datasets, MSRC-21 and PascalVOC-2010. In Python based comparisons,\nour system is insignificantly slower than STF at test-time, yet produces\nsuperior semantic segmentations overall, with just push-button training.",
+                    "authors": "Thanapong Intharah, Gabriel J. Brostow",
+                    "published": "2015-02-17",
+                    "updated": "2015-02-17",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.05321v1",
+                    "title": "Image Segmentation Semantic Communication over Internet of Vehicles",
+                    "abstract": "In this paper, the problem of semantic-based efficient image transmission is\nstudied over the Internet of Vehicles (IoV). In the considered model, a vehicle\nshares massive amount of visual data perceived by its visual sensors to assist\nother vehicles in making driving decisions. However, it is hard to maintain a\nhigh reliable visual data transmission due to the limited spectrum resources.\nTo tackle this problem, a semantic communication approach is introduced to\nreduce the transmission data amount while ensuring the semantic-level accuracy.\nParticularly, an image segmentation semantic communication (ISSC) system is\nproposed, which can extract the semantic features from the perceived images and\ntransmit the features to the receiving vehicle that reconstructs the image\nsegmentations. The ISSC system consists of an encoder and a decoder at the\ntransmitter and the receiver, respectively. To accurately extract the image\nsemantic features, the ISSC system encoder employs a Swin Transformer based\nmulti-scale semantic feature extractor. Then, to resist the wireless noise and\nreconstruct the image segmentation, a semantic feature decoder and a\nreconstructor are designed at the receiver. Simulation results show that the\nproposed ISSC system can reconstruct the image segmentation accurately with a\nhigh compression ratio, and can achieve robust transmission performance against\nchannel noise, especially at the low signal-to-noise ratio (SNR). In terms of\nmean Intersection over Union (mIoU), the ISSC system can achieve an increase by\n75%, compared to the baselines using traditional coding method",
+                    "authors": "Qiang Pan, Haonan Tong, Jie Lv, Tao Luo, Zhilong Zhang, Changchuan Yin, Jianfeng Li",
+                    "published": "2022-10-11",
+                    "updated": "2022-10-11",
+                    "primary_cat": "cs.NI",
+                    "cats": [
+                        "cs.NI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2210.05564v1",
+                    "title": "Hypergraph Convolutional Networks for Weakly-Supervised Semantic Segmentation",
+                    "abstract": "Semantic segmentation is a fundamental topic in computer vision. Several deep\nlearning methods have been proposed for semantic segmentation with outstanding\nresults. However, these models require a lot of densely annotated images. To\naddress this problem, we propose a new algorithm that uses HyperGraph\nConvolutional Networks for Weakly-supervised Semantic Segmentation\n(HyperGCN-WSS). Our algorithm constructs spatial and k-Nearest Neighbor (k-NN)\ngraphs from the images in the dataset to generate the hypergraphs. Then, we\ntrain a specialized HyperGraph Convolutional Network (HyperGCN) architecture\nusing some weak signals. The outputs of the HyperGCN are denominated\npseudo-labels, which are later used to train a DeepLab model for semantic\nsegmentation. HyperGCN-WSS is evaluated on the PASCAL VOC 2012 dataset for\nsemantic segmentation, using scribbles or clicks as weak signals. Our algorithm\nshows competitive performance against previous methods.",
+                    "authors": "Jhony H. Giraldo, Vincenzo Scarrica, Antonino Staiano, Francesco Camastra, Thierry Bouwmans",
+                    "published": "2022-10-11",
+                    "updated": "2022-10-11",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1911.00679v3",
+                    "title": "Cooperative Semantic Segmentation and Image Restoration in Adverse Environmental Conditions",
+                    "abstract": "Most state-of-the-art semantic segmentation approaches only achieve high\naccuracy in good conditions. In practically-common but less-discussed adverse\nenvironmental conditions, their performance can decrease enormously. Existing\nstudies usually cast the handling of segmentation in adverse conditions as a\nseparate post-processing step after signal restoration, making the segmentation\nperformance largely depend on the quality of restoration. In this paper, we\npropose a novel deep-learning framework to tackle semantic segmentation and\nimage restoration in adverse environmental conditions in a holistic manner. The\nproposed approach contains two components: Semantically-Guided Adaptation,\nwhich exploits semantic information from degraded images to refine the\nsegmentation; and Exemplar-Guided Synthesis, which restores images from\nsemantic label maps given degraded exemplars as the guidance. Our method\ncooperatively leverages the complementarity and interdependence of low-level\nrestoration and high-level segmentation in adverse environmental conditions.\nExtensive experiments on various datasets demonstrate that our approach can not\nonly improve the accuracy of semantic segmentation with degradation cues, but\nalso boost the perceptual quality and structural similarity of image\nrestoration with semantic guidance.",
+                    "authors": "Weihao Xia, Zhanglin Cheng, Yujiu Yang, Jing-Hao Xue",
+                    "published": "2019-11-02",
+                    "updated": "2020-03-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "eess.IV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1809.10245v1",
+                    "title": "Cylindrical Transform: 3D Semantic Segmentation of Kidneys With Limited Annotated Images",
+                    "abstract": "In this paper, we propose a novel technique for sampling sequential images\nusing a cylindrical transform in a cylindrical coordinate system for kidney\nsemantic segmentation in abdominal computed tomography (CT). The images\ngenerated from a cylindrical transform augment a limited annotated set of\nimages in three dimensions. This approach enables us to train contemporary\nclassification deep convolutional neural networks (DCNNs) instead of fully\nconvolutional networks (FCNs) for semantic segmentation. Typical semantic\nsegmentation models segment a sequential set of images (e.g. CT or video) by\nsegmenting each image independently. However, the proposed method not only\nconsiders the spatial dependency in the x-y plane, but also the spatial\nsequential dependency along the z-axis. The results show that classification\nDCNNs, trained on cylindrical transformed images, can achieve a higher\nsegmentation performance value than FCNs using a limited number of annotated\nimages.",
+                    "authors": "Hojjat Salehinejad, Sumeya Naqvi, Errol Colak, Joseph Barfett, Shahrokh Valaee",
+                    "published": "2018-09-24",
+                    "updated": "2018-09-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.NE"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2108.13118v1",
+                    "title": "Automatic Preprocessing and Ensemble Learning for Low Quality Cell Image Segmentation",
+                    "abstract": "We propose an automatic preprocessing and ensemble learning for segmentation\nof cell images with low quality. It is difficult to capture cells with strong\nlight. Therefore, the microscopic images of cells tend to have low image\nquality but these images are not good for semantic segmentation. Here we\npropose a method to translate an input image to the images that are easy to\nrecognize by deep learning. The proposed method consists of two deep neural\nnetworks. The first network is the usual training for semantic segmentation,\nand penultimate feature maps of the first network are used as filters to\ntranslate an input image to the images that emphasize each class. This is the\nautomatic preprocessing and translated cell images are easily classified. The\ninput cell image with low quality is translated by the feature maps in the\nfirst network, and the translated images are fed into the second network for\nsemantic segmentation. Since the outputs of the second network are multiple\nsegmentation results, we conduct the weighted ensemble of those segmentation\nimages. Two networks are trained by end-to-end manner, and we do not need to\nprepare images with high quality for the translation. We confirmed that our\nproposed method can translate cell images with low quality to the images that\nare easy to segment, and segmentation accuracy has improved using the weighted\nensemble learning.",
+                    "authors": "Sota Kato, Kazuhiro Hotta",
+                    "published": "2021-08-30",
+                    "updated": "2021-08-30",
+                    "primary_cat": "eess.IV",
+                    "cats": [
+                        "eess.IV",
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1609.09220v1",
+                    "title": "CNN-aware Binary Map for General Semantic Segmentation",
+                    "abstract": "In this paper we introduce a novel method for general semantic segmentation\nthat can benefit from general semantics of Convolutional Neural Network (CNN).\nOur segmentation proposes visually and semantically coherent image segments. We\nuse binary encoding of CNN features to overcome the difficulty of the\nclustering on the high-dimensional CNN feature space. These binary codes are\nvery robust against noise and non-semantic changes in the image. These binary\nencoding can be embedded into the CNN as an extra layer at the end of the\nnetwork. This results in real-time segmentation. To the best of our knowledge\nour method is the first attempt on general semantic image segmentation using\nCNN. All the previous papers were limited to few number of category of the\nimages (e.g. PASCAL VOC). Experiments show that our segmentation algorithm\noutperform the state-of-the-art non-semantic segmentation methods by large\nmargin.",
+                    "authors": "Mahdyar Ravanbakhsh, Hossein Mousavi, Moin Nabi, Mohammad Rastegari, Carlo Regazzoni",
+                    "published": "2016-09-29",
+                    "updated": "2016-09-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1807.11857v1",
+                    "title": "Joint Learning of Intrinsic Images and Semantic Segmentation",
+                    "abstract": "Semantic segmentation of outdoor scenes is problematic when there are\nvariations in imaging conditions. It is known that albedo (reflectance) is\ninvariant to all kinds of illumination effects. Thus, using reflectance images\nfor semantic segmentation task can be favorable. Additionally, not only\nsegmentation may benefit from reflectance, but also segmentation may be useful\nfor reflectance computation. Therefore, in this paper, the tasks of semantic\nsegmentation and intrinsic image decomposition are considered as a combined\nprocess by exploring their mutual relationship in a joint fashion. To that end,\nwe propose a supervised end-to-end CNN architecture to jointly learn intrinsic\nimage decomposition and semantic segmentation. We analyze the gains of\naddressing those two problems jointly. Moreover, new cascade CNN architectures\nfor intrinsic-for-segmentation and segmentation-for-intrinsic are proposed as\nsingle tasks. Furthermore, a dataset of 35K synthetic images of natural\nenvironments is created with corresponding albedo and shading (intrinsics), as\nwell as semantic labels (segmentation) assigned to each object/scene. The\nexperiments show that joint learning of intrinsic image decomposition and\nsemantic segmentation is beneficial for both tasks for natural scenes. Dataset\nand models are available at: https://ivi.fnwi.uva.nl/cv/intrinseg",
+                    "authors": "Anil S. Baslamisli, Thomas T. Groenestege, Partha Das, Hoang-An Le, Sezer Karaoglu, Theo Gevers",
+                    "published": "2018-07-31",
+                    "updated": "2018-07-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2012.10782v2",
+                    "title": "Three Ways to Improve Semantic Segmentation with Self-Supervised Depth Estimation",
+                    "abstract": "Training deep networks for semantic segmentation requires large amounts of\nlabeled training data, which presents a major challenge in practice, as\nlabeling segmentation masks is a highly labor-intensive process. To address\nthis issue, we present a framework for semi-supervised semantic segmentation,\nwhich is enhanced by self-supervised monocular depth estimation from unlabeled\nimage sequences. In particular, we propose three key contributions: (1) We\ntransfer knowledge from features learned during self-supervised depth\nestimation to semantic segmentation, (2) we implement a strong data\naugmentation by blending images and labels using the geometry of the scene, and\n(3) we utilize the depth feature diversity as well as the level of difficulty\nof learning depth in a student-teacher framework to select the most useful\nsamples to be annotated for semantic segmentation. We validate the proposed\nmodel on the Cityscapes dataset, where all three modules demonstrate\nsignificant performance gains, and we achieve state-of-the-art results for\nsemi-supervised semantic segmentation. The implementation is available at\nhttps://github.com/lhoyer/improving_segmentation_with_selfsupervised_depth.",
+                    "authors": "Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian K\u00f6ring, Suman Saha, Luc Van Gool",
+                    "published": "2020-12-19",
+                    "updated": "2021-04-05",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2305.17091v1",
+                    "title": "SSSegmenation: An Open Source Supervised Semantic Segmentation Toolbox Based on PyTorch",
+                    "abstract": "This paper presents SSSegmenation, which is an open source supervised\nsemantic image segmentation toolbox based on PyTorch. The design of this\ntoolbox is motivated by MMSegmentation while it is easier to use because of\nfewer dependencies and achieves superior segmentation performance under a\ncomparable training and testing setup. Moreover, the toolbox also provides\nplenty of trained weights for popular and contemporary semantic segmentation\nmethods, including Deeplab, PSPNet, OCRNet, MaskFormer, \\emph{etc}. We expect\nthat this toolbox can contribute to the future development of semantic\nsegmentation. Codes and model zoos are available at\n\\href{https://github.com/SegmentationBLWX/sssegmentation/}{SSSegmenation}.",
+                    "authors": "Zhenchao Jin",
+                    "published": "2023-05-26",
+                    "updated": "2023-05-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2402.04618v1",
+                    "title": "Multi-Scale Semantic Segmentation with Modified MBConv Blocks",
+                    "abstract": "Recently, MBConv blocks, initially designed for efficiency in\nresource-limited settings and later adapted for cutting-edge image\nclassification performances, have demonstrated significant potential in image\nclassification tasks. Despite their success, their application in semantic\nsegmentation has remained relatively unexplored. This paper introduces a novel\nadaptation of MBConv blocks specifically tailored for semantic segmentation.\nOur modification stems from the insight that semantic segmentation requires the\nextraction of more detailed spatial information than image classification. We\nargue that to effectively perform multi-scale semantic segmentation, each\nbranch of a U-Net architecture, regardless of its resolution, should possess\nequivalent segmentation capabilities. By implementing these changes, our\napproach achieves impressive mean Intersection over Union (IoU) scores of 84.5%\nand 84.0% on the Cityscapes test and validation datasets, respectively,\ndemonstrating the efficacy of our proposed modifications in enhancing semantic\nsegmentation performance.",
+                    "authors": "Xi Chen, Yang Cai, Yuan Wu, Bo Xiong, Taesung Park",
+                    "published": "2024-02-07",
+                    "updated": "2024-02-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2308.11506v1",
+                    "title": "LCCo: Lending CLIP to Co-Segmentation",
+                    "abstract": "This paper studies co-segmenting the common semantic object in a set of\nimages. Existing works either rely on carefully engineered networks to mine the\nimplicit semantic information in visual features or require extra data (i.e.,\nclassification labels) for training. In this paper, we leverage the contrastive\nlanguage-image pre-training framework (CLIP) for the task. With a backbone\nsegmentation network that independently processes each image from the set, we\nintroduce semantics from CLIP into the backbone features, refining them in a\ncoarse-to-fine manner with three key modules: i) an image set feature\ncorrespondence module, encoding global consistent semantic information of the\nimage set; ii) a CLIP interaction module, using CLIP-mined common semantics of\nthe image set to refine the backbone feature; iii) a CLIP regularization\nmodule, drawing CLIP towards this co-segmentation task, identifying the best\nCLIP semantic and using it to regularize the backbone feature. Experiments on\nfour standard co-segmentation benchmark datasets show that the performance of\nour method outperforms state-of-the-art methods.",
+                    "authors": "Xin Duan, Yan Yang, Liyuan Pan, Xiabi Liu",
+                    "published": "2023-08-22",
+                    "updated": "2023-08-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2307.13215v1",
+                    "title": "Image Segmentation Keras : Implementation of Segnet, FCN, UNet, PSPNet and other models in Keras",
+                    "abstract": "Semantic segmentation plays a vital role in computer vision tasks, enabling\nprecise pixel-level understanding of images. In this paper, we present a\ncomprehensive library for semantic segmentation, which contains implementations\nof popular segmentation models like SegNet, FCN, UNet, and PSPNet. We also\nevaluate and compare these models on several datasets, offering researchers and\npractitioners a powerful toolset for tackling diverse segmentation challenges.",
+                    "authors": "Divam Gupta",
+                    "published": "2023-07-25",
+                    "updated": "2023-07-25",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1804.04882v2",
+                    "title": "Learning to Exploit the Prior Network Knowledge for Weakly-Supervised Semantic Segmentation",
+                    "abstract": "Training a Convolutional Neural Network (CNN) for semantic segmentation\ntypically requires to collect a large amount of accurate pixel-level\nannotations, a hard and expensive task. In contrast, simple image tags are\neasier to gather. With this paper we introduce a novel weakly-supervised\nsemantic segmentation model able to learn from image labels, and just image\nlabels. Our model uses the prior knowledge of a network trained for image\nrecognition, employing these image annotations as an attention mechanism to\nidentify semantic regions in the images. We then present a methodology that\nbuilds accurate class-specific segmentation masks from these regions, where\nneither external objectness nor saliency algorithms are required. We describe\nhow to incorporate this mask generation strategy into a fully end-to-end\ntrainable process where the network jointly learns to classify and segment\nimages. Our experiments on PASCAL VOC 2012 dataset show that exploiting these\ngenerated class-specific masks in conjunction with our novel end-to-end\nlearning process outperforms several recent weakly-supervised semantic\nsegmentation methods that use image tags only, and even some models that\nleverage additional supervision or training data.",
+                    "authors": "Carolina Redondo-Cabrera, Marcos Baptista-R\u00edos, Roberto J. L\u00f3pez-Sastre",
+                    "published": "2018-04-13",
+                    "updated": "2019-02-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2102.12095v1",
+                    "title": "Synergy Between Semantic Segmentation and Image Denoising via Alternate Boosting",
+                    "abstract": "The capability of image semantic segmentation may be deteriorated due to\nnoisy input image, where image denoising prior to segmentation helps. Both\nimage denoising and semantic segmentation have been developed significantly\nwith the advance of deep learning. Thus, we are interested in the synergy\nbetween them by using a holistic deep model. We observe that not only denoising\nhelps combat the drop of segmentation accuracy due to noise, but also\npixel-wise semantic information boosts the capability of denoising. We then\npropose a boosting network to perform denoising and segmentation alternately.\nThe proposed network is composed of multiple segmentation and denoising blocks\n(SDBs), each of which estimates semantic map then uses the map to regularize\ndenoising. Experimental results show that the denoised image quality is\nimproved substantially and the segmentation accuracy is improved to close to\nthat of clean images. Our code and models will be made publicly available.",
+                    "authors": "Shunxin Xu, Ke Sun, Dong Liu, Zhiwei Xiong, Zheng-Jun Zha",
+                    "published": "2021-02-24",
+                    "updated": "2021-02-24",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1904.05457v1",
+                    "title": "Instance Segmentation based Semantic Matting for Compositing Applications",
+                    "abstract": "Image compositing is a key step in film making and image editing that aims to\nsegment a foreground object and combine it with a new background. Automatic\nimage compositing can be done easily in a studio using chroma-keying when the\nbackground is pure blue or green. However, image compositing in natural scenes\nwith complex backgrounds remains a tedious task, requiring experienced artists\nto hand-segment. In order to achieve automatic compositing in natural scenes,\nwe propose a fully automated method that integrates instance segmentation and\nimage matting processes to generate high-quality semantic mattes that can be\nused for image editing task. Our approach can be seen both as a refinement of\nexisting instance segmentation algorithms and as a fully automated semantic\nimage matting method. It extends automatic image compositing techniques such as\nchroma-keying to scenes with complex natural backgrounds without the need for\nany kind of user interaction. The output of our approach can be considered as\nboth refined instance segmentations and alpha mattes with semantic meanings. We\nprovide experimental results which show improved performance results as\ncompared to existing approaches.",
+                    "authors": "Guanqing Hu, James J. Clark",
+                    "published": "2019-04-10",
+                    "updated": "2019-04-10",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2003.04404v1",
+                    "title": "FusionLane: Multi-Sensor Fusion for Lane Marking Semantic Segmentation Using Deep Neural Networks",
+                    "abstract": "It is a crucial step to achieve effective semantic segmentation of lane\nmarking during the construction of the lane level high-precision map. In recent\nyears, many image semantic segmentation methods have been proposed. These\nmethods mainly focus on the image from camera, due to the limitation of the\nsensor itself, the accurate three-dimensional spatial position of the lane\nmarking cannot be obtained, so the demand for the lane level high-precision map\nconstruction cannot be met. This paper proposes a lane marking semantic\nsegmentation method based on LIDAR and camera fusion deep neural network.\nDifferent from other methods, in order to obtain accurate position information\nof the segmentation results, the semantic segmentation object of this paper is\na bird's eye view converted from a LIDAR points cloud instead of an image\ncaptured by a camera. This method first uses the deeplabv3+ [\\ref{ref:1}]\nnetwork to segment the image captured by the camera, and the segmentation\nresult is merged with the point clouds collected by the LIDAR as the input of\nthe proposed network. In this neural network, we also add a long short-term\nmemory (LSTM) structure to assist the network for semantic segmentation of lane\nmarkings by using the the time series information. The experiments on more than\n14,000 image datasets which we have manually labeled and expanded have shown\nthe proposed method has better performance on the semantic segmentation of the\npoints cloud bird's eye view. Therefore, the automation of high-precision map\nconstruction can be significantly improved. Our code is available at\nhttps://github.com/rolandying/FusionLane.",
+                    "authors": "Ruochen Yin, Biao Yu, Huapeng Wu, Yutao Song, Runxin Niu",
+                    "published": "2020-03-09",
+                    "updated": "2020-03-09",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.AI"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2106.04108v3",
+                    "title": "Fully Transformer Networks for Semantic Image Segmentation",
+                    "abstract": "Transformers have shown impressive performance in various natural language\nprocessing and computer vision tasks, due to the capability of modeling\nlong-range dependencies. Recent progress has demonstrated that combining such\nTransformers with CNN-based semantic image segmentation models is very\npromising. However, it is not well studied yet on how well a pure Transformer\nbased approach can achieve for image segmentation. In this work, we explore a\nnovel framework for semantic image segmentation, which is encoder-decoder based\nFully Transformer Networks (FTN). Specifically, we first propose a Pyramid\nGroup Transformer (PGT) as the encoder for progressively learning hierarchical\nfeatures, meanwhile reducing the computation complexity of the standard Visual\nTransformer (ViT). Then, we propose a Feature Pyramid Transformer (FPT) to fuse\nsemantic-level and spatial-level information from multiple levels of the PGT\nencoder for semantic image segmentation. Surprisingly, this simple baseline can\nachieve better results on multiple challenging semantic segmentation and face\nparsing benchmarks, including PASCAL Context, ADE20K, COCOStuff, and\nCelebAMask-HQ. The source code will be released on\nhttps://github.com/BR-IDL/PaddleViT.",
+                    "authors": "Sitong Wu, Tianyi Wu, Fangjian Lin, Shengwei Tian, Guodong Guo",
+                    "published": "2021-06-08",
+                    "updated": "2021-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1812.10885v1",
+                    "title": "Coarse-to-fine Semantic Segmentation from Image-level Labels",
+                    "abstract": "Deep neural network-based semantic segmentation generally requires\nlarge-scale cost extensive annotations for training to obtain better\nperformance. To avoid pixel-wise segmentation annotations which are needed for\nmost methods, recently some researchers attempted to use object-level labels\n(e.g. bounding boxes) or image-level labels (e.g. image categories). In this\npaper, we propose a novel recursive coarse-to-fine semantic segmentation\nframework based on only image-level category labels. For each image, an initial\ncoarse mask is first generated by a convolutional neural network-based\nunsupervised foreground segmentation model and then is enhanced by a graph\nmodel. The enhanced coarse mask is fed to a fully convolutional neural network\nto be recursively refined. Unlike existing image-level label-based semantic\nsegmentation methods which require to label all categories for images contain\nmultiple types of objects, our framework only needs one label for each image\nand can handle images contains multi-category objects. With only trained on\nImageNet, our framework achieves comparable performance on PASCAL VOC dataset\nas other image-level label-based state-of-the-arts of semantic segmentation.\nFurthermore, our framework can be easily extended to foreground object\nsegmentation task and achieves comparable performance with the state-of-the-art\nsupervised methods on the Internet Object dataset.",
+                    "authors": "Longlong Jing, Yucheng Chen, Yingli Tian",
+                    "published": "2018-12-28",
+                    "updated": "2018-12-28",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2306.02095v1",
+                    "title": "Content-aware Token Sharing for Efficient Semantic Segmentation with Vision Transformers",
+                    "abstract": "This paper introduces Content-aware Token Sharing (CTS), a token reduction\napproach that improves the computational efficiency of semantic segmentation\nnetworks that use Vision Transformers (ViTs). Existing works have proposed\ntoken reduction approaches to improve the efficiency of ViT-based image\nclassification networks, but these methods are not directly applicable to\nsemantic segmentation, which we address in this work. We observe that, for\nsemantic segmentation, multiple image patches can share a token if they contain\nthe same semantic class, as they contain redundant information. Our approach\nleverages this by employing an efficient, class-agnostic policy network that\npredicts if image patches contain the same semantic class, and lets them share\na token if they do. With experiments, we explore the critical design choices of\nCTS and show its effectiveness on the ADE20K, Pascal Context and Cityscapes\ndatasets, various ViT backbones, and different segmentation decoders. With\nContent-aware Token Sharing, we are able to reduce the number of processed\ntokens by up to 44%, without diminishing the segmentation quality.",
+                    "authors": "Chenyang Lu, Daan de Geus, Gijs Dubbelman",
+                    "published": "2023-06-03",
+                    "updated": "2023-06-03",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/1709.01956v1",
+                    "title": "Learning Dilation Factors for Semantic Segmentation of Street Scenes",
+                    "abstract": "Contextual information is crucial for semantic segmentation. However, finding\nthe optimal trade-off between keeping desired fine details and at the same time\nproviding sufficiently large receptive fields is non trivial. This is even more\nso, when objects or classes present in an image significantly vary in size.\nDilated convolutions have proven valuable for semantic segmentation, because\nthey allow to increase the size of the receptive field without sacrificing\nimage resolution. However, in current state-of-the-art methods, dilation\nparameters are hand-tuned and fixed. In this paper, we present an approach for\nlearning dilation parameters adaptively per channel, consistently improving\nsemantic segmentation results on street-scene datasets like Cityscapes and\nCamvid.",
+                    "authors": "Yang He, Margret Keuper, Bernt Schiele, Mario Fritz",
+                    "published": "2017-09-06",
+                    "updated": "2017-09-06",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2104.10834v1",
+                    "title": "DANNet: A One-Stage Domain Adaptation Network for Unsupervised Nighttime Semantic Segmentation",
+                    "abstract": "Semantic segmentation of nighttime images plays an equally important role as\nthat of daytime images in autonomous driving, but the former is much more\nchallenging due to poor illuminations and arduous human annotations. In this\npaper, we propose a novel domain adaptation network (DANNet) for nighttime\nsemantic segmentation without using labeled nighttime image data. It employs an\nadversarial training with a labeled daytime dataset and an unlabeled dataset\nthat contains coarsely aligned day-night image pairs. Specifically, for the\nunlabeled day-night image pairs, we use the pixel-level predictions of static\nobject categories on a daytime image as a pseudo supervision to segment its\ncounterpart nighttime image. We further design a re-weighting strategy to\nhandle the inaccuracy caused by misalignment between day-night image pairs and\nwrong predictions of daytime images, as well as boost the prediction accuracy\nof small objects. The proposed DANNet is the first one stage adaptation\nframework for nighttime semantic segmentation, which does not train additional\nday-night image transfer models as a separate pre-processing stage. Extensive\nexperiments on Dark Zurich and Nighttime Driving datasets show that our method\nachieves state-of-the-art performance for nighttime semantic segmentation.",
+                    "authors": "Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, Song Wang",
+                    "published": "2021-04-22",
+                    "updated": "2021-04-22",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "category": "Semantic AND Segmentation AND Image"
+                }
+            ]
+        ]
+    },
+    {
+        "url": "http://arxiv.org/abs/2308.14604v3",
+        "title": "SAM-PARSER: Fine-tuning SAM Efficiently by Parameter Space Reconstruction",
+        "abstract": "Segment Anything Model (SAM) has received remarkable attention as it offers a\npowerful and versatile solution for object segmentation in images. However,\nfine-tuning SAM for downstream segmentation tasks under different scenarios\nremains a challenge, as the varied characteristics of different scenarios\nnaturally requires diverse model parameter spaces. Most existing fine-tuning\nmethods attempt to bridge the gaps among different scenarios by introducing a\nset of new parameters to modify SAM's original parameter space. Unlike these\nworks, in this paper, we propose fine-tuning SAM efficiently by parameter space\nreconstruction (SAM-PARSER), which introduce nearly zero trainable parameters\nduring fine-tuning. In SAM-PARSER, we assume that SAM's original parameter\nspace is relatively complete, so that its bases are able to reconstruct the\nparameter space of a new scenario. We obtain the bases by matrix decomposition,\nand fine-tuning the coefficients to reconstruct the parameter space tailored to\nthe new scenario by an optimal linear combination of the bases. Experimental\nresults show that SAM-PARSER exhibits superior segmentation performance across\nvarious scenarios, while reducing the number of trainable parameters by\n$\\approx 290$ times compared with current parameter-efficient fine-tuning\nmethods.",
+        "authors": "Zelin Peng, Zhengqin Xu, Zhilin Zeng, Xiaokang Yang, Wei Shen",
+        "published": "2023-08-28",
+        "updated": "2023-12-18",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2306.13731v1",
+        "title": "How to Efficiently Adapt Large Segmentation Model(SAM) to Medical Images",
+        "abstract": "The emerging scale segmentation model, Segment Anything (SAM), exhibits\nimpressive capabilities in zero-shot segmentation for natural images. However,\nwhen applied to medical images, SAM suffers from noticeable performance drop.\nTo make SAM a real ``foundation model\" for the computer vision community, it is\ncritical to find an efficient way to customize SAM for medical image dataset.\nIn this work, we propose to freeze SAM encoder and finetune a lightweight\ntask-specific prediction head, as most of weights in SAM are contributed by the\nencoder. In addition, SAM is a promptable model, while prompt is not\nnecessarily available in all application cases, and precise prompts for\nmultiple class segmentation are also time-consuming. Therefore, we explore\nthree types of prompt-free prediction heads in this work, include ViT, CNN, and\nlinear layers. For ViT head, we remove the prompt tokens in the mask decoder of\nSAM, which is named AutoSAM. AutoSAM can also generate masks for different\nclasses with one single inference after modification. To evaluate the\nlabel-efficiency of our finetuning method, we compare the results of these\nthree prediction heads on a public medical image segmentation dataset with\nlimited labeled data. Experiments demonstrate that finetuning SAM significantly\nimproves its performance on medical image dataset, even with just one labeled\nvolume. Moreover, AutoSAM and CNN prediction head also has better segmentation\naccuracy than training from scratch and self-supervised learning approaches\nwhen there is a shortage of annotations.",
+        "authors": "Xinrong Hu, Xiaowei Xu, Yiyu Shi",
+        "published": "2023-06-23",
+        "updated": "2023-06-23",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2310.07183v2",
+        "title": "SAM-OCTA: Prompting Segment-Anything for OCTA Image Segmentation",
+        "abstract": "Segmenting specific targets or biomarkers is necessary to analyze optical\ncoherence tomography angiography (OCTA) images. Previous methods typically\nsegment all the targets in an OCTA sample, such as retinal vessels (RVs).\nAlthough these methods perform well in accuracy and precision, OCTA analyses\noften focusing local information within the images which has not been\nfulfilled. In this paper, we propose a method called SAM-OCTA for local\nsegmentation in OCTA images. The method fine-tunes a pre-trained segment\nanything model (SAM) using low-rank adaptation (LoRA) and utilizes prompt\npoints for local RVs, arteries, and veins segmentation in OCTA. To explore the\neffect and mechanism of prompt points, we set up global and local segmentation\nmodes with two prompt point generation strategies, namely random selection and\nspecial annotation. Considering practical usage, we conducted extended\nexperiments with different model scales and analyzed the model performance\nbefore and after fine-tuning besides the general segmentation task. From\ncomprehensive experimental results with the OCTA-500 dataset, our SAM-OCTA\nmethod has achieved state-of-the-art performance in common OCTA segmentation\ntasks related to RV and FAZ, and it also performs accurate segmentation of\nartery-vein and local vessels. The code is available at\nhttps://github.com/ShellRedia/SAM-OCTA-extend.",
+        "authors": "Xinrun Chen, Chengliang Wang, Haojian Ning, Shiying Li, Mei Shen",
+        "published": "2023-10-11",
+        "updated": "2024-03-20",
+        "primary_cat": "cs.LG",
+        "cats": [
+            "cs.LG"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2305.00035v1",
+        "title": "SAM on Medical Images: A Comprehensive Study on Three Prompt Modes",
+        "abstract": "The Segment Anything Model (SAM) made an eye-catching debut recently and\ninspired many researchers to explore its potential and limitation in terms of\nzero-shot generalization capability. As the first promptable foundation model\nfor segmentation tasks, it was trained on a large dataset with an unprecedented\nnumber of images and annotations. This large-scale dataset and its promptable\nnature endow the model with strong zero-shot generalization. Although the SAM\nhas shown competitive performance on several datasets, we still want to\ninvestigate its zero-shot generalization on medical images. As we know, the\nacquisition of medical image annotation usually requires a lot of effort from\nprofessional practitioners. Therefore, if there exists a foundation model that\ncan give high-quality mask prediction simply based on a few point prompts, this\nmodel will undoubtedly become the game changer for medical image analysis. To\nevaluate whether SAM has the potential to become the foundation model for\nmedical image segmentation tasks, we collected more than 12 public medical\nimage datasets that cover various organs and modalities. We also explore what\nkind of prompt can lead to the best zero-shot performance with different\nmodalities. Furthermore, we find that a pattern shows that the perturbation of\nthe box size will significantly change the prediction accuracy. Finally,\nExtensive experiments show that the predicted mask quality varied a lot among\ndifferent datasets. And providing proper prompts, such as bounding boxes, to\nthe SAM will significantly increase its performance.",
+        "authors": "Dongjie Cheng, Ziyuan Qin, Zekun Jiang, Shaoting Zhang, Qicheng Lao, Kang Li",
+        "published": "2023-04-28",
+        "updated": "2023-04-28",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2304.13785v2",
+        "title": "Customized Segment Anything Model for Medical Image Segmentation",
+        "abstract": "We propose SAMed, a general solution for medical image segmentation.\nDifferent from the previous methods, SAMed is built upon the large-scale image\nsegmentation model, Segment Anything Model (SAM), to explore the new research\nparadigm of customizing large-scale models for medical image segmentation.\nSAMed applies the low-rank-based (LoRA) finetuning strategy to the SAM image\nencoder and finetunes it together with the prompt encoder and the mask decoder\non labeled medical image segmentation datasets. We also observe the warmup\nfinetuning strategy and the AdamW optimizer lead SAMed to successful\nconvergence and lower loss. Different from SAM, SAMed could perform semantic\nsegmentation on medical images. Our trained SAMed model achieves 81.88 DSC and\n20.64 HD on the Synapse multi-organ segmentation dataset, which is on par with\nthe state-of-the-art methods. We conduct extensive experiments to validate the\neffectiveness of our design. Since SAMed only updates a small fraction of the\nSAM parameters, its deployment cost and storage cost are quite marginal in\npractical usage. The code of SAMed is available at\nhttps://github.com/hitachinsk/SAMed.",
+        "authors": "Kaidong Zhang, Dong Liu",
+        "published": "2023-04-26",
+        "updated": "2023-10-17",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/1810.04805v2",
+        "title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
+        "abstract": "We introduce a new language representation model called BERT, which stands\nfor Bidirectional Encoder Representations from Transformers. Unlike recent\nlanguage representation models, BERT is designed to pre-train deep\nbidirectional representations from unlabeled text by jointly conditioning on\nboth left and right context in all layers. As a result, the pre-trained BERT\nmodel can be fine-tuned with just one additional output layer to create\nstate-of-the-art models for a wide range of tasks, such as question answering\nand language inference, without substantial task-specific architecture\nmodifications.\n  BERT is conceptually simple and empirically powerful. It obtains new\nstate-of-the-art results on eleven natural language processing tasks, including\npushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI\naccuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering\nTest F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1\n(5.1 point absolute improvement).",
+        "authors": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova",
+        "published": "2018-10-11",
+        "updated": "2019-05-24",
+        "primary_cat": "cs.CL",
+        "cats": [
+            "cs.CL"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2310.18642v1",
+        "title": "One-shot Localization and Segmentation of Medical Images with Foundation Models",
+        "abstract": "Recent advances in Vision Transformers (ViT) and Stable Diffusion (SD) models\nwith their ability to capture rich semantic features of the image have been\nused for image correspondence tasks on natural images. In this paper, we\nexamine the ability of a variety of pre-trained ViT (DINO, DINOv2, SAM, CLIP)\nand SD models, trained exclusively on natural images, for solving the\ncorrespondence problems on medical images. While many works have made a case\nfor in-domain training, we show that the models trained on natural images can\noffer good performance on medical images across different modalities\n(CT,MR,Ultrasound) sourced from various manufacturers, over multiple anatomical\nregions (brain, thorax, abdomen, extremities), and on wide variety of tasks.\nFurther, we leverage the correspondence with respect to a template image to\nprompt a Segment Anything (SAM) model to arrive at single shot segmentation,\nachieving dice range of 62%-90% across tasks, using just one image as\nreference. We also show that our single-shot method outperforms the recently\nproposed few-shot segmentation method - UniverSeg (Dice range 47%-80%) on most\nof the semantic segmentation tasks(six out of seven) across medical imaging\nmodalities.",
+        "authors": "Deepa Anand, Gurunath Reddy M, Vanika Singhal, Dattesh D. Shanbhag, Shriram KS, Uday Patil, Chitresh Bhushan, Kavitha Manickam, Dawei Gui, Rakesh Mullick, Avinash Gopal, Parminder Bhatia, Taha Kass-Hout",
+        "published": "2023-10-28",
+        "updated": "2023-10-28",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2304.09324v3",
+        "title": "Computer-Vision Benchmark Segment-Anything Model (SAM) in Medical Images: Accuracy in 12 Datasets",
+        "abstract": "Background: The segment-anything model (SAM), introduced in April 2023, shows\npromise as a benchmark model and a universal solution to segment various\nnatural images. It comes without previously-required re-training or fine-tuning\nspecific to each new dataset.\n  Purpose: To test SAM's accuracy in various medical image segmentation tasks\nand investigate potential factors that may affect its accuracy in medical\nimages.\n  Methods: SAM was tested on 12 public medical image segmentation datasets\ninvolving 7,451 subjects. The accuracy was measured by the Dice overlap between\nthe algorithm-segmented and ground-truth masks. SAM was compared with five\nstate-of-the-art algorithms specifically designed for medical image\nsegmentation tasks. Associations of SAM's accuracy with six factors were\ncomputed, independently and jointly, including segmentation difficulties as\nmeasured by segmentation ability score and by Dice overlap in U-Net, image\ndimension, size of the target region, image modality, and contrast.\n  Results: The Dice overlaps from SAM were significantly lower than the five\nmedical-image-based algorithms in all 12 medical image segmentation datasets,\nby a margin of 0.1-0.5 and even 0.6-0.7 Dice. SAM-Semantic was significantly\nassociated with medical image segmentation difficulty and the image modality,\nand SAM-Point and SAM-Box were significantly associated with image segmentation\ndifficulty, image dimension, target region size, and target-vs-background\ncontrast. All these 3 variations of SAM were more accurate in 2D medical\nimages, larger target region sizes, easier cases with a higher Segmentation\nAbility score and higher U-Net Dice, and higher foreground-background contrast.",
+        "authors": "Sheng He, Rina Bao, Jingpeng Li, Jeffrey Stout, Atle Bjornerud, P. Ellen Grant, Yangming Ou",
+        "published": "2023-04-18",
+        "updated": "2023-05-05",
+        "primary_cat": "eess.IV",
+        "cats": [
+            "eess.IV",
+            "cs.CV"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2010.11929v2",
+        "title": "An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale",
+        "abstract": "While the Transformer architecture has become the de-facto standard for\nnatural language processing tasks, its applications to computer vision remain\nlimited. In vision, attention is either applied in conjunction with\nconvolutional networks, or used to replace certain components of convolutional\nnetworks while keeping their overall structure in place. We show that this\nreliance on CNNs is not necessary and a pure transformer applied directly to\nsequences of image patches can perform very well on image classification tasks.\nWhen pre-trained on large amounts of data and transferred to multiple mid-sized\nor small image recognition benchmarks (ImageNet, CIFAR-100, VTAB, etc.), Vision\nTransformer (ViT) attains excellent results compared to state-of-the-art\nconvolutional networks while requiring substantially fewer computational\nresources to train.",
+        "authors": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby",
+        "published": "2020-10-22",
+        "updated": "2021-06-03",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI",
+            "cs.LG"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2308.14133v1",
+        "title": "Cheap Lunch for Medical Image Segmentation by Fine-tuning SAM on Few Exemplars",
+        "abstract": "The Segment Anything Model (SAM) has demonstrated remarkable capabilities of\nscaled-up segmentation models, enabling zero-shot generalization across a\nvariety of domains. By leveraging large-scale foundational models as\npre-trained models, it is a natural progression to fine-tune SAM for specific\ndomains to further enhance performances. However, the adoption of foundational\nmodels in the medical domain presents a challenge due to the difficulty and\nexpense of labeling sufficient data for adaptation within hospital systems. In\nthis paper, we introduce an efficient and practical approach for fine-tuning\nSAM using a limited number of exemplars, making it suitable for such scenarios.\nOur approach combines two established techniques from the literature: an\nexemplar-guided synthesis module and the widely recognized Low-Rank Adaptation\n(LoRA) fine-tuning strategy, serving as data-level and model-level attempts\nrespectively. Interestingly, our empirical findings suggest that SAM can be\neffectively aligned within the medical domain even with few labeled data. We\nvalidate our approach through experiments on brain tumor segmentation (BraTS)\nand multi-organ CT segmentation (Synapse). The comprehensive results underscore\nthe feasibility and effectiveness of such an approach, paving the way for the\npractical application of SAM in the medical domain.",
+        "authors": "Weijia Feng, Lingting Zhu, Lequan Yu",
+        "published": "2023-08-27",
+        "updated": "2023-08-27",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2102.12092v2",
+        "title": "Zero-Shot Text-to-Image Generation",
+        "abstract": "Text-to-image generation has traditionally focused on finding better modeling\nassumptions for training on a fixed dataset. These assumptions might involve\ncomplex architectures, auxiliary losses, or side information such as object\npart labels or segmentation masks supplied during training. We describe a\nsimple approach for this task based on a transformer that autoregressively\nmodels the text and image tokens as a single stream of data. With sufficient\ndata and scale, our approach is competitive with previous domain-specific\nmodels when evaluated in a zero-shot fashion.",
+        "authors": "Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, Ilya Sutskever",
+        "published": "2021-02-24",
+        "updated": "2021-02-26",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.LG"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2304.12306v3",
+        "title": "Segment Anything in Medical Images",
+        "abstract": "Medical image segmentation is a critical component in clinical practice,\nfacilitating accurate diagnosis, treatment planning, and disease monitoring.\nHowever, existing methods, often tailored to specific modalities or disease\ntypes, lack generalizability across the diverse spectrum of medical image\nsegmentation tasks. Here we present MedSAM, a foundation model designed for\nbridging this gap by enabling universal medical image segmentation. The model\nis developed on a large-scale medical image dataset with 1,570,263 image-mask\npairs, covering 10 imaging modalities and over 30 cancer types. We conduct a\ncomprehensive evaluation on 86 internal validation tasks and 60 external\nvalidation tasks, demonstrating better accuracy and robustness than\nmodality-wise specialist models. By delivering accurate and efficient\nsegmentation across a wide spectrum of tasks, MedSAM holds significant\npotential to expedite the evolution of diagnostic tools and the personalization\nof treatment plans.",
+        "authors": "Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, Bo Wang",
+        "published": "2023-04-24",
+        "updated": "2024-04-01",
+        "primary_cat": "eess.IV",
+        "cats": [
+            "eess.IV",
+            "cs.CV"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2308.14936v2",
+        "title": "Auto-Prompting SAM for Mobile Friendly 3D Medical Image Segmentation",
+        "abstract": "Segment Anything Model (SAM) has rapidly been adopted for segmenting a wide\nrange of natural images. However, recent studies have indicated that SAM\nexhibits subpar performance on 3D medical image segmentation tasks. In addition\nto the domain gaps between natural and medical images, disparities in the\nspatial arrangement between 2D and 3D images, the substantial computational\nburden imposed by powerful GPU servers, and the time-consuming manual prompt\ngeneration impede the extension of SAM to a broader spectrum of medical image\nsegmentation applications. To mitigate these challenges, we introduce a novel\nmethod, AutoSAM Adapter, designed specifically for 3D multi-organ CT-based\nsegmentation. This approach utilizes parameter-efficient adaptation techniques\nand an automatic prompt learning paradigm, transforming SAM's capabilities for\n3D medical image segmentation. It eliminates the need for manual prompts and\nachieves SOTA performance in CT-based multi-organ segmentation tasks.\nFurthermore, we successfully transfer the acquired knowledge of the AutoSAM\nAdapter to other lightweight models tailored for 3D medical image analysis with\nenhanced performance. Through extensive experiments, the AutoSAM Adapter has\nbeen demonstrated as an effective method to adapt the foundational SAM-based 2D\nnatural image segmentation model for 3D medical image segmentation tasks.",
+        "authors": "Chengyin Li, Prashant Khanduri, Yao Qiang, Rafi Ibn Sultan, Indrin Chetty, Dongxiao Zhu",
+        "published": "2023-08-28",
+        "updated": "2023-11-21",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2304.10517v3",
+        "title": "Segment Anything Model for Medical Image Analysis: an Experimental Study",
+        "abstract": "Training segmentation models for medical images continues to be challenging\ndue to the limited availability of data annotations. Segment Anything Model\n(SAM) is a foundation model that is intended to segment user-defined objects of\ninterest in an interactive manner. While the performance on natural images is\nimpressive, medical image domains pose their own set of challenges. Here, we\nperform an extensive evaluation of SAM's ability to segment medical images on a\ncollection of 19 medical imaging datasets from various modalities and\nanatomies. We report the following findings: (1) SAM's performance based on\nsingle prompts highly varies depending on the dataset and the task, from\nIoU=0.1135 for spine MRI to IoU=0.8650 for hip X-ray. (2) Segmentation\nperformance appears to be better for well-circumscribed objects with prompts\nwith less ambiguity and poorer in various other scenarios such as the\nsegmentation of brain tumors. (3) SAM performs notably better with box prompts\nthan with point prompts. (4) SAM outperforms similar methods RITM, SimpleClick,\nand FocalClick in almost all single-point prompt settings. (5) When\nmultiple-point prompts are provided iteratively, SAM's performance generally\nimproves only slightly while other methods' performance improves to the level\nthat surpasses SAM's point-based performance. We also provide several\nillustrations for SAM's performance on all tested datasets, iterative\nsegmentation, and SAM's behavior given prompt ambiguity. We conclude that SAM\nshows impressive zero-shot segmentation performance for certain medical imaging\ndatasets, but moderate to poor performance for others. SAM has the potential to\nmake a significant impact in automated medical image segmentation in medical\nimaging, but appropriate care needs to be applied when using it.",
+        "authors": "Maciej A. Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, Yixin Zhang",
+        "published": "2023-04-20",
+        "updated": "2023-05-17",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI",
+            "cs.LG"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2106.09685v2",
+        "title": "LoRA: Low-Rank Adaptation of Large Language Models",
+        "abstract": "An important paradigm of natural language processing consists of large-scale\npre-training on general domain data and adaptation to particular tasks or\ndomains. As we pre-train larger models, full fine-tuning, which retrains all\nmodel parameters, becomes less feasible. Using GPT-3 175B as an example --\ndeploying independent instances of fine-tuned models, each with 175B\nparameters, is prohibitively expensive. We propose Low-Rank Adaptation, or\nLoRA, which freezes the pre-trained model weights and injects trainable rank\ndecomposition matrices into each layer of the Transformer architecture, greatly\nreducing the number of trainable parameters for downstream tasks. Compared to\nGPT-3 175B fine-tuned with Adam, LoRA can reduce the number of trainable\nparameters by 10,000 times and the GPU memory requirement by 3 times. LoRA\nperforms on-par or better than fine-tuning in model quality on RoBERTa,\nDeBERTa, GPT-2, and GPT-3, despite having fewer trainable parameters, a higher\ntraining throughput, and, unlike adapters, no additional inference latency. We\nalso provide an empirical investigation into rank-deficiency in language model\nadaptation, which sheds light on the efficacy of LoRA. We release a package\nthat facilitates the integration of LoRA with PyTorch models and provide our\nimplementations and model checkpoints for RoBERTa, DeBERTa, and GPT-2 at\nhttps://github.com/microsoft/LoRA.",
+        "authors": "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen",
+        "published": "2021-06-17",
+        "updated": "2021-10-16",
+        "primary_cat": "cs.CL",
+        "cats": [
+            "cs.CL",
+            "cs.AI",
+            "cs.LG"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/1706.05587v3",
+        "title": "Rethinking Atrous Convolution for Semantic Image Segmentation",
+        "abstract": "In this work, we revisit atrous convolution, a powerful tool to explicitly\nadjust filter's field-of-view as well as control the resolution of feature\nresponses computed by Deep Convolutional Neural Networks, in the application of\nsemantic image segmentation. To handle the problem of segmenting objects at\nmultiple scales, we design modules which employ atrous convolution in cascade\nor in parallel to capture multi-scale context by adopting multiple atrous\nrates. Furthermore, we propose to augment our previously proposed Atrous\nSpatial Pyramid Pooling module, which probes convolutional features at multiple\nscales, with image-level features encoding global context and further boost\nperformance. We also elaborate on implementation details and share our\nexperience on training our system. The proposed `DeepLabv3' system\nsignificantly improves over our previous DeepLab versions without DenseCRF\npost-processing and attains comparable performance with other state-of-art\nmodels on the PASCAL VOC 2012 semantic image segmentation benchmark.",
+        "authors": "Liang-Chieh Chen, George Papandreou, Florian Schroff, Hartwig Adam",
+        "published": "2017-06-17",
+        "updated": "2017-12-05",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "label": "Related Work"
+    },
+    {
+        "url": "http://arxiv.org/abs/2012.10122v2",
+        "title": "Weakly-supervised Semantic Segmentation in Cityscape via Hyperspectral Image",
+        "abstract": "High-resolution hyperspectral images (HSIs) contain the response of each\npixel in different spectral bands, which can be used to effectively distinguish\nvarious objects in complex scenes. While HSI cameras have become low cost,\nalgorithms based on it have not been well exploited. In this paper, we focus on\na novel topic, weakly-supervised semantic segmentation in cityscape via HSIs.\nIt is based on the idea that high-resolution HSIs in city scenes contain rich\nspectral information, which can be easily associated to semantics without\nmanual labeling. Therefore, it enables low cost, highly reliable semantic\nsegmentation in complex scenes. Specifically, in this paper, we theoretically\nanalyze the HSIs and introduce a weakly-supervised HSI semantic segmentation\nframework, which utilizes spectral information to improve the coarse labels to\na finer degree. The experimental results show that our method can obtain highly\ncompetitive labels and even have higher edge fineness than artificial fine\nlabels in some classes. At the same time, the results also show that the\nrefined labels can effectively improve the effect of semantic segmentation. The\ncombination of HSIs and semantic segmentation proves that HSIs have great\npotential in high-level visual tasks.",
+        "authors": "Yuxing Huang, Shaodi You, Ying Fu, Qiu Shen",
+        "published": "2020-12-18",
+        "updated": "2021-07-20",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2210.02270v1",
+        "title": "Weak-shot Semantic Segmentation via Dual Similarity Transfer",
+        "abstract": "Semantic segmentation is an important and prevalent task, but severely\nsuffers from the high cost of pixel-level annotations when extending to more\nclasses in wider applications. To this end, we focus on the problem named\nweak-shot semantic segmentation, where the novel classes are learnt from\ncheaper image-level labels with the support of base classes having\noff-the-shelf pixel-level labels. To tackle this problem, we propose SimFormer,\nwhich performs dual similarity transfer upon MaskFormer. Specifically,\nMaskFormer disentangles the semantic segmentation task into two sub-tasks:\nproposal classification and proposal segmentation for each proposal. Proposal\nsegmentation allows proposal-pixel similarity transfer from base classes to\nnovel classes, which enables the mask learning of novel classes. We also learn\npixel-pixel similarity from base classes and distill such class-agnostic\nsemantic similarity to the semantic masks of novel classes, which regularizes\nthe segmentation model with pixel-level semantic relationship across images. In\naddition, we propose a complementary loss to facilitate the learning of novel\nclasses. Comprehensive experiments on the challenging COCO-Stuff-10K and ADE20K\ndatasets demonstrate the effectiveness of our method. Codes are available at\nhttps://github.com/bcmi/SimFormer-Weak-Shot-Semantic-Segmentation.",
+        "authors": "Junjie Chen, Li Niu, Siyuan Zhou, Jianlou Si, Chen Qian, Liqing Zhang",
+        "published": "2022-10-05",
+        "updated": "2022-10-05",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2009.12232v4",
+        "title": "From Pixel to Patch: Synthesize Context-aware Features for Zero-shot Semantic Segmentation",
+        "abstract": "Zero-shot learning has been actively studied for image classification task to\nrelieve the burden of annotating image labels. Interestingly, semantic\nsegmentation task requires more labor-intensive pixel-wise annotation, but\nzero-shot semantic segmentation has only attracted limited research interest.\nThus, we focus on zero-shot semantic segmentation, which aims to segment unseen\nobjects with only category-level semantic representations provided for unseen\ncategories. In this paper, we propose a novel Context-aware feature Generation\nNetwork (CaGNet), which can synthesize context-aware pixel-wise visual features\nfor unseen categories based on category-level semantic representations and\npixel-wise contextual information. The synthesized features are used to\nfinetune the classifier to enable segmenting unseen objects. Furthermore, we\nextend pixel-wise feature generation and finetuning to patch-wise feature\ngeneration and finetuning, which additionally considers inter-pixel\nrelationship. Experimental results on Pascal-VOC, Pascal-Context, and\nCOCO-stuff show that our method significantly outperforms the existing\nzero-shot semantic segmentation methods. Code is available at\nhttps://github.com/bcmi/CaGNetv2-Zero-Shot-Semantic-Segmentation.",
+        "authors": "Zhangxuan Gu, Siyuan Zhou, Li Niu, Zihan Zhao, Liqing Zhang",
+        "published": "2020-09-25",
+        "updated": "2022-01-21",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2111.11103v1",
+        "title": "Improving Semantic Image Segmentation via Label Fusion in Semantically Textured Meshes",
+        "abstract": "Models for semantic segmentation require a large amount of hand-labeled\ntraining data which is costly and time-consuming to produce. For this purpose,\nwe present a label fusion framework that is capable of improving semantic pixel\nlabels of video sequences in an unsupervised manner. We make use of a 3D mesh\nrepresentation of the environment and fuse the predictions of different frames\ninto a consistent representation using semantic mesh textures. Rendering the\nsemantic mesh using the original intrinsic and extrinsic camera parameters\nyields a set of improved semantic segmentation images. Due to our optimized\nCUDA implementation, we are able to exploit the entire $c$-dimensional\nprobability distribution of annotations over $c$ classes in an\nuncertainty-aware manner. We evaluate our method on the Scannet dataset where\nwe improve annotations produced by the state-of-the-art segmentation network\nESANet from $52.05 \\%$ to $58.25 \\%$ pixel accuracy. We publish the source code\nof our framework online to foster future research in this area\n(\\url{https://github.com/fferflo/semantic-meshes}). To the best of our\nknowledge, this is the first publicly available label fusion framework for\nsemantic image segmentation based on meshes with semantic textures.",
+        "authors": "Florian Fervers, Timo Breuer, Gregor Stachowiak, Sebastian Bullinger, Christoph Bodensteiner, Michael Arens",
+        "published": "2021-11-22",
+        "updated": "2021-11-22",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1705.09052v3",
+        "title": "Weakly Supervised Semantic Segmentation Based on Web Image Co-segmentation",
+        "abstract": "Training a Fully Convolutional Network (FCN) for semantic segmentation\nrequires a large number of masks with pixel level labelling, which involves a\nlarge amount of human labour and time for annotation. In contrast, web images\nand their image-level labels are much easier and cheaper to obtain. In this\nwork, we propose a novel method for weakly supervised semantic segmentation\nwith only image-level labels. The method utilizes the internet to retrieve a\nlarge number of images and uses a large scale co-segmentation framework to\ngenerate masks for the retrieved images. We first retrieve images from search\nengines, e.g. Flickr and Google, using semantic class names as queries, e.g.\nclass names in the dataset PASCAL VOC 2012. We then use high quality masks\nproduced by co-segmentation on the retrieved images as well as the target\ndataset images with image level labels to train segmentation networks. We\nobtain an IoU score of 56.9 on test set of PASCAL VOC 2012, which reaches the\nstate-of-the-art performance.",
+        "authors": "Tong Shen, Guosheng Lin, Lingqiao Liu, Chunhua Shen, Ian Reid",
+        "published": "2017-05-25",
+        "updated": "2017-08-06",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2309.05840v2",
+        "title": "Self-Correlation and Cross-Correlation Learning for Few-Shot Remote Sensing Image Semantic Segmentation",
+        "abstract": "Remote sensing image semantic segmentation is an important problem for remote\nsensing image interpretation. Although remarkable progress has been achieved,\nexisting deep neural network methods suffer from the reliance on massive\ntraining data. Few-shot remote sensing semantic segmentation aims at learning\nto segment target objects from a query image using only a few annotated support\nimages of the target class. Most existing few-shot learning methods stem\nprimarily from their sole focus on extracting information from support images,\nthereby failing to effectively address the large variance in appearance and\nscales of geographic objects. To tackle these challenges, we propose a\nSelf-Correlation and Cross-Correlation Learning Network for the few-shot remote\nsensing image semantic segmentation. Our model enhances the generalization by\nconsidering both self-correlation and cross-correlation between support and\nquery images to make segmentation predictions. To further explore the\nself-correlation with the query image, we propose to adopt a classical spectral\nmethod to produce a class-agnostic segmentation mask based on the basic visual\ninformation of the image. Extensive experiments on two remote sensing image\ndatasets demonstrate the effectiveness and superiority of our model in few-shot\nremote sensing image semantic segmentation. Code and models will be accessed at\nhttps://github.com/linhanwang/SCCNet.",
+        "authors": "Linhan Wang, Shuo Lei, Jianfeng He, Shengkun Wang, Min Zhang, Chang-Tien Lu",
+        "published": "2023-09-11",
+        "updated": "2023-09-15",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2211.08352v1",
+        "title": "Visual Semantic Segmentation Based on Few/Zero-Shot Learning: An Overview",
+        "abstract": "Visual semantic segmentation aims at separating a visual sample into diverse\nblocks with specific semantic attributes and identifying the category for each\nblock, and it plays a crucial role in environmental perception. Conventional\nlearning-based visual semantic segmentation approaches count heavily on\nlarge-scale training data with dense annotations and consistently fail to\nestimate accurate semantic labels for unseen categories. This obstruction spurs\na craze for studying visual semantic segmentation with the assistance of\nfew/zero-shot learning. The emergence and rapid progress of few/zero-shot\nvisual semantic segmentation make it possible to learn unseen-category from a\nfew labeled or zero-labeled samples, which advances the extension to practical\napplications. Therefore, this paper focuses on the recently published\nfew/zero-shot visual semantic segmentation methods varying from 2D to 3D space\nand explores the commonalities and discrepancies of technical settlements under\ndifferent segmentation circumstances. Specifically, the preliminaries on\nfew/zero-shot visual semantic segmentation, including the problem definitions,\ntypical datasets, and technical remedies, are briefly reviewed and discussed.\nMoreover, three typical instantiations are involved to uncover the interactions\nof few/zero-shot learning with visual semantic segmentation, including image\nsemantic segmentation, video object segmentation, and 3D segmentation. Finally,\nthe future challenges of few/zero-shot visual semantic segmentation are\ndiscussed.",
+        "authors": "Wenqi Ren, Yang Tang, Qiyu Sun, Chaoqiang Zhao, Qing-Long Han",
+        "published": "2022-11-13",
+        "updated": "2022-11-13",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2207.08455v3",
+        "title": "Open-world Semantic Segmentation via Contrasting and Clustering Vision-Language Embedding",
+        "abstract": "To bridge the gap between supervised semantic segmentation and real-world\napplications that acquires one model to recognize arbitrary new concepts,\nrecent zero-shot segmentation attracts a lot of attention by exploring the\nrelationships between unseen and seen object categories, yet requiring large\namounts of densely-annotated data with diverse base classes. In this paper, we\npropose a new open-world semantic segmentation pipeline that makes the first\nattempt to learn to segment semantic objects of various open-world categories\nwithout any efforts on dense annotations, by purely exploiting the\nimage-caption data that naturally exist on the Internet. Our method,\nVision-language-driven Semantic Segmentation (ViL-Seg), employs an image and a\ntext encoder to generate visual and text embeddings for the image-caption data,\nwith two core components that endow its segmentation ability: First, the image\nencoder is jointly trained with a vision-based contrasting and a cross-modal\ncontrasting, which encourage the visual embeddings to preserve both\nfine-grained semantics and high-level category information that are crucial for\nthe segmentation task. Furthermore, an online clustering head is devised over\nthe image encoder, which allows to dynamically segment the visual embeddings\ninto distinct semantic groups such that they can be classified by comparing\nwith various text embeddings to complete our segmentation pipeline. Experiments\nshow that without using any data with dense annotations, our method can\ndirectly segment objects of arbitrary categories, outperforming zero-shot\nsegmentation methods that require data labeling on three benchmark datasets.",
+        "authors": "Quande Liu, Youpeng Wen, Jianhua Han, Chunjing Xu, Hang Xu, Xiaodan Liang",
+        "published": "2022-07-18",
+        "updated": "2022-10-29",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2203.15943v1",
+        "title": "Self-Supervised Leaf Segmentation under Complex Lighting Conditions",
+        "abstract": "As an essential prerequisite task in image-based plant phenotyping, leaf\nsegmentation has garnered increasing attention in recent years. While\nself-supervised learning is emerging as an effective alternative to various\ncomputer vision tasks, its adaptation for image-based plant phenotyping remains\nrather unexplored. In this work, we present a self-supervised leaf segmentation\nframework consisting of a self-supervised semantic segmentation model, a\ncolor-based leaf segmentation algorithm, and a self-supervised color correction\nmodel. The self-supervised semantic segmentation model groups the semantically\nsimilar pixels by iteratively referring to the self-contained information,\nallowing the pixels of the same semantic object to be jointly considered by the\ncolor-based leaf segmentation algorithm for identifying the leaf regions.\nAdditionally, we propose to use a self-supervised color correction model for\nimages taken under complex illumination conditions. Experimental results on\ndatasets of different plant species demonstrate the potential of the proposed\nself-supervised framework in achieving effective and generalizable leaf\nsegmentation.",
+        "authors": "Xufeng Lin, Chang-Tsun Li, Scott Adams, Abbas Kouzani, Richard Jiang, Ligang He, Yongjian Hu, Michael Vernon, Egan Doeven, Lawrence Webb, Todd Mcclellan, Adam Guskic",
+        "published": "2022-03-29",
+        "updated": "2022-03-29",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "eess.IV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1502.04983v1",
+        "title": "Context Tricks for Cheap Semantic Segmentation",
+        "abstract": "Accurate semantic labeling of image pixels is difficult because intra-class\nvariability is often greater than inter-class variability. In turn, fast\nsemantic segmentation is hard because accurate models are usually too\ncomplicated to also run quickly at test-time. Our experience with building and\nrunning semantic segmentation systems has also shown a reasonably obvious\nbottleneck on model complexity, imposed by small training datasets. We\ntherefore propose two simple complementary strategies that leverage context to\ngive better semantic segmentation, while scaling up or down to train on\ndifferent-sized datasets.\n  As easy modifications for existing semantic segmentation algorithms, we\nintroduce Decorrelated Semantic Texton Forests, and the Context Sensitive Image\nLevel Prior. The proposed modifications are tested using a Semantic Texton\nForest (STF) system, and the modifications are validated on two standard\nbenchmark datasets, MSRC-21 and PascalVOC-2010. In Python based comparisons,\nour system is insignificantly slower than STF at test-time, yet produces\nsuperior semantic segmentations overall, with just push-button training.",
+        "authors": "Thanapong Intharah, Gabriel J. Brostow",
+        "published": "2015-02-17",
+        "updated": "2015-02-17",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1812.10885v1",
+        "title": "Coarse-to-fine Semantic Segmentation from Image-level Labels",
+        "abstract": "Deep neural network-based semantic segmentation generally requires\nlarge-scale cost extensive annotations for training to obtain better\nperformance. To avoid pixel-wise segmentation annotations which are needed for\nmost methods, recently some researchers attempted to use object-level labels\n(e.g. bounding boxes) or image-level labels (e.g. image categories). In this\npaper, we propose a novel recursive coarse-to-fine semantic segmentation\nframework based on only image-level category labels. For each image, an initial\ncoarse mask is first generated by a convolutional neural network-based\nunsupervised foreground segmentation model and then is enhanced by a graph\nmodel. The enhanced coarse mask is fed to a fully convolutional neural network\nto be recursively refined. Unlike existing image-level label-based semantic\nsegmentation methods which require to label all categories for images contain\nmultiple types of objects, our framework only needs one label for each image\nand can handle images contains multi-category objects. With only trained on\nImageNet, our framework achieves comparable performance on PASCAL VOC dataset\nas other image-level label-based state-of-the-arts of semantic segmentation.\nFurthermore, our framework can be easily extended to foreground object\nsegmentation task and achieves comparable performance with the state-of-the-art\nsupervised methods on the Internet Object dataset.",
+        "authors": "Longlong Jing, Yucheng Chen, Yingli Tian",
+        "published": "2018-12-28",
+        "updated": "2018-12-28",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2306.02095v1",
+        "title": "Content-aware Token Sharing for Efficient Semantic Segmentation with Vision Transformers",
+        "abstract": "This paper introduces Content-aware Token Sharing (CTS), a token reduction\napproach that improves the computational efficiency of semantic segmentation\nnetworks that use Vision Transformers (ViTs). Existing works have proposed\ntoken reduction approaches to improve the efficiency of ViT-based image\nclassification networks, but these methods are not directly applicable to\nsemantic segmentation, which we address in this work. We observe that, for\nsemantic segmentation, multiple image patches can share a token if they contain\nthe same semantic class, as they contain redundant information. Our approach\nleverages this by employing an efficient, class-agnostic policy network that\npredicts if image patches contain the same semantic class, and lets them share\na token if they do. With experiments, we explore the critical design choices of\nCTS and show its effectiveness on the ADE20K, Pascal Context and Cityscapes\ndatasets, various ViT backbones, and different segmentation decoders. With\nContent-aware Token Sharing, we are able to reduce the number of processed\ntokens by up to 44%, without diminishing the segmentation quality.",
+        "authors": "Chenyang Lu, Daan de Geus, Gijs Dubbelman",
+        "published": "2023-06-03",
+        "updated": "2023-06-03",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2308.06974v1",
+        "title": "A One Stop 3D Target Reconstruction and multilevel Segmentation Method",
+        "abstract": "3D object reconstruction and multilevel segmentation are fundamental to\ncomputer vision research. Existing algorithms usually perform 3D scene\nreconstruction and target objects segmentation independently, and the\nperformance is not fully guaranteed due to the challenge of the 3D\nsegmentation. Here we propose an open-source one stop 3D target reconstruction\nand multilevel segmentation framework (OSTRA), which performs segmentation on\n2D images, tracks multiple instances with segmentation labels in the image\nsequence, and then reconstructs labelled 3D objects or multiple parts with\nMulti-View Stereo (MVS) or RGBD-based 3D reconstruction methods. We extend\nobject tracking and 3D reconstruction algorithms to support continuous\nsegmentation labels to leverage the advances in the 2D image segmentation,\nespecially the Segment-Anything Model (SAM) which uses the pretrained neural\nnetwork without additional training for new scenes, for 3D object segmentation.\nOSTRA supports most popular 3D object models including point cloud, mesh and\nvoxel, and achieves high performance for semantic segmentation, instance\nsegmentation and part segmentation on several 3D datasets. It even surpasses\nthe manual segmentation in scenes with complex structures and occlusions. Our\nmethod opens up a new avenue for reconstructing 3D targets embedded with rich\nmulti-scale segmentation information in complex scenes. OSTRA is available from\nhttps://github.com/ganlab/OSTRA.",
+        "authors": "Jiexiong Xu, Weikun Zhao, Zhiyan Tang, Xiangchao Gan",
+        "published": "2023-08-14",
+        "updated": "2023-08-14",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2202.04754v2",
+        "title": "Wireless Transmission of Images With The Assistance of Multi-level Semantic Information",
+        "abstract": "Semantic-oriented communication has been considered as a promising to boost\nthe bandwidth efficiency by only transmitting the semantics of the data. In\nthis paper, we propose a multi-level semantic aware communication system for\nwireless image transmission, named MLSC-image, which is based on the deep\nlearning techniques and trained in an end to end manner. In particular, the\nproposed model includes a multilevel semantic feature extractor, that extracts\nboth the highlevel semantic information, such as the text semantics and the\nsegmentation semantics, and the low-level semantic information, such as local\nspatial details of the images. We employ a pretrained image caption to capture\nthe text semantics and a pretrained image segmentation model to obtain the\nsegmentation semantics. These high-level and low-level semantic features are\nthen combined and encoded by a joint semantic and channel encoder into symbols\nto transmit over the physical channel. The numerical results validate the\neffectiveness and efficiency of the proposed semantic communication system,\nespecially under the limited bandwidth condition, which indicates the\nadvantages of the high-level semantics in the compression of images.",
+        "authors": "Zhenguo Zhang, Qianqian Yang, Shibo He, Mingyang Sun, Jiming Chen",
+        "published": "2022-02-08",
+        "updated": "2023-12-08",
+        "primary_cat": "eess.IV",
+        "cats": [
+            "eess.IV",
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1611.08408v1",
+        "title": "Semantic Segmentation using Adversarial Networks",
+        "abstract": "Adversarial training has been shown to produce state of the art results for\ngenerative image modeling. In this paper we propose an adversarial training\napproach to train semantic segmentation models. We train a convolutional\nsemantic segmentation network along with an adversarial network that\ndiscriminates segmentation maps coming either from the ground truth or from the\nsegmentation network. The motivation for our approach is that it can detect and\ncorrect higher-order inconsistencies between ground truth segmentation maps and\nthe ones produced by the segmentation net. Our experiments show that our\nadversarial training approach leads to improved accuracy on the Stanford\nBackground and PASCAL VOC 2012 datasets.",
+        "authors": "Pauline Luc, Camille Couprie, Soumith Chintala, Jakob Verbeek",
+        "published": "2016-11-25",
+        "updated": "2016-11-25",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1904.03983v1",
+        "title": "Weakly Supervised Semantic Segmentation of Satellite Images",
+        "abstract": "When one wants to train a neural network to perform semantic segmentation,\ncreating pixel-level annotations for each of the images in the database is a\ntedious task. If he works with aerial or satellite images, which are usually\nvery large, it is even worse. With that in mind, we investigate how to use\nimage-level annotations in order to perform semantic segmentation. Image-level\nannotations are much less expensive to acquire than pixel-level annotations,\nbut we lose a lot of information for the training of the model. From the\nannotations of the images, the model must find by itself how to classify the\ndifferent regions of the image. In this work, we use the method proposed by Anh\nand Kwak [1] to produce pixel-level annotation from image level annotation. We\ncompare the overall quality of our generated dataset with the original dataset.\nIn addition, we propose an adaptation of the AffinityNet that allows us to\ndirectly perform a semantic segmentation. Our results show that the generated\nlabels lead to the same performances for the training of several segmentation\nnetworks. Also, the quality of semantic segmentation performed directly by the\nAffinityNet and the Random Walk is close to the one of the best\nfully-supervised approaches.",
+        "authors": "Adrien Nivaggioli, Hicham Randrianarivo",
+        "published": "2019-04-08",
+        "updated": "2019-04-08",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2309.01369v2",
+        "title": "Exploring Limits of Diffusion-Synthetic Training with Weakly Supervised Semantic Segmentation",
+        "abstract": "The advance of generative models for images has inspired various training\ntechniques for image recognition utilizing synthetic images. In semantic\nsegmentation, one promising approach is extracting pseudo-masks from attention\nmaps in text-to-image diffusion models, which enables\nreal-image-and-annotation-free training. However, the pioneering training\nmethod using the diffusion-synthetic images and pseudo-masks, i.e., DiffuMask\nhas limitations in terms of mask quality, scalability, and ranges of applicable\ndomains. To overcome these limitations, this work introduces three techniques\nfor diffusion-synthetic semantic segmentation training. First,\nreliability-aware robust training, originally used in weakly supervised\nlearning, helps segmentation with insufficient synthetic mask quality. %Second,\nlarge-scale pretraining of whole segmentation models, not only backbones, on\nsynthetic ImageNet-1k-class images with pixel-labels benefits downstream\nsegmentation tasks. Second, we introduce prompt augmentation, data augmentation\nto the prompt text set to scale up and diversify training images with a limited\ntext resources. Finally, LoRA-based adaptation of Stable Diffusion enables the\ntransfer to a distant domain, e.g., auto-driving images. Experiments in PASCAL\nVOC, ImageNet-S, and Cityscapes show that our method effectively closes gap\nbetween real and synthetic training in semantic segmentation.",
+        "authors": "Ryota Yoshihashi, Yuya Otsuka, Kenji Doi, Tomohiro Tanaka, Hirokatsu Kataoka",
+        "published": "2023-09-04",
+        "updated": "2024-04-15",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2201.05869v2",
+        "title": "Prototype Guided Network for Anomaly Segmentation",
+        "abstract": "Semantic segmentation methods can not directly identify abnormal objects in\nimages. Anomaly Segmentation algorithm from this realistic setting can\ndistinguish between in-distribution objects and Out-Of-Distribution (OOD)\nobjects and output the anomaly probability for pixels. In this paper, a\nPrototype Guided Anomaly segmentation Network (PGAN) is proposed to extract\nsemantic prototypes for in-distribution training data from limited annotated\nimages. In the model, prototypes are used to model the hierarchical category\nsemantic information and distinguish OOD pixels. The proposed PGAN model\nincludes a semantic segmentation network and a prototype extraction network.\nSimilarity measures are adopted to optimize the prototypes. The learned\nsemantic prototypes are used as category semantics to compare the similarity\nwith features extracted from test images and then to generate semantic\nsegmentation prediction. The proposed prototype extraction network can also be\nintegrated into most semantic segmentation networks and recognize OOD pixels.\nOn the StreetHazards dataset, the proposed PGAN model produced mIoU of 53.4%\nfor anomaly segmentation. The experimental results demonstrate PGAN may achieve\nthe SOTA performance in the anomaly segmentation tasks.",
+        "authors": "Yiqing Hao, Yi Jin, Gaoyun An",
+        "published": "2022-01-15",
+        "updated": "2022-03-15",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2210.05564v1",
+        "title": "Hypergraph Convolutional Networks for Weakly-Supervised Semantic Segmentation",
+        "abstract": "Semantic segmentation is a fundamental topic in computer vision. Several deep\nlearning methods have been proposed for semantic segmentation with outstanding\nresults. However, these models require a lot of densely annotated images. To\naddress this problem, we propose a new algorithm that uses HyperGraph\nConvolutional Networks for Weakly-supervised Semantic Segmentation\n(HyperGCN-WSS). Our algorithm constructs spatial and k-Nearest Neighbor (k-NN)\ngraphs from the images in the dataset to generate the hypergraphs. Then, we\ntrain a specialized HyperGraph Convolutional Network (HyperGCN) architecture\nusing some weak signals. The outputs of the HyperGCN are denominated\npseudo-labels, which are later used to train a DeepLab model for semantic\nsegmentation. HyperGCN-WSS is evaluated on the PASCAL VOC 2012 dataset for\nsemantic segmentation, using scribbles or clicks as weak signals. Our algorithm\nshows competitive performance against previous methods.",
+        "authors": "Jhony H. Giraldo, Vincenzo Scarrica, Antonino Staiano, Francesco Camastra, Thierry Bouwmans",
+        "published": "2022-10-11",
+        "updated": "2022-10-11",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1911.00679v3",
+        "title": "Cooperative Semantic Segmentation and Image Restoration in Adverse Environmental Conditions",
+        "abstract": "Most state-of-the-art semantic segmentation approaches only achieve high\naccuracy in good conditions. In practically-common but less-discussed adverse\nenvironmental conditions, their performance can decrease enormously. Existing\nstudies usually cast the handling of segmentation in adverse conditions as a\nseparate post-processing step after signal restoration, making the segmentation\nperformance largely depend on the quality of restoration. In this paper, we\npropose a novel deep-learning framework to tackle semantic segmentation and\nimage restoration in adverse environmental conditions in a holistic manner. The\nproposed approach contains two components: Semantically-Guided Adaptation,\nwhich exploits semantic information from degraded images to refine the\nsegmentation; and Exemplar-Guided Synthesis, which restores images from\nsemantic label maps given degraded exemplars as the guidance. Our method\ncooperatively leverages the complementarity and interdependence of low-level\nrestoration and high-level segmentation in adverse environmental conditions.\nExtensive experiments on various datasets demonstrate that our approach can not\nonly improve the accuracy of semantic segmentation with degradation cues, but\nalso boost the perceptual quality and structural similarity of image\nrestoration with semantic guidance.",
+        "authors": "Weihao Xia, Zhanglin Cheng, Yujiu Yang, Jing-Hao Xue",
+        "published": "2019-11-02",
+        "updated": "2020-03-03",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "eess.IV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2207.13297v5",
+        "title": "GPS-GLASS: Learning Nighttime Semantic Segmentation Using Daytime Video and GPS data",
+        "abstract": "Semantic segmentation for autonomous driving should be robust against various\nin-the-wild environments. Nighttime semantic segmentation is especially\nchallenging due to a lack of annotated nighttime images and a large domain gap\nfrom daytime images with sufficient annotation. In this paper, we propose a\nnovel GPS-based training framework for nighttime semantic segmentation. Given\nGPS-aligned pairs of daytime and nighttime images, we perform cross-domain\ncorrespondence matching to obtain pixel-level pseudo supervision. Moreover, we\nconduct flow estimation between daytime video frames and apply GPS-based\nscaling to acquire another pixel-level pseudo supervision. Using these pseudo\nsupervisions with a confidence map, we train a nighttime semantic segmentation\nnetwork without any annotation from nighttime images. Experimental results\ndemonstrate the effectiveness of the proposed method on several nighttime\nsemantic segmentation datasets. Our source code is available at\nhttps://github.com/jimmy9704/GPS-GLASS.",
+        "authors": "Hongjae Lee, Changwoo Han, Jun-Sang Yoo, Seung-Won Jung",
+        "published": "2022-07-27",
+        "updated": "2023-08-18",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2405.04913v1",
+        "title": "Weakly-supervised Semantic Segmentation via Dual-stream Contrastive Learning of Cross-image Contextual Information",
+        "abstract": "Weakly supervised semantic segmentation (WSSS) aims at learning a semantic\nsegmentation model with only image-level tags. Despite intensive research on\ndeep learning approaches over a decade, there is still a significant\nperformance gap between WSSS and full semantic segmentation. Most current WSSS\nmethods always focus on a limited single image (pixel-wise) information while\nignoring the valuable inter-image (semantic-wise) information. From this\nperspective, a novel end-to-end WSSS framework called DSCNet is developed along\nwith two innovations: i) pixel-wise group contrast and semantic-wise graph\ncontrast are proposed and introduced into the WSSS framework; ii) a novel\ndual-stream contrastive learning (DSCL) mechanism is designed to jointly handle\npixel-wise and semantic-wise context information for better WSSS performance.\nSpecifically, the pixel-wise group contrast learning (PGCL) and semantic-wise\ngraph contrast learning (SGCL) tasks form a more comprehensive solution.\nExtensive experiments on PASCAL VOC and MS COCO benchmarks verify the\nsuperiority of DSCNet over SOTA approaches and baseline models.",
+        "authors": "Qi Lai, Chi-Man Vong",
+        "published": "2024-05-08",
+        "updated": "2024-05-08",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2308.00949v3",
+        "title": "Synthetic Instance Segmentation from Semantic Image Segmentation Masks",
+        "abstract": "In recent years, the development of instance segmentation has garnered\nsignificant attention in a wide range of applications. However, the training of\na fully-supervised instance segmentation model requires costly both\ninstance-level and pixel-level annotations. In contrast, weakly-supervised\ninstance segmentation methods (i.e., with image-level class labels or point\nlabels) struggle to satisfy the accuracy and recall requirements of practical\nscenarios. In this paper, we propose a novel paradigm called synthetic instance\nsegmentation (SISeg), which achieves Instance Segmentation results from image\nmasks predicted using off-the-shelf semantic segmentation models. SISeg does\nnot require training a semantic or/and instance segmentation model and avoids\nthe need for instance-level image annotations. Therefore, it is highly\nefficient. Specifically, we first obtain a semantic segmentation mask of the\ninput image via a trained semantic segmentation model. Then, we calculate a\ndisplacement field vector for each pixel based on the segmentation mask, which\ncan indicate representations belonging to the same class but different\ninstances, i.e., obtaining the instance-level object information. Finally,\ninstance segmentation results are obtained after being refined by a learnable\ncategory-agnostic object boundary branch. Extensive experimental results on two\nchallenging datasets and representative semantic segmentation baselines\n(including CNNs and Transformers) demonstrate that SISeg can achieve\ncompetitive results compared to the state-of-the-art fully-supervised instance\nsegmentation methods without the need for additional human resources or\nincreased computational costs. The code is available at: SISeg",
+        "authors": "Yuchen Shen, Dong Zhang, Yuhui Zheng, Zechao Li, Liyong Fu, Qiaolin Ye",
+        "published": "2023-08-02",
+        "updated": "2023-10-31",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1606.01481v1",
+        "title": "Better Image Segmentation by Exploiting Dense Semantic Predictions",
+        "abstract": "It is well accepted that image segmentation can benefit from utilizing\nmultilevel cues. The paper focuses on utilizing the FCNN-based dense semantic\npredictions in the bottom-up image segmentation, arguing to take semantic cues\ninto account from the very beginning. By this we can avoid merging regions of\nsimilar appearance but distinct semantic categories as possible. The semantic\ninefficiency problem is handled. We also propose a straightforward way to use\nthe contour cues to suppress the noise in multilevel cues, thus to improve the\nsegmentation robustness. The evaluation on the BSDS500 shows that we obtain the\ncompetitive region and boundary performance. Furthermore, since all individual\nregions can be assigned with appropriate semantic labels during the\ncomputation, we are capable of extracting the adjusted semantic segmentations.\nThe experiment on Pascal VOC 2012 shows our improvement to the original\nsemantic segmentations which derives directly from the dense predictions.",
+        "authors": "Qiyang Zhao, Lewis D Griffin",
+        "published": "2016-06-05",
+        "updated": "2016-06-05",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2210.08988v1",
+        "title": "Heterogeneous Feature Distillation Network for SAR Image Semantic Segmentation",
+        "abstract": "Semantic segmentation for SAR (Synthetic Aperture Radar) images has attracted\nincreasing attention in the remote sensing community recently, due to SAR's\nall-time and all-weather imaging capability. However, SAR images are generally\nmore difficult to be segmented than their EO (Electro-Optical) counterparts,\nsince speckle noises and layovers are inevitably involved in SAR images. To\naddress this problem, we investigate how to introduce EO features to assist the\ntraining of a SAR-segmentation model, and propose a heterogeneous feature\ndistillation network for segmenting SAR images, called HFD-Net, where a\nSAR-segmentation student model gains knowledge from a pre-trained\nEO-segmentation teacher model. In the proposed HFD-Net, both the student and\nteacher models employ an identical architecture but different parameter\nconfigurations, and a heterogeneous feature distillation model is explored for\ntransferring latent EO features from the teacher model to the student model and\nthen enhancing the ability of the student model for SAR image segmentation. In\naddition, a heterogeneous feature alignment module is explored to aggregate\nmulti-scale features for segmentation in each of the student model and teacher\nmodel. Extensive experimental results on two public datasets demonstrate that\nthe proposed HFD-Net outperforms seven state-of-the-art SAR image semantic\nsegmentation methods.",
+        "authors": "Gao Mengyu, Dong Qiulei",
+        "published": "2022-10-17",
+        "updated": "2022-10-17",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "eess.IV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2007.02361v1",
+        "title": "Self-supervised Depth Estimation to Regularise Semantic Segmentation in Knee Arthroscopy",
+        "abstract": "Intra-operative automatic semantic segmentation of knee joint structures can\nassist surgeons during knee arthroscopy in terms of situational awareness.\nHowever, due to poor imaging conditions (e.g., low texture, overexposure,\netc.), automatic semantic segmentation is a challenging scenario, which\njustifies the scarce literature on this topic. In this paper, we propose a\nnovel self-supervised monocular depth estimation to regularise the training of\nthe semantic segmentation in knee arthroscopy. To further regularise the depth\nestimation, we propose the use of clean training images captured by the stereo\narthroscope of routine objects (presenting none of the poor imaging conditions\nand with rich texture information) to pre-train the model. We fine-tune such\nmodel to produce both the semantic segmentation and self-supervised monocular\ndepth using stereo arthroscopic images taken from inside the knee. Using a data\nset containing 3868 arthroscopic images captured during cadaveric knee\narthroscopy with semantic segmentation annotations, 2000 stereo image pairs of\ncadaveric knee arthroscopy, and 2150 stereo image pairs of routine objects, we\nshow that our semantic segmentation regularised by self-supervised depth\nestimation produces a more accurate segmentation than a state-of-the-art\nsemantic segmentation approach modeled exclusively with semantic segmentation\nannotation.",
+        "authors": "Fengbei Liu, Yaqub Jonmohamadi, Gabriel Maicas, Ajay K. Pandey, Gustavo Carneiro",
+        "published": "2020-07-05",
+        "updated": "2020-07-05",
+        "primary_cat": "eess.IV",
+        "cats": [
+            "eess.IV",
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2111.02444v2",
+        "title": "Panoptic 3D Scene Reconstruction From a Single RGB Image",
+        "abstract": "Understanding 3D scenes from a single image is fundamental to a wide variety\nof tasks, such as for robotics, motion planning, or augmented reality. Existing\nworks in 3D perception from a single RGB image tend to focus on geometric\nreconstruction only, or geometric reconstruction with semantic segmentation or\ninstance segmentation. Inspired by 2D panoptic segmentation, we propose to\nunify the tasks of geometric reconstruction, 3D semantic segmentation, and 3D\ninstance segmentation into the task of panoptic 3D scene reconstruction - from\na single RGB image, predicting the complete geometric reconstruction of the\nscene in the camera frustum of the image, along with semantic and instance\nsegmentations. We thus propose a new approach for holistic 3D scene\nunderstanding from a single RGB image which learns to lift and propagate 2D\nfeatures from an input image to a 3D volumetric scene representation. We\ndemonstrate that this holistic view of joint scene reconstruction, semantic,\nand instance segmentation is beneficial over treating the tasks independently,\nthus outperforming alternative approaches.",
+        "authors": "Manuel Dahnert, Ji Hou, Matthias Nie\u00dfner, Angela Dai",
+        "published": "2021-11-03",
+        "updated": "2022-05-16",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1609.09220v1",
+        "title": "CNN-aware Binary Map for General Semantic Segmentation",
+        "abstract": "In this paper we introduce a novel method for general semantic segmentation\nthat can benefit from general semantics of Convolutional Neural Network (CNN).\nOur segmentation proposes visually and semantically coherent image segments. We\nuse binary encoding of CNN features to overcome the difficulty of the\nclustering on the high-dimensional CNN feature space. These binary codes are\nvery robust against noise and non-semantic changes in the image. These binary\nencoding can be embedded into the CNN as an extra layer at the end of the\nnetwork. This results in real-time segmentation. To the best of our knowledge\nour method is the first attempt on general semantic image segmentation using\nCNN. All the previous papers were limited to few number of category of the\nimages (e.g. PASCAL VOC). Experiments show that our segmentation algorithm\noutperform the state-of-the-art non-semantic segmentation methods by large\nmargin.",
+        "authors": "Mahdyar Ravanbakhsh, Hossein Mousavi, Moin Nabi, Mohammad Rastegari, Carlo Regazzoni",
+        "published": "2016-09-29",
+        "updated": "2016-09-29",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1905.08748v3",
+        "title": "RIU-Net: Embarrassingly simple semantic segmentation of 3D LiDAR point cloud",
+        "abstract": "This paper proposes RIU-Net (for Range-Image U-Net), the adaptation of a\npopular semantic segmentation network for the semantic segmentation of a 3D\nLiDAR point cloud. The point cloud is turned into a 2D range-image by\nexploiting the topology of the sensor. This image is then used as input to a\nU-net. This architecture has already proved its efficiency for the task of\nsemantic segmentation of medical images. We demonstrate how it can also be used\nfor the accurate semantic segmentation of a 3D LiDAR point cloud and how it\nrepresents a valid bridge between image processing and 3D point cloud\nprocessing. Our model is trained on range-images built from KITTI 3D object\ndetection dataset. Experiments show that RIU-Net, despite being very simple,\noffers results that are comparable to the state-of-the-art of range-image based\nmethods. Finally, we demonstrate that this architecture is able to operate at\n90fps on a single GPU, which enables deployment for real-time segmentation.",
+        "authors": "Pierre Biasutti, Aur\u00e9lie Bugeau, Jean-Fran\u00e7ois Aujol, Mathieu Br\u00e9dif",
+        "published": "2019-05-21",
+        "updated": "2019-06-17",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1904.08465v2",
+        "title": "DeepAtlas: Joint Semi-Supervised Learning of Image Registration and Segmentation",
+        "abstract": "Deep convolutional neural networks (CNNs) are state-of-the-art for semantic\nimage segmentation, but typically require many labeled training samples.\nObtaining 3D segmentations of medical images for supervised training is\ndifficult and labor intensive. Motivated by classical approaches for joint\nsegmentation and registration we therefore propose a deep learning framework\nthat jointly learns networks for image registration and image segmentation. In\ncontrast to previous work on deep unsupervised image registration, which showed\nthe benefit of weak supervision via image segmentations, our approach can use\nexisting segmentations when available and computes them via the segmentation\nnetwork otherwise, thereby providing the same registration benefit. Conversely,\nsegmentation network training benefits from the registration, which essentially\nprovides a realistic form of data augmentation. Experiments on knee and brain\n3D magnetic resonance (MR) images show that our approach achieves large\nsimultaneous improvements of segmentation and registration accuracy (over\nindependently trained networks) and allows training high-quality models with\nvery limited training data. Specifically, in a one-shot-scenario (with only one\nmanually labeled image) our approach increases Dice scores (%) over an\nunsupervised registration network by 2.7 and 1.8 on the knee and brain images\nrespectively.",
+        "authors": "Zhenlin Xu, Marc Niethammer",
+        "published": "2019-04-17",
+        "updated": "2019-07-26",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2308.04829v2",
+        "title": "MixReorg: Cross-Modal Mixed Patch Reorganization is a Good Mask Learner for Open-World Semantic Segmentation",
+        "abstract": "Recently, semantic segmentation models trained with image-level text\nsupervision have shown promising results in challenging open-world scenarios.\nHowever, these models still face difficulties in learning fine-grained semantic\nalignment at the pixel level and predicting accurate object masks. To address\nthis issue, we propose MixReorg, a novel and straightforward pre-training\nparadigm for semantic segmentation that enhances a model's ability to\nreorganize patches mixed across images, exploring both local visual relevance\nand global semantic coherence. Our approach involves generating fine-grained\npatch-text pairs data by mixing image patches while preserving the\ncorrespondence between patches and text. The model is then trained to minimize\nthe segmentation loss of the mixed images and the two contrastive losses of the\noriginal and restored features. With MixReorg as a mask learner, conventional\ntext-supervised semantic segmentation models can achieve highly generalizable\npixel-semantic alignment ability, which is crucial for open-world segmentation.\nAfter training with large-scale image-text data, MixReorg models can be applied\ndirectly to segment visual objects of arbitrary categories, without the need\nfor further fine-tuning. Our proposed framework demonstrates strong performance\non popular zero-shot semantic segmentation benchmarks, outperforming GroupViT\nby significant margins of 5.0%, 6.2%, 2.5%, and 3.4% mIoU on PASCAL VOC2012,\nPASCAL Context, MS COCO, and ADE20K, respectively.",
+        "authors": "Kaixin Cai, Pengzhen Ren, Yi Zhu, Hang Xu, Jianzhuang Liu, Changlin Li, Guangrun Wang, Xiaodan Liang",
+        "published": "2023-08-09",
+        "updated": "2024-03-13",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2303.11316v2",
+        "title": "Generative Semantic Segmentation",
+        "abstract": "We present Generative Semantic Segmentation (GSS), a generative learning\napproach for semantic segmentation. Uniquely, we cast semantic segmentation as\nan image-conditioned mask generation problem. This is achieved by replacing the\nconventional per-pixel discriminative learning with a latent prior learning\nprocess. Specifically, we model the variational posterior distribution of\nlatent variables given the segmentation mask. To that end, the segmentation\nmask is expressed with a special type of image (dubbed as maskige). This\nposterior distribution allows to generate segmentation masks unconditionally.\nTo achieve semantic segmentation on a given image, we further introduce a\nconditioning network. It is optimized by minimizing the divergence between the\nposterior distribution of maskige (i.e., segmentation masks) and the latent\nprior distribution of input training images. Extensive experiments on standard\nbenchmarks show that our GSS can perform competitively to prior art\nalternatives in the standard semantic segmentation setting, whilst achieving a\nnew state of the art in the more challenging cross-domain setting.",
+        "authors": "Jiaqi Chen, Jiachen Lu, Xiatian Zhu, Li Zhang",
+        "published": "2023-03-20",
+        "updated": "2023-08-10",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1807.03138v1",
+        "title": "Quantity beats quality for semantic segmentation of corrosion in images",
+        "abstract": "Dataset creation is typically one of the first steps when applying Artificial\nIntelligence methods to a new task; and the real world performance of models\nhinges on the quality and quantity of data available. Producing an image\ndataset for semantic segmentation is resource intensive, particularly for\nspecialist subjects where class segmentation is not able to be effectively\nfarmed out. The benefit of producing a large, but poorly labelled, dataset\nversus a small, expertly segmented dataset for semantic segmentation is an open\nquestion. Here we show that a large, noisy dataset outperforms a small,\nexpertly segmented dataset for training a Fully Convolutional Network model for\nsemantic segmentation of corrosion in images. A large dataset of 250 images\nwith segmentations labelled by undergraduates and a second dataset of just 10\nimages, with segmentations labelled by subject matter experts were produced.\nThe mean Intersection over Union and micro F-score metrics were compared after\ntraining for 50,000 epochs. This work is illustrative for researchers setting\nout to develop deep learning models for detection and location of specialist\nfeatures.",
+        "authors": "Will Nash, Tom Drummond, Nick Birbilis",
+        "published": "2018-06-30",
+        "updated": "2018-06-30",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2001.00335v1",
+        "title": "Graph-FCN for image semantic segmentation",
+        "abstract": "Semantic segmentation with deep learning has achieved great progress in\nclassifying the pixels in the image. However, the local location information is\nusually ignored in the high-level feature extraction by the deep learning,\nwhich is important for image semantic segmentation. To avoid this problem, we\npropose a graph model initialized by a fully convolutional network (FCN) named\nGraph-FCN for image semantic segmentation. Firstly, the image grid data is\nextended to graph structure data by a convolutional network, which transforms\nthe semantic segmentation problem into a graph node classification problem.\nThen we apply graph convolutional network to solve this graph node\nclassification problem. As far as we know, it is the first time that we apply\nthe graph convolutional network in image semantic segmentation. Our method\nachieves competitive performance in mean intersection over union (mIOU) on the\nVOC dataset(about 1.34% improvement), compared to the original FCN model.",
+        "authors": "Yi Lu, Yaran Chen, Dongbin Zhao, Jianxin Chen",
+        "published": "2020-01-02",
+        "updated": "2020-01-02",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1911.12903v1",
+        "title": "Land Cover Change Detection via Semantic Segmentation",
+        "abstract": "This paper presents a change detection method that identifies land cover\nchanges from aerial imagery, using semantic segmentation, a machine learning\napproach. We present a land cover classification training pipeline with Deeplab\nv3+, state-of-the-art semantic segmentation technology, including data\npreparation, model training for seven land cover types, and model exporting\nmodules. In the land cover change detection system, the inputs are images\nretrieved from Google Earth at the same location but from different times. The\nsystem then predicts semantic segmentation results on these images using the\ntrained model and calculates the land cover class percentage for each input\nimage. We see an improvement in the accuracy of the land cover semantic\nsegmentation model, with a mean IoU of 0.756 compared to 0.433, as reported in\nthe DeepGlobe land cover classification challenge. The land cover change\ndetection system that leverages the state-of-the-art semantic segmentation\ntechnology is proposed and can be used for deforestation analysis, land\nmanagement, and urban planning.",
+        "authors": "Renee Su, Rong Chen",
+        "published": "2019-11-28",
+        "updated": "2019-11-28",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2107.13978v3",
+        "title": "Personalized Image Semantic Segmentation",
+        "abstract": "Semantic segmentation models trained on public datasets have achieved great\nsuccess in recent years. However, these models didn't consider the\npersonalization issue of segmentation though it is important in practice. In\nthis paper, we address the problem of personalized image segmentation. The\nobjective is to generate more accurate segmentation results on unlabeled\npersonalized images by investigating the data's personalized traits. To open up\nfuture research in this area, we collect a large dataset containing various\nusers' personalized images called PIS (Personalized Image Semantic\nSegmentation). We also survey some recent researches related to this problem\nand report their performance on our dataset. Furthermore, by observing the\ncorrelation among a user's personalized images, we propose a baseline method\nthat incorporates the inter-image context when segmenting certain images.\nExtensive experiments show that our method outperforms the existing methods on\nthe proposed dataset. The code and the PIS dataset will be made publicly\navailable.",
+        "authors": "Yu Zhang, Chang-Bin Zhang, Peng-Tao Jiang, Ming-Ming Cheng, Feng Mao",
+        "published": "2021-07-24",
+        "updated": "2021-09-04",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1804.04882v2",
+        "title": "Learning to Exploit the Prior Network Knowledge for Weakly-Supervised Semantic Segmentation",
+        "abstract": "Training a Convolutional Neural Network (CNN) for semantic segmentation\ntypically requires to collect a large amount of accurate pixel-level\nannotations, a hard and expensive task. In contrast, simple image tags are\neasier to gather. With this paper we introduce a novel weakly-supervised\nsemantic segmentation model able to learn from image labels, and just image\nlabels. Our model uses the prior knowledge of a network trained for image\nrecognition, employing these image annotations as an attention mechanism to\nidentify semantic regions in the images. We then present a methodology that\nbuilds accurate class-specific segmentation masks from these regions, where\nneither external objectness nor saliency algorithms are required. We describe\nhow to incorporate this mask generation strategy into a fully end-to-end\ntrainable process where the network jointly learns to classify and segment\nimages. Our experiments on PASCAL VOC 2012 dataset show that exploiting these\ngenerated class-specific masks in conjunction with our novel end-to-end\nlearning process outperforms several recent weakly-supervised semantic\nsegmentation methods that use image tags only, and even some models that\nleverage additional supervision or training data.",
+        "authors": "Carolina Redondo-Cabrera, Marcos Baptista-R\u00edos, Roberto J. L\u00f3pez-Sastre",
+        "published": "2018-04-13",
+        "updated": "2019-02-22",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1807.11857v1",
+        "title": "Joint Learning of Intrinsic Images and Semantic Segmentation",
+        "abstract": "Semantic segmentation of outdoor scenes is problematic when there are\nvariations in imaging conditions. It is known that albedo (reflectance) is\ninvariant to all kinds of illumination effects. Thus, using reflectance images\nfor semantic segmentation task can be favorable. Additionally, not only\nsegmentation may benefit from reflectance, but also segmentation may be useful\nfor reflectance computation. Therefore, in this paper, the tasks of semantic\nsegmentation and intrinsic image decomposition are considered as a combined\nprocess by exploring their mutual relationship in a joint fashion. To that end,\nwe propose a supervised end-to-end CNN architecture to jointly learn intrinsic\nimage decomposition and semantic segmentation. We analyze the gains of\naddressing those two problems jointly. Moreover, new cascade CNN architectures\nfor intrinsic-for-segmentation and segmentation-for-intrinsic are proposed as\nsingle tasks. Furthermore, a dataset of 35K synthetic images of natural\nenvironments is created with corresponding albedo and shading (intrinsics), as\nwell as semantic labels (segmentation) assigned to each object/scene. The\nexperiments show that joint learning of intrinsic image decomposition and\nsemantic segmentation is beneficial for both tasks for natural scenes. Dataset\nand models are available at: https://ivi.fnwi.uva.nl/cv/intrinseg",
+        "authors": "Anil S. Baslamisli, Thomas T. Groenestege, Partha Das, Hoang-An Le, Sezer Karaoglu, Theo Gevers",
+        "published": "2018-07-31",
+        "updated": "2018-07-31",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1509.02441v1",
+        "title": "Semantic Video Segmentation : Exploring Inference Efficiency",
+        "abstract": "We explore the efficiency of the CRF inference beyond image level semantic\nsegmentation and perform joint inference in video frames. The key idea is to\ncombine best of two worlds: semantic co-labeling and more expressive models.\nOur formulation enables us to perform inference over ten thousand images within\nseconds and makes the system amenable to perform video semantic segmentation\nmost effectively. On CamVid dataset, with TextonBoost unaries, our proposed\nmethod achieves up to 8% improvement in accuracy over individual semantic image\nsegmentation without additional time overhead. The source code is available at\nhttps://github.com/subtri/video_inference",
+        "authors": "Subarna Tripathi, Serge Belongie, Youngbae Hwang, Truong Nguyen",
+        "published": "2015-09-04",
+        "updated": "2015-09-04",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1709.01956v1",
+        "title": "Learning Dilation Factors for Semantic Segmentation of Street Scenes",
+        "abstract": "Contextual information is crucial for semantic segmentation. However, finding\nthe optimal trade-off between keeping desired fine details and at the same time\nproviding sufficiently large receptive fields is non trivial. This is even more\nso, when objects or classes present in an image significantly vary in size.\nDilated convolutions have proven valuable for semantic segmentation, because\nthey allow to increase the size of the receptive field without sacrificing\nimage resolution. However, in current state-of-the-art methods, dilation\nparameters are hand-tuned and fixed. In this paper, we present an approach for\nlearning dilation parameters adaptively per channel, consistently improving\nsemantic segmentation results on street-scene datasets like Cityscapes and\nCamvid.",
+        "authors": "Yang He, Margret Keuper, Bernt Schiele, Mario Fritz",
+        "published": "2017-09-06",
+        "updated": "2017-09-06",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2007.13138v1",
+        "title": "Virtual Multi-view Fusion for 3D Semantic Segmentation",
+        "abstract": "Semantic segmentation of 3D meshes is an important problem for 3D scene\nunderstanding. In this paper we revisit the classic multiview representation of\n3D meshes and study several techniques that make them effective for 3D semantic\nsegmentation of meshes. Given a 3D mesh reconstructed from RGBD sensors, our\nmethod effectively chooses different virtual views of the 3D mesh and renders\nmultiple 2D channels for training an effective 2D semantic segmentation model.\nFeatures from multiple per view predictions are finally fused on 3D mesh\nvertices to predict mesh semantic segmentation labels. Using the large scale\nindoor 3D semantic segmentation benchmark of ScanNet, we show that our virtual\nviews enable more effective training of 2D semantic segmentation networks than\nprevious multiview approaches. When the 2D per pixel predictions are aggregated\non 3D surfaces, our virtual multiview fusion method is able to achieve\nsignificantly better 3D semantic segmentation results compared to all prior\nmultiview approaches and competitive with recent 3D convolution approaches.",
+        "authors": "Abhijit Kundu, Xiaoqi Yin, Alireza Fathi, David Ross, Brian Brewington, Thomas Funkhouser, Caroline Pantofaru",
+        "published": "2020-07-26",
+        "updated": "2020-07-26",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "eess.IV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2101.09642v2",
+        "title": "Image Compression with Encoder-Decoder Matched Semantic Segmentation",
+        "abstract": "In recent years, layered image compression is demonstrated to be a promising\ndirection, which encodes a compact representation of the input image and apply\nan up-sampling network to reconstruct the image. To further improve the quality\nof the reconstructed image, some works transmit the semantic segment together\nwith the compressed image data. Consequently, the compression ratio is also\ndecreased because extra bits are required for transmitting the semantic\nsegment. To solve this problem, we propose a new layered image compression\nframework with encoder-decoder matched semantic segmentation (EDMS). And then,\nfollowed by the semantic segmentation, a special convolution neural network is\nused to enhance the inaccurate semantic segment. As a result, the accurate\nsemantic segment can be obtained in the decoder without requiring extra bits.\nThe experimental results show that the proposed EDMS framework can get up to\n35.31% BD-rate reduction over the HEVC-based (BPG) codec, 5% bitrate, and 24%\nencoding time saving compare to the state-of-the-art semantic-based image\ncodec.",
+        "authors": "Trinh Man Hoang, Jinjia Zhou, Yibo Fan",
+        "published": "2021-01-24",
+        "updated": "2021-01-30",
+        "primary_cat": "eess.IV",
+        "cats": [
+            "eess.IV",
+            "cs.CV",
+            "cs.MM"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1707.02432v2",
+        "title": "Deep Semantic Segmentation for Automated Driving: Taxonomy, Roadmap and Challenges",
+        "abstract": "Semantic segmentation was seen as a challenging computer vision problem few\nyears ago. Due to recent advancements in deep learning, relatively accurate\nsolutions are now possible for its use in automated driving. In this paper, the\nsemantic segmentation problem is explored from the perspective of automated\ndriving. Most of the current semantic segmentation algorithms are designed for\ngeneric images and do not incorporate prior structure and end goal for\nautomated driving. First, the paper begins with a generic taxonomic survey of\nsemantic segmentation algorithms and then discusses how it fits in the context\nof automated driving. Second, the particular challenges of deploying it into a\nsafety system which needs high level of accuracy and robustness are listed.\nThird, different alternatives instead of using an independent semantic\nsegmentation module are explored. Finally, an empirical evaluation of various\nsemantic segmentation architectures was performed on CamVid dataset in terms of\naccuracy and speed. This paper is a preliminary shorter version of a more\ndetailed survey which is work in progress.",
+        "authors": "Mennatullah Siam, Sara Elkerdawy, Martin Jagersand, Senthil Yogamani",
+        "published": "2017-07-08",
+        "updated": "2017-08-03",
+        "primary_cat": "stat.ML",
+        "cats": [
+            "stat.ML",
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2101.08418v2",
+        "title": "Rethinking Semantic Segmentation Evaluation for Explainability and Model Selection",
+        "abstract": "Semantic segmentation aims to robustly predict coherent class labels for\nentire regions of an image. It is a scene understanding task that powers\nreal-world applications (e.g., autonomous navigation). One important\napplication, the use of imagery for automated semantic understanding of\npedestrian environments, provides remote mapping of accessibility features in\nstreet environments. This application (and others like it) require detailed\ngeometric information of geographical objects. Semantic segmentation is a\nprerequisite for this task since it maps contiguous regions of the same class\nas single entities. Importantly, semantic segmentation uses like ours are not\npixel-wise outcomes; however, most of their quantitative evaluation metrics\n(e.g., mean Intersection Over Union) are based on pixel-wise similarities to a\nground-truth, which fails to emphasize over- and under-segmentation properties\nof a segmentation model. Here, we introduce a new metric to assess region-based\nover- and under-segmentation. We analyze and compare it to other metrics,\ndemonstrating that the use of our metric lends greater explainability to\nsemantic segmentation model performance in real-world applications.",
+        "authors": "Yuxiang Zhang, Sachin Mehta, Anat Caspi",
+        "published": "2021-01-21",
+        "updated": "2023-02-15",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2307.13215v1",
+        "title": "Image Segmentation Keras : Implementation of Segnet, FCN, UNet, PSPNet and other models in Keras",
+        "abstract": "Semantic segmentation plays a vital role in computer vision tasks, enabling\nprecise pixel-level understanding of images. In this paper, we present a\ncomprehensive library for semantic segmentation, which contains implementations\nof popular segmentation models like SegNet, FCN, UNet, and PSPNet. We also\nevaluate and compare these models on several datasets, offering researchers and\npractitioners a powerful toolset for tackling diverse segmentation challenges.",
+        "authors": "Divam Gupta",
+        "published": "2023-07-25",
+        "updated": "2023-07-25",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2306.06753v1",
+        "title": "3rd Place Solution for PVUW Challenge 2023: Video Panoptic Segmentation",
+        "abstract": "In order to deal with the task of video panoptic segmentation in the wild, we\npropose a robust integrated video panoptic segmentation solution. In our\nsolution, we regard the video panoptic segmentation task as a segmentation\ntarget querying task, represent both semantic and instance targets as a set of\nqueries, and then combine these queries with video features extracted by neural\nnetworks to predict segmentation masks. In order to improve the learning\naccuracy and convergence speed of the solution, we add additional tasks of\nvideo semantic segmentation and video instance segmentation for joint training.\nIn addition, we also add an additional image semantic segmentation model to\nfurther improve the performance of semantic classes. In addition, we also add\nsome additional operations to improve the robustness of the model. Extensive\nexperiments on the VIPSeg dataset show that the proposed solution achieves\nstate-of-the-art performance with 50.04\\% VPQ on the VIPSeg test set, which is\n3rd place on the video panoptic segmentation track of the PVUW Challenge 2023.",
+        "authors": "Jinming Su, Wangwang Yang, Junfeng Luo, Xiaolin Wei",
+        "published": "2023-06-11",
+        "updated": "2023-06-11",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2308.11506v1",
+        "title": "LCCo: Lending CLIP to Co-Segmentation",
+        "abstract": "This paper studies co-segmenting the common semantic object in a set of\nimages. Existing works either rely on carefully engineered networks to mine the\nimplicit semantic information in visual features or require extra data (i.e.,\nclassification labels) for training. In this paper, we leverage the contrastive\nlanguage-image pre-training framework (CLIP) for the task. With a backbone\nsegmentation network that independently processes each image from the set, we\nintroduce semantics from CLIP into the backbone features, refining them in a\ncoarse-to-fine manner with three key modules: i) an image set feature\ncorrespondence module, encoding global consistent semantic information of the\nimage set; ii) a CLIP interaction module, using CLIP-mined common semantics of\nthe image set to refine the backbone feature; iii) a CLIP regularization\nmodule, drawing CLIP towards this co-segmentation task, identifying the best\nCLIP semantic and using it to regularize the backbone feature. Experiments on\nfour standard co-segmentation benchmark datasets show that the performance of\nour method outperforms state-of-the-art methods.",
+        "authors": "Xin Duan, Yan Yang, Liyuan Pan, Xiabi Liu",
+        "published": "2023-08-22",
+        "updated": "2023-08-22",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2304.02713v1",
+        "title": "NUMSnet: Nested-U Multi-class Segmentation network for 3D Medical Image Stacks",
+        "abstract": "Semantic segmentation for medical 3D image stacks enables accurate volumetric\nreconstructions, computer-aided diagnostics and follow up treatment planning.\nIn this work, we present a novel variant of the Unet model called the NUMSnet\nthat transmits pixel neighborhood features across scans through nested layers\nto achieve accurate multi-class semantic segmentations with minimal training\ndata. We analyze the semantic segmentation performance of the NUMSnet model in\ncomparison with several Unet model variants to segment 3-7 regions of interest\nusing only 10% of images for training per Lung-CT and Heart-CT volumetric image\nstacks. The proposed NUMSnet model achieves up to 20% improvement in\nsegmentation recall with 4-9% improvement in Dice scores for Lung-CT stacks and\n2.5-10% improvement in Dice scores for Heart-CT stacks when compared to the\nUnet++ model. The NUMSnet model needs to be trained by ordered images around\nthe central scan of each volumetric stack. Propagation of image feature\ninformation from the 6 nested layers of the Unet++ model are found to have\nbetter computation and segmentation performances than propagation of all\nup-sampling layers in a Unet++ model. The NUMSnet model achieves comparable\nsegmentation performances to existing works, while being trained on as low as\n5\\% of the training images. Also, transfer learning allows faster convergence\nof the NUMSnet model for multi-class semantic segmentation from pathology in\nLung-CT images to cardiac segmentations in Heart-CT stacks. Thus, the proposed\nmodel can standardize multi-class semantic segmentation on a variety of\nvolumetric image stacks with minimal training dataset. This can significantly\nreduce the cost, time and inter-observer variabilities associated with\ncomputer-aided detections and treatment.",
+        "authors": "Sohini Roychowdhury",
+        "published": "2023-04-05",
+        "updated": "2023-04-05",
+        "primary_cat": "eess.IV",
+        "cats": [
+            "eess.IV",
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2006.07601v1",
+        "title": "NoPeopleAllowed: The Three-Step Approach to Weakly Supervised Semantic Segmentation",
+        "abstract": "We propose a novel approach to weakly supervised semantic segmentation, which\nconsists of three consecutive steps. The first two steps extract high-quality\npseudo masks from image-level annotated data, which are then used to train a\nsegmentation model on the third step. The presented approach also addresses two\nproblems in the data: class imbalance and missing labels. Using only\nimage-level annotations as supervision, our method is capable of segmenting\nvarious classes and complex objects. It achieves 37.34 mean IoU on the test\nset, placing 3rd at the LID Challenge in the task of weakly supervised semantic\nsegmentation.",
+        "authors": "Mariia Dobko, Ostap Viniavskyi, Oles Dobosevych",
+        "published": "2020-06-13",
+        "updated": "2020-06-13",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2310.13026v1",
+        "title": "Weakly-Supervised Semantic Segmentation with Image-Level Labels: from Traditional Models to Foundation Models",
+        "abstract": "The rapid development of deep learning has driven significant progress in the\nfield of image semantic segmentation - a fundamental task in computer vision.\nSemantic segmentation algorithms often depend on the availability of\npixel-level labels (i.e., masks of objects), which are expensive,\ntime-consuming, and labor-intensive. Weakly-supervised semantic segmentation\n(WSSS) is an effective solution to avoid such labeling. It utilizes only\npartial or incomplete annotations and provides a cost-effective alternative to\nfully-supervised semantic segmentation. In this paper, we focus on the WSSS\nwith image-level labels, which is the most challenging form of WSSS. Our work\nhas two parts. First, we conduct a comprehensive survey on traditional methods,\nprimarily focusing on those presented at premier research conferences. We\ncategorize them into four groups based on where their methods operate:\npixel-wise, image-wise, cross-image, and external data. Second, we investigate\nthe applicability of visual foundation models, such as the Segment Anything\nModel (SAM), in the context of WSSS. We scrutinize SAM in two intriguing\nscenarios: text prompting and zero-shot learning. We provide insights into the\npotential and challenges associated with deploying visual foundational models\nfor WSSS, facilitating future developments in this exciting research area.",
+        "authors": "Zhaozheng Chen, Qianru Sun",
+        "published": "2023-10-19",
+        "updated": "2023-10-19",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1610.01706v1",
+        "title": "Exploiting Depth from Single Monocular Images for Object Detection and Semantic Segmentation",
+        "abstract": "Augmenting RGB data with measured depth has been shown to improve the\nperformance of a range of tasks in computer vision including object detection\nand semantic segmentation. Although depth sensors such as the Microsoft Kinect\nhave facilitated easy acquisition of such depth information, the vast majority\nof images used in vision tasks do not contain depth information. In this paper,\nwe show that augmenting RGB images with estimated depth can also improve the\naccuracy of both object detection and semantic segmentation. Specifically, we\nfirst exploit the recent success of depth estimation from monocular images and\nlearn a deep depth estimation model. Then we learn deep depth features from the\nestimated depth and combine with RGB features for object detection and semantic\nsegmentation. Additionally, we propose an RGB-D semantic segmentation method\nwhich applies a multi-task training scheme: semantic label prediction and depth\nvalue regression. We test our methods on several datasets and demonstrate that\nincorporating information from estimated depth improves the performance of\nobject detection and semantic segmentation remarkably.",
+        "authors": "Yuanzhouhan Cao, Chunhua Shen, Heng Tao Shen",
+        "published": "2016-10-06",
+        "updated": "2016-10-06",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2306.02094v1",
+        "title": "Segment Anything Meets Semantic Communication",
+        "abstract": "In light of the diminishing returns of traditional methods for enhancing\ntransmission rates, the domain of semantic communication presents promising new\nfrontiers. Focusing on image transmission, this paper explores the application\nof foundation models, particularly the Segment Anything Model (SAM) developed\nby Meta AI Research, to improve semantic communication. SAM is a promptable\nimage segmentation model that has gained attention for its ability to perform\nzero-shot segmentation tasks without explicit training or domain-specific\nknowledge. By employing SAM's segmentation capability and lightweight neural\nnetwork architecture for semantic coding, we propose a practical approach to\nsemantic communication. We demonstrate that this approach retains critical\nsemantic features, achieving higher image reconstruction quality and reducing\ncommunication overhead. This practical solution eliminates the\nresource-intensive stage of training a segmentation model and can be applied to\nany semantic coding architecture, paving the way for real-world applications.",
+        "authors": "Shehbaz Tariq, Brian Estadimas Arfeto, Chaoning Zhang, Hyundong Shin",
+        "published": "2023-06-03",
+        "updated": "2023-06-03",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1809.10198v1",
+        "title": "Recent progress in semantic image segmentation",
+        "abstract": "Semantic image segmentation, which becomes one of the key applications in\nimage processing and computer vision domain, has been used in multiple domains\nsuch as medical area and intelligent transportation. Lots of benchmark datasets\nare released for researchers to verify their algorithms. Semantic segmentation\nhas been studied for many years. Since the emergence of Deep Neural Network\n(DNN), segmentation has made a tremendous progress. In this paper, we divide\nsemantic image segmentation methods into two categories: traditional and recent\nDNN method. Firstly, we briefly summarize the traditional method as well as\ndatasets released for segmentation, then we comprehensively investigate recent\nmethods based on DNN which are described in the eight aspects: fully\nconvolutional network, upsample ways, FCN joint with CRF methods, dilated\nconvolution approaches, progresses in backbone network, pyramid methods,\nMulti-level feature and multi-stage method, supervised, weakly-supervised and\nunsupervised methods. Finally, a conclusion in this area is drawn.",
+        "authors": "Xiaolong Liu, Zhidong Deng, Yuhan Yang",
+        "published": "2018-09-20",
+        "updated": "2018-09-20",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2003.04404v1",
+        "title": "FusionLane: Multi-Sensor Fusion for Lane Marking Semantic Segmentation Using Deep Neural Networks",
+        "abstract": "It is a crucial step to achieve effective semantic segmentation of lane\nmarking during the construction of the lane level high-precision map. In recent\nyears, many image semantic segmentation methods have been proposed. These\nmethods mainly focus on the image from camera, due to the limitation of the\nsensor itself, the accurate three-dimensional spatial position of the lane\nmarking cannot be obtained, so the demand for the lane level high-precision map\nconstruction cannot be met. This paper proposes a lane marking semantic\nsegmentation method based on LIDAR and camera fusion deep neural network.\nDifferent from other methods, in order to obtain accurate position information\nof the segmentation results, the semantic segmentation object of this paper is\na bird's eye view converted from a LIDAR points cloud instead of an image\ncaptured by a camera. This method first uses the deeplabv3+ [\\ref{ref:1}]\nnetwork to segment the image captured by the camera, and the segmentation\nresult is merged with the point clouds collected by the LIDAR as the input of\nthe proposed network. In this neural network, we also add a long short-term\nmemory (LSTM) structure to assist the network for semantic segmentation of lane\nmarkings by using the the time series information. The experiments on more than\n14,000 image datasets which we have manually labeled and expanded have shown\nthe proposed method has better performance on the semantic segmentation of the\npoints cloud bird's eye view. Therefore, the automation of high-precision map\nconstruction can be significantly improved. Our code is available at\nhttps://github.com/rolandying/FusionLane.",
+        "authors": "Ruochen Yin, Biao Yu, Huapeng Wu, Yutao Song, Runxin Niu",
+        "published": "2020-03-09",
+        "updated": "2020-03-09",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2210.12417v2",
+        "title": "SLAMs: Semantic Learning based Activation Map for Weakly Supervised Semantic Segmentation",
+        "abstract": "Recent mainstream weakly-supervised semantic segmentation (WSSS) approaches\nmainly relies on image-level classification learning, which has limited\nrepresentation capacity. In this paper, we propose a novel semantic learning\nbased framework, named SLAMs (Semantic Learning based Activation Map), for\nWSSS.",
+        "authors": "Junliang Chen, Xiaodong Zhao, Minmin Liu, Linlin Shen",
+        "published": "2022-10-22",
+        "updated": "2022-11-10",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2104.13395v3",
+        "title": "ACDC: The Adverse Conditions Dataset with Correspondences for Semantic Driving Scene Understanding",
+        "abstract": "Level 5 autonomy for self-driving cars requires a robust visual perception\nsystem that can parse input images under any visual condition. However,\nexisting semantic segmentation datasets are either dominated by images captured\nunder normal conditions or are small in scale. To address this, we introduce\nACDC, the Adverse Conditions Dataset with Correspondences for training and\ntesting semantic segmentation methods on adverse visual conditions. ACDC\nconsists of a large set of 4006 images which are equally distributed between\nfour common adverse conditions: fog, nighttime, rain, and snow. Each\nadverse-condition image comes with a high-quality fine pixel-level semantic\nannotation, a corresponding image of the same scene taken under normal\nconditions, and a binary mask that distinguishes between intra-image regions of\nclear and uncertain semantic content. Thus, ACDC supports both standard\nsemantic segmentation and the newly introduced uncertainty-aware semantic\nsegmentation. A detailed empirical study demonstrates the challenges that the\nadverse domains of ACDC pose to state-of-the-art supervised and unsupervised\napproaches and indicates the value of our dataset in steering future progress\nin the field. Our dataset and benchmark are publicly available.",
+        "authors": "Christos Sakaridis, Dengxin Dai, Luc Van Gool",
+        "published": "2021-04-27",
+        "updated": "2021-09-01",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2104.00487v1",
+        "title": "Linear Semantics in Generative Adversarial Networks",
+        "abstract": "Generative Adversarial Networks (GANs) are able to generate high-quality\nimages, but it remains difficult to explicitly specify the semantics of\nsynthesized images. In this work, we aim to better understand the semantic\nrepresentation of GANs, and thereby enable semantic control in GAN's generation\nprocess. Interestingly, we find that a well-trained GAN encodes image semantics\nin its internal feature maps in a surprisingly simple way: a linear\ntransformation of feature maps suffices to extract the generated image\nsemantics. To verify this simplicity, we conduct extensive experiments on\nvarious GANs and datasets; and thanks to this simplicity, we are able to learn\na semantic segmentation model for a trained GAN from a small number (e.g., 8)\nof labeled images. Last but not least, leveraging our findings, we propose two\nfew-shot image editing approaches, namely Semantic-Conditional Sampling and\nSemantic Image Editing. Given a trained GAN and as few as eight semantic\nannotations, the user is able to generate diverse images subject to a\nuser-provided semantic layout, and control the synthesized image semantics. We\nhave made the code publicly available.",
+        "authors": "Jianjin Xu, Changxi Zheng",
+        "published": "2021-04-01",
+        "updated": "2021-04-01",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2106.04108v3",
+        "title": "Fully Transformer Networks for Semantic Image Segmentation",
+        "abstract": "Transformers have shown impressive performance in various natural language\nprocessing and computer vision tasks, due to the capability of modeling\nlong-range dependencies. Recent progress has demonstrated that combining such\nTransformers with CNN-based semantic image segmentation models is very\npromising. However, it is not well studied yet on how well a pure Transformer\nbased approach can achieve for image segmentation. In this work, we explore a\nnovel framework for semantic image segmentation, which is encoder-decoder based\nFully Transformer Networks (FTN). Specifically, we first propose a Pyramid\nGroup Transformer (PGT) as the encoder for progressively learning hierarchical\nfeatures, meanwhile reducing the computation complexity of the standard Visual\nTransformer (ViT). Then, we propose a Feature Pyramid Transformer (FPT) to fuse\nsemantic-level and spatial-level information from multiple levels of the PGT\nencoder for semantic image segmentation. Surprisingly, this simple baseline can\nachieve better results on multiple challenging semantic segmentation and face\nparsing benchmarks, including PASCAL Context, ADE20K, COCOStuff, and\nCelebAMask-HQ. The source code will be released on\nhttps://github.com/BR-IDL/PaddleViT.",
+        "authors": "Sitong Wu, Tianyi Wu, Fangjian Lin, Shengwei Tian, Guodong Guo",
+        "published": "2021-06-08",
+        "updated": "2021-12-28",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2305.17091v1",
+        "title": "SSSegmenation: An Open Source Supervised Semantic Segmentation Toolbox Based on PyTorch",
+        "abstract": "This paper presents SSSegmenation, which is an open source supervised\nsemantic image segmentation toolbox based on PyTorch. The design of this\ntoolbox is motivated by MMSegmentation while it is easier to use because of\nfewer dependencies and achieves superior segmentation performance under a\ncomparable training and testing setup. Moreover, the toolbox also provides\nplenty of trained weights for popular and contemporary semantic segmentation\nmethods, including Deeplab, PSPNet, OCRNet, MaskFormer, \\emph{etc}. We expect\nthat this toolbox can contribute to the future development of semantic\nsegmentation. Codes and model zoos are available at\n\\href{https://github.com/SegmentationBLWX/sssegmentation/}{SSSegmenation}.",
+        "authors": "Zhenchao Jin",
+        "published": "2023-05-26",
+        "updated": "2023-05-26",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2305.03273v1",
+        "title": "Semantic Segmentation using Vision Transformers: A survey",
+        "abstract": "Semantic segmentation has a broad range of applications in a variety of\ndomains including land coverage analysis, autonomous driving, and medical image\nanalysis. Convolutional neural networks (CNN) and Vision Transformers (ViTs)\nprovide the architecture models for semantic segmentation. Even though ViTs\nhave proven success in image classification, they cannot be directly applied to\ndense prediction tasks such as image segmentation and object detection since\nViT is not a general purpose backbone due to its patch partitioning scheme. In\nthis survey, we discuss some of the different ViT architectures that can be\nused for semantic segmentation and how their evolution managed the above-stated\nchallenge. The rise of ViT and its performance with a high success rate\nmotivated the community to slowly replace the traditional convolutional neural\nnetworks in various computer vision tasks. This survey aims to review and\ncompare the performances of ViT architectures designed for semantic\nsegmentation using benchmarking datasets. This will be worthwhile for the\ncommunity to yield knowledge regarding the implementations carried out in\nsemantic segmentation and to discover more efficient methodologies using ViTs.",
+        "authors": "Hans Thisanke, Chamli Deshan, Kavindu Chamith, Sachith Seneviratne, Rajith Vidanaarachchi, Damayanthi Herath",
+        "published": "2023-05-05",
+        "updated": "2023-05-05",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI",
+            "cs.LG"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2107.02525v1",
+        "title": "Semantic Segmentation Alternative Technique: Segmentation Domain Generation",
+        "abstract": "Detecting objects of interest in images was always a compelling task to\nautomate. In recent years this task was more and more explored using deep\nlearning techniques, mostly using region-based convolutional networks. In this\nproject we propose an alternative semantic segmentation technique making use of\nGenerative Adversarial Networks. We consider semantic segmentation to be a\ndomain transfer problem. Thus, we train a feed forward network (FFNN) to\nreceive as input a seed real image and generate as output its segmentation\nmask.",
+        "authors": "Ana-Cristina Rogoz, Radu Muntean, Stefan Cobeli",
+        "published": "2021-07-06",
+        "updated": "2021-07-06",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.LG",
+            "eess.IV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2012.10782v2",
+        "title": "Three Ways to Improve Semantic Segmentation with Self-Supervised Depth Estimation",
+        "abstract": "Training deep networks for semantic segmentation requires large amounts of\nlabeled training data, which presents a major challenge in practice, as\nlabeling segmentation masks is a highly labor-intensive process. To address\nthis issue, we present a framework for semi-supervised semantic segmentation,\nwhich is enhanced by self-supervised monocular depth estimation from unlabeled\nimage sequences. In particular, we propose three key contributions: (1) We\ntransfer knowledge from features learned during self-supervised depth\nestimation to semantic segmentation, (2) we implement a strong data\naugmentation by blending images and labels using the geometry of the scene, and\n(3) we utilize the depth feature diversity as well as the level of difficulty\nof learning depth in a student-teacher framework to select the most useful\nsamples to be annotated for semantic segmentation. We validate the proposed\nmodel on the Cityscapes dataset, where all three modules demonstrate\nsignificant performance gains, and we achieve state-of-the-art results for\nsemi-supervised semantic segmentation. The implementation is available at\nhttps://github.com/lhoyer/improving_segmentation_with_selfsupervised_depth.",
+        "authors": "Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian K\u00f6ring, Suman Saha, Luc Van Gool",
+        "published": "2020-12-19",
+        "updated": "2021-04-05",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2110.04487v1",
+        "title": "Colour augmentation for improved semi-supervised semantic segmentation",
+        "abstract": "Consistency regularization describes a class of approaches that have yielded\nstate-of-the-art results for semi-supervised classification. While\nsemi-supervised semantic segmentation proved to be more challenging, a number\nof successful approaches have been recently proposed. Recent work explored the\nchallenges involved in using consistency regularization for segmentation\nproblems. In their self-supervised work Chen et al. found that colour\naugmentation prevents a classification network from using image colour\nstatistics as a short-cut for self-supervised learning via instance\ndiscrimination. Drawing inspiration from this we find that a similar problem\nimpedes semi-supervised semantic segmentation and offer colour augmentation as\na solution, improving semi-supervised semantic segmentation performance on\nchallenging photographic imagery.",
+        "authors": "Geoff French, Michal Mackiewicz",
+        "published": "2021-10-09",
+        "updated": "2021-10-09",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.LG"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1809.10245v1",
+        "title": "Cylindrical Transform: 3D Semantic Segmentation of Kidneys With Limited Annotated Images",
+        "abstract": "In this paper, we propose a novel technique for sampling sequential images\nusing a cylindrical transform in a cylindrical coordinate system for kidney\nsemantic segmentation in abdominal computed tomography (CT). The images\ngenerated from a cylindrical transform augment a limited annotated set of\nimages in three dimensions. This approach enables us to train contemporary\nclassification deep convolutional neural networks (DCNNs) instead of fully\nconvolutional networks (FCNs) for semantic segmentation. Typical semantic\nsegmentation models segment a sequential set of images (e.g. CT or video) by\nsegmenting each image independently. However, the proposed method not only\nconsiders the spatial dependency in the x-y plane, but also the spatial\nsequential dependency along the z-axis. The results show that classification\nDCNNs, trained on cylindrical transformed images, can achieve a higher\nsegmentation performance value than FCNs using a limited number of annotated\nimages.",
+        "authors": "Hojjat Salehinejad, Sumeya Naqvi, Errol Colak, Joseph Barfett, Shahrokh Valaee",
+        "published": "2018-09-24",
+        "updated": "2018-09-24",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.NE"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2401.07589v1",
+        "title": "Semantic Scene Segmentation for Robotics",
+        "abstract": "Comprehensive scene understanding is a critical enabler of robot autonomy.\nSemantic segmentation is one of the key scene understanding tasks which is\npivotal for several robotics applications including autonomous driving,\ndomestic service robotics, last mile delivery, amongst many others. Semantic\nsegmentation is a dense prediction task that aims to provide a scene\nrepresentation in which each pixel of an image is assigned a semantic class\nlabel. Therefore, semantic segmentation considers the full scene context,\nincorporating the object category, location, and shape of all the scene\nelements, including the background. Numerous algorithms have been proposed for\nsemantic segmentation over the years. However, the recent advances in deep\nlearning combined with the boost in the computational capacity and the\navailability of large-scale labeled datasets have led to significant advances\nin semantic segmentation. In this chapter, we introduce the task of semantic\nsegmentation and present the deep learning techniques that have been proposed\nto address this task over the years. We first define the task of semantic\nsegmentation and contrast it with other closely related scene understanding\nproblems. We detail different algorithms and architectures for semantic\nsegmentation and the commonly employed loss functions. Furthermore, we present\nan overview of datasets, benchmarks, and metrics that are used in semantic\nsegmentation. We conclude the chapter with a discussion of challenges and\nopportunities for further research in this area.",
+        "authors": "Juana Valeria Hurtado, Abhinav Valada",
+        "published": "2024-01-15",
+        "updated": "2024-01-15",
+        "primary_cat": "cs.RO",
+        "cats": [
+            "cs.RO"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2108.02840v1",
+        "title": "Attention-based fusion of semantic boundary and non-boundary information to improve semantic segmentation",
+        "abstract": "This paper introduces a method for image semantic segmentation grounded on a\nnovel fusion scheme, which takes place inside a deep convolutional neural\nnetwork. The main goal of our proposal is to explore object boundary\ninformation to improve the overall segmentation performance. Unlike previous\nworks that combine boundary and segmentation features, or those that use\nboundary information to regularize semantic segmentation, we instead propose a\nnovel approach that embodies boundary information onto segmentation. For that,\nour semantic segmentation method uses two streams, which are combined through\nan attention gate, forming an end-to-end Y-model. To the best of our knowledge,\nours is the first work to show that boundary detection can improve semantic\nsegmentation when fused through a semantic fusion gate (attention model). We\nperformed an extensive evaluation of our method over public data sets. We found\ncompetitive results on all data sets after comparing our proposed model with\nother twelve state-of-the-art segmenters, considering the same training\nconditions. Our proposed model achieved the best mIoU on the CityScapes,\nCamVid, and Pascal Context data sets, and the second best on Mapillary Vistas.",
+        "authors": "Jefferson Fontinele, Gabriel Lefundes, Luciano Oliveira",
+        "published": "2021-08-05",
+        "updated": "2021-08-05",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2207.04297v1",
+        "title": "SHDM-NET: Heat Map Detail Guidance with Image Matting for Industrial Weld Semantic Segmentation Network",
+        "abstract": "In actual industrial production, the assessment of the steel plate welding\neffect is an important task, and the segmentation of the weld section is the\nbasis of the assessment. This paper proposes an industrial weld segmentation\nnetwork based on a deep learning semantic segmentation algorithm fused with\nheatmap detail guidance and Image Matting to solve the automatic segmentation\nproblem of weld regions. In the existing semantic segmentation networks, the\nboundary information can be preserved by fusing the features of both high-level\nand low-level layers. However, this method can lead to insufficient expression\nof the spatial information in the low-level layer, resulting in inaccurate\nsegmentation boundary positioning. We propose a detailed guidance module based\non heatmaps to fully express the segmented region boundary information in the\nlow-level network to address this problem. Specifically, the expression of\nboundary information can be enhanced by adding a detailed branch to predict\nsegmented boundary and then matching it with the boundary heat map generated by\nmask labels to calculate the mean square error loss. In addition, although deep\nlearning has achieved great success in the field of semantic segmentation, the\nprecision of the segmentation boundary region is not high due to the loss of\ndetailed information caused by the classical segmentation network in the\nprocess of encoding and decoding process. This paper introduces a matting\nalgorithm to calibrate the boundary of the segmentation region of the semantic\nsegmentation network to solve this problem. Through many experiments on\nindustrial weld data sets, the effectiveness of our method is demonstrated, and\nthe MIOU reaches 97.93%. It is worth noting that this performance is comparable\nto human manual segmentation ( MIOU 97.96%).",
+        "authors": "Qi Wang, Jingwu Mei",
+        "published": "2022-07-09",
+        "updated": "2022-07-09",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2112.03241v1",
+        "title": "Unsupervised Domain Adaptation for Semantic Image Segmentation: a Comprehensive Survey",
+        "abstract": "Semantic segmentation plays a fundamental role in a broad variety of computer\nvision applications, providing key information for the global understanding of\nan image. Yet, the state-of-the-art models rely on large amount of annotated\nsamples, which are more expensive to obtain than in tasks such as image\nclassification. Since unlabelled data is instead significantly cheaper to\nobtain, it is not surprising that Unsupervised Domain Adaptation reached a\nbroad success within the semantic segmentation community.\n  This survey is an effort to summarize five years of this incredibly rapidly\ngrowing field, which embraces the importance of semantic segmentation itself\nand a critical need of adapting segmentation models to new environments. We\npresent the most important semantic segmentation methods; we provide a\ncomprehensive survey on domain adaptation techniques for semantic segmentation;\nwe unveil newer trends such as multi-domain learning, domain generalization,\ntest-time adaptation or source-free domain adaptation; we conclude this survey\nby describing datasets and benchmarks most widely used in semantic segmentation\nresearch. We hope that this survey will provide researchers across academia and\nindustry with a comprehensive reference guide and will help them in fostering\nnew research directions in the field.",
+        "authors": "Gabriela Csurka, Riccardo Volpi, Boris Chidlovskii",
+        "published": "2021-12-06",
+        "updated": "2021-12-06",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI",
+            "I.4.6; I.2"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1511.06988v1",
+        "title": "Learning High-level Prior with Convolutional Neural Networks for Semantic Segmentation",
+        "abstract": "This paper proposes a convolutional neural network that can fuse high-level\nprior for semantic image segmentation. Motivated by humans' vision recognition\nsystem, our key design is a three-layer generative structure consisting of\nhigh-level coding, middle-level segmentation and low-level image to introduce\nglobal prior for semantic segmentation. Based on this structure, we proposed a\ngenerative model called conditional variational auto-encoder (CVAE) that can\nbuild up the links behind these three layers. These important links include an\nimage encoder that extracts high level info from image, a segmentation encoder\nthat extracts high level info from segmentation, and a hybrid decoder that\noutputs semantic segmentation from the high level prior and input image. We\ntheoretically derive the semantic segmentation as an optimization problem\nparameterized by these links. Finally, the optimization problem enables us to\ntake advantage of state-of-the-art fully convolutional network structure for\nthe implementation of the above encoders and decoder. Experimental results on\nseveral representative datasets demonstrate our supreme performance for\nsemantic segmentation.",
+        "authors": "Haitian Zheng, Yebin Liu, Mengqi Ji, Feng Wu, Lu Fang",
+        "published": "2015-11-22",
+        "updated": "2015-11-22",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1803.10464v2",
+        "title": "Learning Pixel-level Semantic Affinity with Image-level Supervision for Weakly Supervised Semantic Segmentation",
+        "abstract": "The deficiency of segmentation labels is one of the main obstacles to\nsemantic segmentation in the wild. To alleviate this issue, we present a novel\nframework that generates segmentation labels of images given their image-level\nclass labels. In this weakly supervised setting, trained models have been known\nto segment local discriminative parts rather than the entire object area. Our\nsolution is to propagate such local responses to nearby areas which belong to\nthe same semantic entity. To this end, we propose a Deep Neural Network (DNN)\ncalled AffinityNet that predicts semantic affinity between a pair of adjacent\nimage coordinates. The semantic propagation is then realized by random walk\nwith the affinities predicted by AffinityNet. More importantly, the supervision\nemployed to train AffinityNet is given by the initial discriminative part\nsegmentation, which is incomplete as a segmentation annotation but sufficient\nfor learning semantic affinities within small image areas. Thus the entire\nframework relies only on image-level class labels and does not require any\nextra data or annotations. On the PASCAL VOC 2012 dataset, a DNN learned with\nsegmentation labels generated by our method outperforms previous models trained\nwith the same level of supervision, and is even as competitive as those relying\non stronger supervision.",
+        "authors": "Jiwoon Ahn, Suha Kwak",
+        "published": "2018-03-28",
+        "updated": "2018-04-09",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2304.11332v2",
+        "title": "Input Augmentation with SAM: Boosting Medical Image Segmentation with Segmentation Foundation Model",
+        "abstract": "The Segment Anything Model (SAM) is a recently developed large model for\ngeneral-purpose segmentation for computer vision tasks. SAM was trained using\n11 million images with over 1 billion masks and can produce segmentation\nresults for a wide range of objects in natural scene images. SAM can be viewed\nas a general perception model for segmentation (partitioning images into\nsemantically meaningful regions). Thus, how to utilize such a large foundation\nmodel for medical image segmentation is an emerging research target. This paper\nshows that although SAM does not immediately give high-quality segmentation for\nmedical image data, its generated masks, features, and stability scores are\nuseful for building and training better medical image segmentation models. In\nparticular, we demonstrate how to use SAM to augment image input for\ncommonly-used medical image segmentation models (e.g., U-Net). Experiments on\nthree segmentation tasks show the effectiveness of our proposed SAMAug method.\nThe code is available at \\url{https://github.com/yizhezhang2000/SAMAug}.",
+        "authors": "Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Danny Z. Chen",
+        "published": "2023-04-22",
+        "updated": "2023-06-21",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI",
+            "cs.LG"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2303.07898v5",
+        "title": "ISLE: A Framework for Image Level Semantic Segmentation Ensemble",
+        "abstract": "One key bottleneck of employing state-of-the-art semantic segmentation\nnetworks in the real world is the availability of training labels. Conventional\nsemantic segmentation networks require massive pixel-wise annotated labels to\nreach state-of-the-art prediction quality. Hence, several works focus on\nsemantic segmentation networks trained with only image-level annotations.\nHowever, when scrutinizing the results of state-of-the-art in more detail, we\nnotice that they are remarkably close to each other on average prediction\nquality, different approaches perform better in different classes while\nproviding low quality in others. To address this problem, we propose a novel\nframework, ISLE, which employs an ensemble of the \"pseudo-labels\" for a given\nset of different semantic segmentation techniques on a class-wise level.\nPseudo-labels are the pixel-wise predictions of the image-level semantic\nsegmentation frameworks used to train the final segmentation model. Our\npseudo-labels seamlessly combine the strong points of multiple segmentation\ntechniques approaches to reach superior prediction quality. We reach up to 2.4%\nimprovement over ISLE's individual components. An exhaustive analysis was\nperformed to demonstrate ISLE's effectiveness over state-of-the-art frameworks\nfor image-level semantic segmentation.",
+        "authors": "Erik Ostrowski, Muhammad Shafique",
+        "published": "2023-03-14",
+        "updated": "2023-09-20",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1904.05457v1",
+        "title": "Instance Segmentation based Semantic Matting for Compositing Applications",
+        "abstract": "Image compositing is a key step in film making and image editing that aims to\nsegment a foreground object and combine it with a new background. Automatic\nimage compositing can be done easily in a studio using chroma-keying when the\nbackground is pure blue or green. However, image compositing in natural scenes\nwith complex backgrounds remains a tedious task, requiring experienced artists\nto hand-segment. In order to achieve automatic compositing in natural scenes,\nwe propose a fully automated method that integrates instance segmentation and\nimage matting processes to generate high-quality semantic mattes that can be\nused for image editing task. Our approach can be seen both as a refinement of\nexisting instance segmentation algorithms and as a fully automated semantic\nimage matting method. It extends automatic image compositing techniques such as\nchroma-keying to scenes with complex natural backgrounds without the need for\nany kind of user interaction. The output of our approach can be considered as\nboth refined instance segmentations and alpha mattes with semantic meanings. We\nprovide experimental results which show improved performance results as\ncompared to existing approaches.",
+        "authors": "Guanqing Hu, James J. Clark",
+        "published": "2019-04-10",
+        "updated": "2019-04-10",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2402.04618v1",
+        "title": "Multi-Scale Semantic Segmentation with Modified MBConv Blocks",
+        "abstract": "Recently, MBConv blocks, initially designed for efficiency in\nresource-limited settings and later adapted for cutting-edge image\nclassification performances, have demonstrated significant potential in image\nclassification tasks. Despite their success, their application in semantic\nsegmentation has remained relatively unexplored. This paper introduces a novel\nadaptation of MBConv blocks specifically tailored for semantic segmentation.\nOur modification stems from the insight that semantic segmentation requires the\nextraction of more detailed spatial information than image classification. We\nargue that to effectively perform multi-scale semantic segmentation, each\nbranch of a U-Net architecture, regardless of its resolution, should possess\nequivalent segmentation capabilities. By implementing these changes, our\napproach achieves impressive mean Intersection over Union (IoU) scores of 84.5%\nand 84.0% on the Cityscapes test and validation datasets, respectively,\ndemonstrating the efficacy of our proposed modifications in enhancing semantic\nsegmentation performance.",
+        "authors": "Xi Chen, Yang Cai, Yuan Wu, Bo Xiong, Taesung Park",
+        "published": "2024-02-07",
+        "updated": "2024-02-07",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/1605.06885v1",
+        "title": "Bridging Category-level and Instance-level Semantic Image Segmentation",
+        "abstract": "We propose an approach to instance-level image segmentation that is built on\ntop of category-level segmentation. Specifically, for each pixel in a semantic\ncategory mask, its corresponding instance bounding box is predicted using a\ndeep fully convolutional regression network. Thus it follows a different\npipeline to the popular detect-then-segment approaches that first predict\ninstances' bounding boxes, which are the current state-of-the-art in instance\nsegmentation. We show that, by leveraging the strength of our state-of-the-art\nsemantic segmentation models, the proposed method can achieve comparable or\neven better results to detect-then-segment approaches. We make the following\ncontributions. (i) First, we propose a simple yet effective approach to\nsemantic instance segmentation. (ii) Second, we propose an online bootstrapping\nmethod during training, which is critically important for achieving good\nperformance for both semantic category segmentation and instance-level\nsegmentation. (iii) As the performance of semantic category segmentation has a\nsignificant impact on the instance-level segmentation, which is the second step\nof our approach, we train fully convolutional residual networks to achieve the\nbest semantic category segmentation accuracy. On the PASCAL VOC 2012 dataset,\nwe obtain the currently best mean intersection-over-union score of 79.1%. (iv)\nWe also achieve state-of-the-art results for instance-level segmentation.",
+        "authors": "Zifeng Wu, Chunhua Shen, Anton van den Hengel",
+        "published": "2016-05-23",
+        "updated": "2016-05-23",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2102.12095v1",
+        "title": "Synergy Between Semantic Segmentation and Image Denoising via Alternate Boosting",
+        "abstract": "The capability of image semantic segmentation may be deteriorated due to\nnoisy input image, where image denoising prior to segmentation helps. Both\nimage denoising and semantic segmentation have been developed significantly\nwith the advance of deep learning. Thus, we are interested in the synergy\nbetween them by using a holistic deep model. We observe that not only denoising\nhelps combat the drop of segmentation accuracy due to noise, but also\npixel-wise semantic information boosts the capability of denoising. We then\npropose a boosting network to perform denoising and segmentation alternately.\nThe proposed network is composed of multiple segmentation and denoising blocks\n(SDBs), each of which estimates semantic map then uses the map to regularize\ndenoising. Experimental results show that the denoised image quality is\nimproved substantially and the segmentation accuracy is improved to close to\nthat of clean images. Our code and models will be made publicly available.",
+        "authors": "Shunxin Xu, Ke Sun, Dong Liu, Zhiwei Xiong, Zheng-Jun Zha",
+        "published": "2021-02-24",
+        "updated": "2021-02-24",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2004.10126v3",
+        "title": "Generative Synthetic Augmentation using Label-to-Image Translation for Nuclei Image Segmentation",
+        "abstract": "In medical image diagnosis, pathology image analysis using semantic\nsegmentation becomes important for efficient screening as a field of digital\npathology. The spatial augmentation is ordinary used for semantic segmentation.\nTumor images under malignant are rare and to annotate the labels of nuclei\nregion takes much time-consuming. We require an effective use of dataset to\nmaximize the segmentation accuracy. It is expected that some augmentation to\ntransform generalized images influence the segmentation performance. We propose\na synthetic augmentation using label-to-image translation, mapping from a\nsemantic label with the edge structure to a real image. Exactly this paper deal\nwith stain slides of nuclei in tumor. Actually, we demonstrate several\nsegmentation algorithms applied to the initial dataset that contains real\nimages and labels using synthetic augmentation in order to add their\ngeneralized images. We computes and reports that a proposed synthetic\naugmentation procedure improve their accuracy.",
+        "authors": "Takato Yasuno",
+        "published": "2020-04-21",
+        "updated": "2021-03-02",
+        "primary_cat": "cs.LG",
+        "cats": [
+            "cs.LG",
+            "eess.IV",
+            "stat.ML",
+            "I.4.6; I.2.6"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2401.11739v1",
+        "title": "EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models",
+        "abstract": "Diffusion models have recently received increasing research attention for\ntheir remarkable transfer abilities in semantic segmentation tasks. However,\ngenerating fine-grained segmentation masks with diffusion models often requires\nadditional training on annotated datasets, leaving it unclear to what extent\npre-trained diffusion models alone understand the semantic relations of their\ngenerated images. To address this question, we leverage the semantic knowledge\nextracted from Stable Diffusion (SD) and aim to develop an image segmentor\ncapable of generating fine-grained segmentation maps without any additional\ntraining. The primary difficulty stems from the fact that semantically\nmeaningful feature maps typically exist only in the spatially lower-dimensional\nlayers, which poses a challenge in directly extracting pixel-level semantic\nrelations from these feature maps. To overcome this issue, our framework\nidentifies semantic correspondences between image pixels and spatial locations\nof low-dimensional feature maps by exploiting SD's generation process and\nutilizes them for constructing image-resolution segmentation maps. In extensive\nexperiments, the produced segmentation maps are demonstrated to be well\ndelineated and capture detailed parts of the images, indicating the existence\nof highly accurate pixel-level semantic knowledge in diffusion models.",
+        "authors": "Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, Seung Wook Kim",
+        "published": "2024-01-22",
+        "updated": "2024-01-22",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.LG"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2310.17874v1",
+        "title": "SmooSeg: Smoothness Prior for Unsupervised Semantic Segmentation",
+        "abstract": "Unsupervised semantic segmentation is a challenging task that segments images\ninto semantic groups without manual annotation. Prior works have primarily\nfocused on leveraging prior knowledge of semantic consistency or priori\nconcepts from self-supervised learning methods, which often overlook the\ncoherence property of image segments. In this paper, we demonstrate that the\nsmoothness prior, asserting that close features in a metric space share the\nsame semantics, can significantly simplify segmentation by casting unsupervised\nsemantic segmentation as an energy minimization problem. Under this paradigm,\nwe propose a novel approach called SmooSeg that harnesses self-supervised\nlearning methods to model the closeness relationships among observations as\nsmoothness signals. To effectively discover coherent semantic segments, we\nintroduce a novel smoothness loss that promotes piecewise smoothness within\nsegments while preserving discontinuities across different segments.\nAdditionally, to further enhance segmentation quality, we design an asymmetric\nteacher-student style predictor that generates smoothly updated pseudo labels,\nfacilitating an optimal fit between observations and labeling outputs. Thanks\nto the rich supervision cues of the smoothness prior, our SmooSeg significantly\noutperforms STEGO in terms of pixel accuracy on three datasets: COCOStuff\n(+14.9%), Cityscapes (+13.0%), and Potsdam-3 (+5.7%).",
+        "authors": "Mengcheng Lan, Xinjiang Wang, Yiping Ke, Jiaxing Xu, Litong Feng, Wayne Zhang",
+        "published": "2023-10-27",
+        "updated": "2023-10-27",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2305.15608v1",
+        "title": "Semantic Segmentation by Semantic Proportions",
+        "abstract": "Semantic segmentation is a critical task in computer vision that aims to\nidentify and classify individual pixels in an image, with numerous applications\nfor example autonomous driving and medical image analysis. However, semantic\nsegmentation can be super challenging particularly due to the need for large\namounts of annotated data. Annotating images is a time-consuming and costly\nprocess, often requiring expert knowledge and significant effort. In this\npaper, we propose a novel approach for semantic segmentation by eliminating the\nneed of ground-truth segmentation maps. Instead, our approach requires only the\nrough information of individual semantic class proportions, shortened as\nsemantic proportions. It greatly simplifies the data annotation process and\nthus will significantly reduce the annotation time and cost, making it more\nfeasible for large-scale applications. Moreover, it opens up new possibilities\nfor semantic segmentation tasks where obtaining the full ground-truth\nsegmentation maps may not be feasible or practical. Extensive experimental\nresults demonstrate that our approach can achieve comparable and sometimes even\nbetter performance against the benchmark method that relies on the ground-truth\nsegmentation maps. Utilising semantic proportions suggested in this work offers\na promising direction for future research in the field of semantic\nsegmentation.",
+        "authors": "Halil Ibrahim Aysel, Xiaohao Cai, Adam Pr\u00fcgel-Bennett",
+        "published": "2023-05-24",
+        "updated": "2023-05-24",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.AI"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2404.04807v1",
+        "title": "D2SL: Decouple Defogging and Semantic Learning for Foggy Domain-Adaptive Segmentation",
+        "abstract": "We investigated domain adaptive semantic segmentation in foggy weather\nscenarios, which aims to enhance the utilization of unlabeled foggy data and\nimprove the model's adaptability to foggy conditions. Current methods rely on\nclear images as references, jointly learning defogging and segmentation for\nfoggy images. Despite making some progress, there are still two main drawbacks:\n(1) the coupling of segmentation and defogging feature representations,\nresulting in a decrease in semantic representation capability, and (2) the\nfailure to leverage real fog priors in unlabeled foggy data, leading to\ninsufficient model generalization ability. To address these issues, we propose\na novel training framework, Decouple Defogging and Semantic learning, called\nD2SL, aiming to alleviate the adverse impact of defogging tasks on the final\nsegmentation task. In this framework, we introduce a domain-consistent transfer\nstrategy to establish a connection between defogging and segmentation tasks.\nFurthermore, we design a real fog transfer strategy to improve defogging\neffects by fully leveraging the fog priors from real foggy images. Our approach\nenhances the semantic representations required for segmentation during the\ndefogging learning process and maximizes the representation capability of fog\ninvariance by effectively utilizing real fog data. Comprehensive experiments\nvalidate the effectiveness of the proposed method.",
+        "authors": "Xuan Sun, Zhanfu An, Yuyu Liu",
+        "published": "2024-04-07",
+        "updated": "2024-04-07",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV",
+            "cs.MM"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2104.10834v1",
+        "title": "DANNet: A One-Stage Domain Adaptation Network for Unsupervised Nighttime Semantic Segmentation",
+        "abstract": "Semantic segmentation of nighttime images plays an equally important role as\nthat of daytime images in autonomous driving, but the former is much more\nchallenging due to poor illuminations and arduous human annotations. In this\npaper, we propose a novel domain adaptation network (DANNet) for nighttime\nsemantic segmentation without using labeled nighttime image data. It employs an\nadversarial training with a labeled daytime dataset and an unlabeled dataset\nthat contains coarsely aligned day-night image pairs. Specifically, for the\nunlabeled day-night image pairs, we use the pixel-level predictions of static\nobject categories on a daytime image as a pseudo supervision to segment its\ncounterpart nighttime image. We further design a re-weighting strategy to\nhandle the inaccuracy caused by misalignment between day-night image pairs and\nwrong predictions of daytime images, as well as boost the prediction accuracy\nof small objects. The proposed DANNet is the first one stage adaptation\nframework for nighttime semantic segmentation, which does not train additional\nday-night image transfer models as a separate pre-processing stage. Extensive\nexperiments on Dark Zurich and Nighttime Driving datasets show that our method\nachieves state-of-the-art performance for nighttime semantic segmentation.",
+        "authors": "Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, Song Wang",
+        "published": "2021-04-22",
+        "updated": "2021-04-22",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2404.04231v1",
+        "title": "Image-Text Co-Decomposition for Text-Supervised Semantic Segmentation",
+        "abstract": "This paper addresses text-supervised semantic segmentation, aiming to learn a\nmodel capable of segmenting arbitrary visual concepts within images by using\nonly image-text pairs without dense annotations. Existing methods have\ndemonstrated that contrastive learning on image-text pairs effectively aligns\nvisual segments with the meanings of texts. We notice that there is a\ndiscrepancy between text alignment and semantic segmentation: A text often\nconsists of multiple semantic concepts, whereas semantic segmentation strives\nto create semantically homogeneous segments. To address this issue, we propose\na novel framework, Image-Text Co-Decomposition (CoDe), where the paired image\nand text are jointly decomposed into a set of image regions and a set of word\nsegments, respectively, and contrastive learning is developed to enforce\nregion-word alignment. To work with a vision-language model, we present a\nprompt learning mechanism that derives an extra representation to highlight an\nimage segment or a word segment of interest, with which more effective features\ncan be extracted from that segment. Comprehensive experimental results\ndemonstrate that our method performs favorably against existing text-supervised\nsemantic segmentation methods on six benchmark datasets.",
+        "authors": "Ji-Jia Wu, Andy Chia-Hao Chang, Chieh-Yu Chuang, Chun-Pei Chen, Yu-Lun Liu, Min-Hung Chen, Hou-Ning Hu, Yung-Yu Chuang, Yen-Yu Lin",
+        "published": "2024-04-05",
+        "updated": "2024-04-05",
+        "primary_cat": "cs.CV",
+        "cats": [
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    },
+    {
+        "url": "http://arxiv.org/abs/2108.13118v1",
+        "title": "Automatic Preprocessing and Ensemble Learning for Low Quality Cell Image Segmentation",
+        "abstract": "We propose an automatic preprocessing and ensemble learning for segmentation\nof cell images with low quality. It is difficult to capture cells with strong\nlight. Therefore, the microscopic images of cells tend to have low image\nquality but these images are not good for semantic segmentation. Here we\npropose a method to translate an input image to the images that are easy to\nrecognize by deep learning. The proposed method consists of two deep neural\nnetworks. The first network is the usual training for semantic segmentation,\nand penultimate feature maps of the first network are used as filters to\ntranslate an input image to the images that emphasize each class. This is the\nautomatic preprocessing and translated cell images are easily classified. The\ninput cell image with low quality is translated by the feature maps in the\nfirst network, and the translated images are fed into the second network for\nsemantic segmentation. Since the outputs of the second network are multiple\nsegmentation results, we conduct the weighted ensemble of those segmentation\nimages. Two networks are trained by end-to-end manner, and we do not need to\nprepare images with high quality for the translation. We confirmed that our\nproposed method can translate cell images with low quality to the images that\nare easy to segment, and segmentation accuracy has improved using the weighted\nensemble learning.",
+        "authors": "Sota Kato, Kazuhiro Hotta",
+        "published": "2021-08-30",
+        "updated": "2021-08-30",
+        "primary_cat": "eess.IV",
+        "cats": [
+            "eess.IV",
+            "cs.CV"
+        ],
+        "category": "Semantic AND Segmentation AND Image"
+    }
+]
\ No newline at end of file