diff --git "a/abs_29K_G/test_abstract_long_2405.04534v1.json" "b/abs_29K_G/test_abstract_long_2405.04534v1.json"
new file mode 100644--- /dev/null
+++ "b/abs_29K_G/test_abstract_long_2405.04534v1.json"
@@ -0,0 +1,190 @@
+{
+    "url": "http://arxiv.org/abs/2405.04534v1",
+    "title": "Tactile-Augmented Radiance Fields",
+    "abstract": "We present a scene representation, which we call a tactile-augmented radiance\nfield (TaRF), that brings vision and touch into a shared 3D space. This\nrepresentation can be used to estimate the visual and tactile signals for a\ngiven 3D position within a scene. We capture a scene's TaRF from a collection\nof photos and sparsely sampled touch probes. Our approach makes use of two\ninsights: (i) common vision-based touch sensors are built on ordinary cameras\nand thus can be registered to images using methods from multi-view geometry,\nand (ii) visually and structurally similar regions of a scene share the same\ntactile features. We use these insights to register touch signals to a captured\nvisual scene, and to train a conditional diffusion model that, provided with an\nRGB-D image rendered from a neural radiance field, generates its corresponding\ntactile signal. To evaluate our approach, we collect a dataset of TaRFs. This\ndataset contains more touch samples than previous real-world datasets, and it\nprovides spatially aligned visual signals for each captured touch signal. We\ndemonstrate the accuracy of our cross-modal generative model and the utility of\nthe captured visual-tactile data on several downstream tasks. Project page:\nhttps://dou-yiming.github.io/TaRF",
+    "authors": "Yiming Dou, Fengyu Yang, Yi Liu, Antonio Loquercio, Andrew Owens",
+    "published": "2024-05-07",
+    "updated": "2024-05-07",
+    "primary_cat": "cs.CV",
+    "cats": [
+        "cs.CV"
+    ],
+    "label": "Original Paper",
+    "paper_cat": "Diffusion AND Model",
+    "gt": "We present a scene representation, which we call a tactile-augmented radiance\nfield (TaRF), that brings vision and touch into a shared 3D space. This\nrepresentation can be used to estimate the visual and tactile signals for a\ngiven 3D position within a scene. We capture a scene's TaRF from a collection\nof photos and sparsely sampled touch probes. Our approach makes use of two\ninsights: (i) common vision-based touch sensors are built on ordinary cameras\nand thus can be registered to images using methods from multi-view geometry,\nand (ii) visually and structurally similar regions of a scene share the same\ntactile features. We use these insights to register touch signals to a captured\nvisual scene, and to train a conditional diffusion model that, provided with an\nRGB-D image rendered from a neural radiance field, generates its corresponding\ntactile signal. To evaluate our approach, we collect a dataset of TaRFs. This\ndataset contains more touch samples than previous real-world datasets, and it\nprovides spatially aligned visual signals for each captured touch signal. We\ndemonstrate the accuracy of our cross-modal generative model and the utility of\nthe captured visual-tactile data on several downstream tasks. Project page:\nhttps://dou-yiming.github.io/TaRF",
+    "main_content": "Introduction As humans, our ability to perceive the world relies crucially on cross-modal associations between sight and touch [19, 50]. Tactile sensing provides a detailed understanding of material properties and microgeometry, such as the intricate patterns of bumps on rough surfaces and the complex motions that soft objects make when they deform. This type of understanding, which largely eludes today\u2019s computer vision models, is a critical component of applications that require reasoning about physical contact, such as robotic locomotion [3, 24, 31, 34, 37, 38] and manipulation [6, 7, 11, 42, 60], and methods that simulate the behavior of materials [4, 13, 40, 41]. In comparison to many other modalities, collecting tactile data is an expensive and tedious process, since it requires direct physical interaction with the environment. A recent line of work has addressed this problem by having humans or robots probe the environment with touch sensors (see Table 1). Early efforts have been focused on capturing the properties of only a few objects either in simulation [16, 17, 52] or in lab-controlled settings [6, 7, 18, 28, 35, 52, 63], which may not fully convey the diversity of tactile signals in natural environments. Other works have gone beyond a 1 arXiv:2405.04534v1  [cs.CV]  7 May 2024 \fDataset Samples Aligned Scenario Source More Than a Feeling [7] 6.5k \u2715 Tabletop Robot Feeling of Success [6] 9.3k \u2715 Tabletop Robot VisGel [35] 12k \u2715 Tabletop Robot SSVTP [28] 4.6k \u2713 Tabletop Robot ObjectFolder 1.0 [16] \u2013 \u2713 Object Synthetic ObjectFolder 2.0 [17] \u2013 \u2713 Object Synthetic ObjectFolder Real [18] 3.7k \u2715 Object Robot Burka et al. [5] 1.1k \u2715 Sub-scene Human Touch and Go [56] 13.9k \u2715 Sub-scene Human YCB-Slide\u2217[52] \u2713 Object Human Touching a NeRF [63] 1.2k \u2713 Object Robot TaRF (Ours) 19.3k \u2713 Full scene Human Table 1. Dataset comparison. We present the number of real visual-tactile pairs and whether such pairs are visually aligned, i.e., whether the visual image includes an occlusion-free view of the touched surface. \u2217YCB-Slide has real-world touch probes but synthetic images rendered with CAD models of YCB objects on a white background [9]. lab setting and have collected touch from real scenes [5, 56]. However, existing datasets lack aligned visual and tactile information, since the touch sensor and the person (or robot) that holds it often occlude large portions of the visual scene (Fig. 2). These datasets also contain only a sparse set of touch signals for each scene, and it is not clear how the sampled touch signals relate to each other in 3D. In this work, we present a simple and low-cost procedure to capture quasi-dense, scene-level, and spatially-aligned visual and touch data (Fig. 1). We call the resulting scene representation a tactile-augmented radiance field (TaRF). We remove the need for robotic collection by leveraging a 3D scene representation (a NeRF [39]) to synthesize a view of the surface being touched, which results in spatially aligned visual-tactile data (Fig. 2). We collect this data by mounting a touch sensor to a camera with commonly available materials (Fig. 3). To calibrate the pair of sensors, we take advantage of the fact that popular vision-based touch sensors [25, 26, 32, 48] are built on ordinary cameras. The relative pose between the vision and tactile sensors can thus be estimated using traditional methods from multi-view geometry, such as camera resectioning [20]. We use this procedure to collect a large real-world dataset of aligned visual-tactile data. With this dataset, we train a diffusion model [45, 51] to estimate touch at locations not directly probed by a sensor. In contrast to the recent work of Zhong et al. [63], which also estimates touch from 3D NeRF geometry, we create scene-scale reconstructions, we do not require robotic proprioception, and we use diffusion models [51]. This enables us to obtain tactile data at a much larger scale, and with considerably more diversity. Unlike previous visual-tactile diffusion work [57], we condition the model on spatially aligned visual and depth information, enhancing the generated samples\u2019 quality and their usefulness in downstream applications. After training, the diffusion model can be used to predict tactile informaOF 2.0 [17] VisGel [35] OF Real [18] SSVTP [28] TG [56] TaRF (Ours) Figure 2. Visual-tactile examples. In contrast to the visual-tactile data captured in previous work, our approach allows us to sample unobstructed images that are spatially aligned with the touch signal, from arbitrary 3D viewpoints using a NeRF. tion for novel positions in the scene. Analogous to quasidense stereo methods [15, 33], the diffusion model effectively propagates sparse touch samples, obtained by probing, to other visually and structurally similar regions of the scene. We evaluate our visual-tactile model\u2019s ability to accurately perform cross-modal translation using a variety of quality metrics. We also apply it to several downstream tasks, including localizing a touch within a scene and understanding material properties of the touched area. Our experiments suggest: \u2022 Touch signals can be localized in 3D space by exploiting multi-view geometry constraints between sight and touch. \u2022 Estimated touch measurements from novel views are not only qualitatively accurate, but also beneficial on downstream tasks. \u2022 Cross-modal prediction models can accurately estimate touch from sight for natural scenes. \u2022 Visually-acquired 3D scene geometry improves crossmodal prediction. 2. Related Work Visual-tactile datasets. Previous work has either used simulators [16, 17] or robotic arms [6, 8, 18, 35, 63] for data generation. Our work is closely related to that of Zhong et al. [63], which uses a NeRF and captured touch data to generate a tactile field for several small objects. They use the proprioception of an expensive robot to spatially align vision and touch. In contrast, we leverage the properties of the tactile sensor and novel view synthesis to use commonly available material (a smartphone and a selfie stick) to align vision and touch. This enables the collection of a larger, scene-level, and more diverse dataset, on which we train a higher-capacity diffusion model (rather than a conditional GAN). Like several previous works [5, 56], we also collect scene-level data. In contrast to them, we spatially align the signals by registering them in a unified 3D representation, thereby increasing the prediction power of the visual-tactile generative model. Capturing multimodal 3D scenes. Our work is related to methods that capture 3D visual reconstructions of spaces 2 \fusing RGB-D data [12, 49, 55, 59] and multimodal datasets of paired 3D vision and language [1, 2, 10]. Our work is also related to recent methods that localize objects in NeRFs using joint embeddings between images and language [29] or by semantic segmentation [62]. In contrast to language supervision, touch is tied to a precise position in a scene. 3D touch sensing. A variety of works have studied the close relationship between geometry and touch, motivating our use of geometry in imputing touch. Johnson et al. [25, 26] proposed vision-based touch sensing, and showed that highly accurate depth can be estimated from the touch sensor using photometric stereo. Other work has estimated object-scale 3D from touch [54]. By contrast, we combine sparse estimates of touch with quasi-dense tactile signals estimated using generative models. Cross-modal prediction of touch from sight. Recent work has trained generative models that predict touch from images. Li et al. [35] used a GAN to predict touch for images of a robotic arm, while Gao et al. [18] applied them to objects collected on a turntable. Yang et al. [57] used latent diffusion to predict touch from videos of humans touching objects. Our goal is different from these works: we want to predict touch signals that are spatially aligned with a visual signal, to exploit scene-specific information, and to use geometry. Thus, we use a different architecture and conditioning signal, and fit our model to examples from the same scenes at training and test time. Other work has learned joint embeddings between vision and touch [28, 36, 56, 58, 61]. 3. Method We collect visual and tactile examples from a scene and register them together with a 3D visual reconstruction to build a TaRF. Specifically, we capture a NeRF F\u03b8 : (x, r) 7\u2192(c, \u03c3) that maps a 3D point x = (x, y, z) and viewing direction r to its corresponding RGB color c and density \u03c3 [39]. We associate to the visual representation a touch model F\u03d5 : vt 7\u2192\u03c4 that generates the tactile signal that one would obtain by touching at the center of the image vt. In the following, we explain how to estimate F\u03b8 and F\u03d5 and put them into the same shared 3D space. 3.1. Capturing vision and touch signals Obtaining a visual 3D reconstruction. We build the visual NeRF, F\u03b8, closely following previous work [12, 55]. A human data collector moves through a scene and records a video, covering as much of the space as possible. We then estimate camera pose using structure from motion [47] and create a NeRF using off-the-shelf packages [53]. Additional details are provided in the supplement. Capturing and registering touch. We simultaneously collect tactile and visual signals by mounting a touch sensor Visual Camera Tactile Sensor Tactile frames Visual frames Visual-Tactile Correspondences Figure 3. Capturing setup. (a) We record paired vision and touch signals using a camera attached to a touch sensor. (b) We estimate the relative pose between the touch sensor and the camera using correspondences between sight and touch. on a camera (Fig. 3), obtaining synchronized touch signals {\u03c4 i}N i=1 and video frames v. We then estimate the pose of the video frames using off-the-shelf structure from motion methods [47], obtaining poses {pv i }N i=1. Finally, we use the calibration of the mount to obtain the poses {pt i}N i=1 of the tactile measurements with respect to the scene\u2019s global reference frame. As a collection device, we mount an iPhone 14 Pro to one end of a camera rod, and a DIGIT [32] touch sensor to the other end. Note that the devices can be replaced with any RGB-D camera and vision-based tactile sensor. Capturing setup calibration. To find the relative pose between the camera and the touch sensor (Fig. 3), we exploit the fact that arbitrary viewpoints can be synthesized from F\u03b8, and that ubiquitous vision-based touch sensors are based on perspective cameras. In these sensors, an elastomer gel is placed on the lens of a commodity camera, which is illuminated by colored lights. When the gel is pressed into an object, it deforms, and the camera records an image of the deformation; this image is used as the tactile signal. This design allows us to estimate the pose of the tactile sensor through multi-view constraints from visualtactile correspondences: pixels in visual images and tactile images that are of the same physical point. We start the calibration process by synthesizing novel views from F\u03b8. The views are generated at the camera location {pv i }N i=1, but rotated 90\u25e6on the x-axis. This is because the camera is approximately orthogonal to the touch sensor (see Fig. 3). Then, we manually annotate corresponding pixels between the touch measurements and the generated frames (Fig. 3). To simplify and standardize this process, we place a braille board in each scene and probe it with the touch sensor. This will generate a distinctive touch signal that is easy to localize [23]. We formulate the problem of estimating the six degrees of freedom relative pose (R, t) between the touch sensor and the generated frames as a resectioning problem [20]. We use the estimated 3D structure from the NeRF F\u03b8 to obtain 3D points {xi}M i=1 for each of the annotated corre3 \fspondences. Each point has a pixel position ui \u2208R2 in the touch measurement. We find (R, t) by minimizing the reprojection error:   \\ min   _ { { \\ma thbf R } , { \\ma t hbf t}} \\frac {1}{M}\\sum _{i=1}^M \\lVert \\pi ({\\mathbf K}[\\mathbf {R}\\,\\,|\\,\\,\\mathbf {t}], \\mathbf {X}_i) \\bu _i \\rVert _1,  (1) where \u03c0 projects a 3D point using a given projection matrix, K are the known intrinsics of the tactile sensor\u2019s camera, and the point Xi is in the coordinate system of the generated vision frames. We perform the optimization on 6-15 annotated correspondences from the braille board. For robustness, we compute correspondences from multiple frames. We represent the rotation matrix using quaternions and optimize using nonlinear least-squares. Once we have (R, t) with respect to the generated frames, we can derive the relative pose between the camera and the touch sensor. 3.2. Imputing the missing touch We use a generative model to estimate the touch signal (represented as an image from a vision-based touch sensor) for other locations within the scene. Specifically, we train a diffusion model p\u03d5(\u03c4 | v, d, b), where v and d are images and depth maps extracted from F\u03b8 (see Fig. 4). We also pass as input to the diffusion model a background image captured by the touch sensor when it is not in contact with anything, denoted as b. Although not essential, we have observed that this additional input empirically improves the model\u2019s performance (e.g., Fig. 1 the background provides the location of defects in the gel, which appear as black dots). We train the model p\u03d5 on our entire vision-touch dataset (Sec. 4). The training of p\u03d5 is divided into two stages. In the first, we pre-train a cross-modal visual-tactile encoder with self-supervised contrastive learning on our dataset. This stage, initially proposed by [23, 57], is equivalent to the self-supervised encoding pre-training that is common for image generation models [45]. We use a ResNet-50 [21] as the backbone for this contrastive model. In the second stage, we use the contrastive model to generate the input for a conditional latent diffusion model, which is built upon Stable Diffusion [45]. A frozen pretrained VQ-GAN [14] is used to obtain the latent representation with a spatial dimension of 64 \u00d7 64. We start training the diffusion model from scratch and pre-train it on the task of unconditional tactile image generation on the YCBSlide dataset [52]. After this stage, we train the conditional generative model p\u03d5 on our spatially aligned visual-tactile dataset, further fine-tuning the contrastive model end-to-end with the generation task. At inference time, given a novel location in the 3D scene, we first render the visual signals \u02c6 v and \u02c6 d from NeRF, and then estimate the touch signal \u02c6 \u03c4 of the position using the diffusion model. Latent Diffusion Gaussian Noise \u001f\u001e\u001e\u001e\u001e\u001d\u001e\u001e\u001e\u001e\u001c Depth RGB Est. Touch NeRF { Figure 4. Touch estimation. We estimate the tactile signal for a given touch sensor pose (R, t). To do this, we synthesize a viewpoint from the NeRF, along with a depth map. We use conditional latent diffusion to predict the tactile signal from these inputs. 4. A 3D Visual-Tactile Dataset In the following, we show the details of the data collection process and statistics of our dataset. 4.1. Data Collection Procedure The data collection procedure is divided into two stages. First, we collect multiple views from the scene, capturing enough frames around the areas we plan to touch. During this stage, we collect approximately 500 frames. Next, we collect synchronized visual and touch data, maximizing the geometry and texture being touched. We then estimate the camera location of the vision frames collected in the previous two stages using off-the-shelf mapping tools [47]. After estimating the camera poses for the vision frames, the touch measurements\u2019 poses can be derived by using the mount calibration matrix. More details about the pose estimation procedure can be found in the supplement. Finally, we associate each touch sensor with a color image by translating the sensor poses upwards by 0.4 meters and querying the NeRF with such poses. The field of view we use when querying the NeRF is 50\u25e6. This provides us with approximately 1,500 temporally aligned vision-touch image pairs per scene. Note that this collection procedure is scalable since it does not require specific expertise or equipment and generates abundant scene-level samples. 4.2. Dataset Statistics We collect our data in 13 ordinary scenes including two offices, a workroom, a conference room, a corridor, a tabletop, a corridor, a lounge, a room with various clothes and four outdoor scenes with interesting materials. Typically, we collect 1k to 2k tactile probes in each scene, resulting in a total of 19.3k image pairs in the dataset. Some representative samples from the collected dataset are shown in Fig. 5. Our data includes a large variety of geometry (edges, surfaces, corners, etc.) and texture (plastic, clothes, snow, wood, etc.) of different materials in the scene. During capturing process, the collector will try to 4 \fFigure 5. Representative examples from the captured dataset. Our dataset is obtained from nine everyday scenes, such as offices, classrooms, and kitchens. We show three such scenes in the figure above, together with samples of spatially aligned visual and tactile data. In each scene, 1k to 2k tactile probes were collected, resulting in a total of 19.3k image pairs. The data encompasses diverse geometries (edges, surfaces, corners, etc.) and textures (plastic, clothes, snow, wood, etc.) of various materials. The collector systematically probed different objects, covering areas with distinct geometry and texture using different sensor poses. thoroughly probe various objects and cover the interesting areas with more distinguishable geometry and texture with different sensor poses. To the best of our knowledge, our dataset is the first dataset that captures full, scene-scale spatially aligned vision-touch image pairs. We provide more details about the dataset in the supplement. 5. Experiments Leveraging the spatially aligned image and touch pairs from our dataset, we first conduct experiments on dense touch estimation. We then show the effectiveness of both the aligned data pairs and the synthesized touch signals by conducting tactile localization and material classification as two downstream tasks. 5.1. Implementation Details NeRF. We use the Nerfacto method from Nerfstudio [53]. For each scene, we utilize approximately 2,000 images as training set, which thoroughly cover the scene from various view points. We train the network with a base learning rate of 1 \u00d7 10\u22122 using Adam [30] optimizer for 200,000 steps on a single NVIDIA RTX 2080 Ti GPU to achieve optimal performance. Visual-tactile contrastive model. Following prior works [27, 57], we leverage contrastive learning methods to train a ResNet-50 [21] as visual encoder. The visual and tactile encoders share the same architecture but have different weights. We encode visual and tactile data into latent vectors in the resulting shared representation space. We set the dimension of the latent vectors to 32. Similar to CLIP [43], the model is trained on InfoNCE loss obtained from the pairwise dot products of the latent vectors. We train the model for 20 epochs by Adam [30] optimizer with a learning rate of 10\u22124 and batch size of 256 on 4 NVIDIA RTX 2080 Ti GPUs. Visual-tactile generative model. Our implementation of the diffusion model closely follows Stable Diffusion [46], with the difference that we use a ResNet-50 to generate the visual encoding from RGB-D images for conditioning. Specifically, we also add the RGB-D images rendered from the tactile sensors\u2019 poses into the conditioning, which we refer to in Sec. 5.2 as multiscale conditioning. The model is optimized for 30 epochs by Adam [30] optimizer with a base learning rate of 10\u22125. The learning rate is scaled by gpu number \u00d7 batch size. We train the model with batch size of 48 on 4 NVIDIA A40 GPUs. At inference time, the model conducts 200 steps of denoising process with a 7.5 guidance scale. Following prior cross-modal synthesis work [44], we use reranking to improve the prediction quality. We obtain 16 samples from the diffusion model for every instance and re-rank the samples with our pretrained contrastive model. The sample with highest similarity is the final prediction. 5.2. Dense Touch Estimation Experimental setup. We now evaluate the diffusion model\u2019s ability to generate touch images. To reduce overlap between the training and test set, we first split the frames into sequences temporally (following previous work [56]). We split them into sequences of 50 touch samples, then divide these sequences into train/validation/test with a ratio of 8/1/1. We evaluate the generated samples on Frechet Inception Distance (FID), a standard evaluation metric for cross-modal generation [56]. We also include Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM), though we note that these metrics are highly sensitive to spatial position of the generated content, and can be optimized by models that minimize simple pixelwise losses [22]. We also include CVTP metric proposed by prior work [57], which measures the similarity between visual and tactile embeddings of a contrastive model, analogous to 5 \fedge Condition VisGel Condition G.T. Ours L1 Ours G.T. L1 VisGel brick rock chair sofa desk wall surface desk carpet Figure 6. Qualitative touch estimation results. Each model is conditioned on the RGB image and depth map rendered from the NeRF (left). The white box indicates the tactile sensor\u2019s approximate field of view (which is much smaller than the full conditional image). The G.T. column shows the ground truth touch images measured from a DIGIT sensor. L1 and VisGel often generate blurry textures and inaccurate geometry. By contrast, our model better captures the features of the tactile image, e.g., the rock\u2019s microgeometry and complex textures and shapes of furniture. The last row shows two failure cases of our model. In both examples, our model generates a touch image that is geometrically misaligned with the ground truth. All of the examples shown here are at least 10cm away from any training sample. CLIP [43] score. We compare against two baselines: VisGel, the approach from Li et. [35], which trains a GAN for touch generation, and L1, a model with the same architecture of VisGel but trained to minimize an L1 loss in pixel space. Results. As is shown in Table 2, our approach performs much better on the high-level metrics, with up to 4x lower FID and 80x higher CVTP. This indicates that our proposed diffusion model captures the distribution and characteristics of the real tactile data more effectively. On the low-level metrics (PSNR and SSIM), all methods are comparable. In particular, the L1 model slightly outperforms the other methods since the loss it is trained on is highly correlated with low-level, pixel-wise metrics. Fig. 6 qualitatively compares samples from the different models. Indeed, our generated samples exhibit enhanced details in micro-geometry of fabrics and richer textures, including snow, wood and carpeting. However, all methods fail on fine details that are barely visible in the image, such as the tree bark. Ablation study. We evaluate the importance of the main components of our proposed touch generation approach (Table 3). Removing the conditioning on the RGB image results in the most prominent performance drop. This is expected since RGB image uniquely determines the fineModel PSNR \u2191 SSIM \u2191 FID \u2193 CVTP \u2191 L1 24.34 0.82 97.05 0.01 VisGel [35] 23.66 0.81 130.22 0.03 Ours 22.84 0.72 28.97 0.80 Table 2. Quantitative results on touch estimation for novel views. While comparable on low-level metrics with the baselines, our approach captures the characteristics of the real tactile data more effectively, resulting in a lower FID score. grained details of a tactile image. Removing depth image or contrastive pretraining has small effect on CVTP but results in a drop on FID. Contrastive re-ranking largely improves CVTP, indicating the necessity of obtaining multiple samples from the diffusion model. We also find that multiscale conditioning provide a small benefit on FID and CVTP. 5.3. Downstream Task I: Tactile Localization To help understand the quality of the captured TaRFs, we evaluate the performance of the contrastive model (used for conditioning our diffusion model) on the task of tactile localization. Given a tactile signal, our goal is to find the corresponding regions in a 2D image or in a 3D scene that are associated with it, i.e., we ask the question: what part of this image/scene feel like this? We perform the following 6 \fQuery Heatmap Query Query Heatmap Heatmap Query Heatmap Figure 7. Tactile localization heatmaps. Given a tactile query image, the heatmap shows the image patches with a higher affinity to this tactile signal, as measured by a contrastive model trained on our dataset. We use a sliding window and compare each extracted patch with the touch signal. In each case, the center patch is the true position. Our model successfully captures the correlation between the two signals. This enables it to localize a variety of touch signals, including fine-grained geometry, e.g., a cable or a keyboard, various types of corners and edges, and large uniform regions, such as a clothing. This ability enables our diffusion model to effectively propagate sparse touch samples to other visually and structurally similar regions of the scene. Model variation PSNR \u2191SSIM \u2191FID \u2193CVTP \u2191 Full 22.84 0.72 28.97 0.80 No RGB conditioning 22.13 0.70 34.31 0.76 No depth conditioning 22.57 0.71 33.16 0.80 No contrastive pretraining 22.82 0.71 32.98 0.79 No re-ranking 22.92 0.72 29.46 0.61 No multiscale 23.19 0.72 30.89 0.77 Table 3. Ablation study. Since the fine-grained details of touch images can be determined from a RGB image, removing conditioning on the latter results in the largest performance drops. Reranking has notable impact on CVTP, indicating the necessity of obtaining multiple samples from the diffusion model. evaluations on the test set of our dataset. Note that we run no task-specific training. 2D Localization. To determine which part of an image are associated with a given tactile measurement, we follow the same setup of SSVTP [28]. We first split the image into patches and compute their embedding. Then, we generate the tactile embedding of the input touch image. Finally, we compute the pairwise similarities between the tactile and visual embeddings, which we plot as a heatmap. As we can see in Fig. 7, our constrastive encoder can successfully capture the correlations between the visual and tactile data. For instance, the tactile embeddings of edges are associated to edges of similar shape in the visual image. Note that the majority of tactile embeddings are highly ambiguous: all edges with a similar geometry feel the same. 3D Localization. In 3D, the association of an image to tactile measurements becomes less ambiguous. Indeed, since tactile-visual samples are rotation-dependent, objects with similar shapes but different orientations will generate different tactile measurements. Lifting the task to 3D still does not remove all ambiguities (for example, each side of a rectangular table cannot be precisely localized). Nonetheless, we believe it to be a good fit for a quantitative evaluation since it\u2019s rare for two ambiguous parts of the scene to be touched with exactly the same orientation. We use the following experimental setup for 3D localization. Given a tactile image as a query, we compute its distance in embedding space to all visual test images from the same scene. Note that all test images are associated with a 3D location. We define as ground-truth correspondences all test images at a distance of at most r from the 3D location of the test sample. We vary r to account for local ambiguities. As typical in the retrieval literature, we benchmark the performance with metric mean Average Precision (mAP). We consider three baselines: (1) chance, which randomly selects corresponding samples; (2) real, which uses the contrastive model trained on our dataset; and (3) real + estimated, which trains the contrastive model on both dataset samples and a set of synthetic samples generated via the scenes\u2019 NeRF and our touch generation model. Specifically, we render a new image and corresponding touch by interpolating the position of two consecutive frames in the training dataset. This results in a training dataset for the contrastive model that is twice as large. 7 \fr(m) Dataset 0.001 0.005 0.01 0.05 0.1 Chance 3.55 6.82 10.25 18.26 21.33 Real 12.10 22.93 32.10 50.30 57.15 Real + Est. 14.92 26.69 36.17 53.62 60.61 Table 4. Quantitative results on 3D tactile localization. We evaluate using mean Average Precision (mAP) as a metric. Training the contrastive model on our dataset of visually aligned real samples together with estimated samples from new locations in the scene results in the highest performance. The results, presented in Table 4, demonstrate the performance benefit of employing both real and synthetic tactile pairs. Combining synthetic tactile images with the original pairs achieves highest performance on all distance thresholds. Overall, this indicates that touch measurements from novel views are not only qualitatively accurate, but also beneficial for this downstream task. 5.4. Downstream Task II: Material Classification We investigate the efficacy of our visual-tactile dataset for understanding material properties, focusing on the task of material classification. We follow the formulation by Yang et al. [56], which consists of three subtasks: (i) material classification, requiring the distinction of materials among 20 possible classes; (ii) softness classification, a binary problem dividing materials as either hard or soft; and (iii) hardness classification, which requires the classification of materials as either rough or smooth. We follow the same experimental procedure of [56]: we pretrain a contrastive model on a dataset and perform linear probing on the sub-tasks\u2019 training set. Our experiments only vary the pretraining dataset, leaving all architectural choices and hyperparameters the same. We compare against four baselines. A random classifier (chance); the ObjectFolder 2.0 dataset [17]; the VisGel dataset [35]; and the Touch and Go dataset [56]. Note that the touch sensor used in the test data (GelSight) differs from the one used in our dataset (DIGIT). Therefore, we use for pretraining a combination of our dataset and Touch and Go. To ensure a fair comparison, we also compare to the combination of each dataset and Touch and Go. The findings from this evaluation, as shown in Table 5, suggest that our data improves the effectiveness of the contrastive pretraining objective, even though our data is from a different distribution. Moreover, we find that adding estimated touch probes for pretraining results in a higher performance on all the three tasks, especially the smoothness classification. This indicates that not only does our dataset covers a wide range of materials but also our diffusion model captures the distinguishable and useful patterns of different materials. Dataset Material Hard/ Soft Rough/ Smooth Chance 18.6 66.1 56.3 ObjectFolder 2.0 [17] 36.2 72.0 69.0 VisGel [35] 39.1 69.4 70.4 Touch and Go [56] 54.7 77.3 79.4 + ObjectFolder 2.0 [17] 54.6 87.3 84.8 + VisGel [35] 53.1 86.7 83.6 + Ours\u2217(Real) 57.6 88.4 81.7 + Ours\u2217(Real + Estimated) 59.0 88.7 86.1 Table 5. Material classification. We show the downstream material recognition accuracy of models pre-trained on different datasets. The final rows show the performance when combining different datasets with Touch and Go [56]. \u2217The task-specific training and testing datasets for this task are collected with a GelSight sensor. We note that our data comes from a different distribution, since it is collected with a DIGIT sensor [32]. 6.",
+    "additional_graph_info": {
+        "graph": [
+            [
+                "Yiming Dou",
+                "Fengyu Yang"
+            ],
+            [
+                "Yiming Dou",
+                "Andrew Owens"
+            ],
+            [
+                "Fengyu Yang",
+                "Andrew Owens"
+            ],
+            [
+                "Fengyu Yang",
+                "Jiacheng Zhang"
+            ],
+            [
+                "Andrew Owens",
+                "Jiajun Wu"
+            ]
+        ],
+        "node_feat": {
+            "Yiming Dou": [
+                {
+                    "url": "http://arxiv.org/abs/2405.04534v1",
+                    "title": "Tactile-Augmented Radiance Fields",
+                    "abstract": "We present a scene representation, which we call a tactile-augmented radiance\nfield (TaRF), that brings vision and touch into a shared 3D space. This\nrepresentation can be used to estimate the visual and tactile signals for a\ngiven 3D position within a scene. We capture a scene's TaRF from a collection\nof photos and sparsely sampled touch probes. Our approach makes use of two\ninsights: (i) common vision-based touch sensors are built on ordinary cameras\nand thus can be registered to images using methods from multi-view geometry,\nand (ii) visually and structurally similar regions of a scene share the same\ntactile features. We use these insights to register touch signals to a captured\nvisual scene, and to train a conditional diffusion model that, provided with an\nRGB-D image rendered from a neural radiance field, generates its corresponding\ntactile signal. To evaluate our approach, we collect a dataset of TaRFs. This\ndataset contains more touch samples than previous real-world datasets, and it\nprovides spatially aligned visual signals for each captured touch signal. We\ndemonstrate the accuracy of our cross-modal generative model and the utility of\nthe captured visual-tactile data on several downstream tasks. Project page:\nhttps://dou-yiming.github.io/TaRF",
+                    "authors": "Yiming Dou, Fengyu Yang, Yi Liu, Antonio Loquercio, Andrew Owens",
+                    "published": "2024-05-07",
+                    "updated": "2024-05-07",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction As humans, our ability to perceive the world relies crucially on cross-modal associations between sight and touch [19, 50]. Tactile sensing provides a detailed understanding of material properties and microgeometry, such as the intricate patterns of bumps on rough surfaces and the complex motions that soft objects make when they deform. This type of understanding, which largely eludes today\u2019s computer vision models, is a critical component of applications that require reasoning about physical contact, such as robotic locomotion [3, 24, 31, 34, 37, 38] and manipulation [6, 7, 11, 42, 60], and methods that simulate the behavior of materials [4, 13, 40, 41]. In comparison to many other modalities, collecting tactile data is an expensive and tedious process, since it requires direct physical interaction with the environment. A recent line of work has addressed this problem by having humans or robots probe the environment with touch sensors (see Table 1). Early efforts have been focused on capturing the properties of only a few objects either in simulation [16, 17, 52] or in lab-controlled settings [6, 7, 18, 28, 35, 52, 63], which may not fully convey the diversity of tactile signals in natural environments. Other works have gone beyond a 1 arXiv:2405.04534v1  [cs.CV]  7 May 2024 \fDataset Samples Aligned Scenario Source More Than a Feeling [7] 6.5k \u2715 Tabletop Robot Feeling of Success [6] 9.3k \u2715 Tabletop Robot VisGel [35] 12k \u2715 Tabletop Robot SSVTP [28] 4.6k \u2713 Tabletop Robot ObjectFolder 1.0 [16] \u2013 \u2713 Object Synthetic ObjectFolder 2.0 [17] \u2013 \u2713 Object Synthetic ObjectFolder Real [18] 3.7k \u2715 Object Robot Burka et al. [5] 1.1k \u2715 Sub-scene Human Touch and Go [56] 13.9k \u2715 Sub-scene Human YCB-Slide\u2217[52] \u2713 Object Human Touching a NeRF [63] 1.2k \u2713 Object Robot TaRF (Ours) 19.3k \u2713 Full scene Human Table 1. Dataset comparison. We present the number of real visual-tactile pairs and whether such pairs are visually aligned, i.e., whether the visual image includes an occlusion-free view of the touched surface. \u2217YCB-Slide has real-world touch probes but synthetic images rendered with CAD models of YCB objects on a white background [9]. lab setting and have collected touch from real scenes [5, 56]. However, existing datasets lack aligned visual and tactile information, since the touch sensor and the person (or robot) that holds it often occlude large portions of the visual scene (Fig. 2). These datasets also contain only a sparse set of touch signals for each scene, and it is not clear how the sampled touch signals relate to each other in 3D. In this work, we present a simple and low-cost procedure to capture quasi-dense, scene-level, and spatially-aligned visual and touch data (Fig. 1). We call the resulting scene representation a tactile-augmented radiance field (TaRF). We remove the need for robotic collection by leveraging a 3D scene representation (a NeRF [39]) to synthesize a view of the surface being touched, which results in spatially aligned visual-tactile data (Fig. 2). We collect this data by mounting a touch sensor to a camera with commonly available materials (Fig. 3). To calibrate the pair of sensors, we take advantage of the fact that popular vision-based touch sensors [25, 26, 32, 48] are built on ordinary cameras. The relative pose between the vision and tactile sensors can thus be estimated using traditional methods from multi-view geometry, such as camera resectioning [20]. We use this procedure to collect a large real-world dataset of aligned visual-tactile data. With this dataset, we train a diffusion model [45, 51] to estimate touch at locations not directly probed by a sensor. In contrast to the recent work of Zhong et al. [63], which also estimates touch from 3D NeRF geometry, we create scene-scale reconstructions, we do not require robotic proprioception, and we use diffusion models [51]. This enables us to obtain tactile data at a much larger scale, and with considerably more diversity. Unlike previous visual-tactile diffusion work [57], we condition the model on spatially aligned visual and depth information, enhancing the generated samples\u2019 quality and their usefulness in downstream applications. After training, the diffusion model can be used to predict tactile informaOF 2.0 [17] VisGel [35] OF Real [18] SSVTP [28] TG [56] TaRF (Ours) Figure 2. Visual-tactile examples. In contrast to the visual-tactile data captured in previous work, our approach allows us to sample unobstructed images that are spatially aligned with the touch signal, from arbitrary 3D viewpoints using a NeRF. tion for novel positions in the scene. Analogous to quasidense stereo methods [15, 33], the diffusion model effectively propagates sparse touch samples, obtained by probing, to other visually and structurally similar regions of the scene. We evaluate our visual-tactile model\u2019s ability to accurately perform cross-modal translation using a variety of quality metrics. We also apply it to several downstream tasks, including localizing a touch within a scene and understanding material properties of the touched area. Our experiments suggest: \u2022 Touch signals can be localized in 3D space by exploiting multi-view geometry constraints between sight and touch. \u2022 Estimated touch measurements from novel views are not only qualitatively accurate, but also beneficial on downstream tasks. \u2022 Cross-modal prediction models can accurately estimate touch from sight for natural scenes. \u2022 Visually-acquired 3D scene geometry improves crossmodal prediction. 2. Related Work Visual-tactile datasets. Previous work has either used simulators [16, 17] or robotic arms [6, 8, 18, 35, 63] for data generation. Our work is closely related to that of Zhong et al. [63], which uses a NeRF and captured touch data to generate a tactile field for several small objects. They use the proprioception of an expensive robot to spatially align vision and touch. In contrast, we leverage the properties of the tactile sensor and novel view synthesis to use commonly available material (a smartphone and a selfie stick) to align vision and touch. This enables the collection of a larger, scene-level, and more diverse dataset, on which we train a higher-capacity diffusion model (rather than a conditional GAN). Like several previous works [5, 56], we also collect scene-level data. In contrast to them, we spatially align the signals by registering them in a unified 3D representation, thereby increasing the prediction power of the visual-tactile generative model. Capturing multimodal 3D scenes. Our work is related to methods that capture 3D visual reconstructions of spaces 2 \fusing RGB-D data [12, 49, 55, 59] and multimodal datasets of paired 3D vision and language [1, 2, 10]. Our work is also related to recent methods that localize objects in NeRFs using joint embeddings between images and language [29] or by semantic segmentation [62]. In contrast to language supervision, touch is tied to a precise position in a scene. 3D touch sensing. A variety of works have studied the close relationship between geometry and touch, motivating our use of geometry in imputing touch. Johnson et al. [25, 26] proposed vision-based touch sensing, and showed that highly accurate depth can be estimated from the touch sensor using photometric stereo. Other work has estimated object-scale 3D from touch [54]. By contrast, we combine sparse estimates of touch with quasi-dense tactile signals estimated using generative models. Cross-modal prediction of touch from sight. Recent work has trained generative models that predict touch from images. Li et al. [35] used a GAN to predict touch for images of a robotic arm, while Gao et al. [18] applied them to objects collected on a turntable. Yang et al. [57] used latent diffusion to predict touch from videos of humans touching objects. Our goal is different from these works: we want to predict touch signals that are spatially aligned with a visual signal, to exploit scene-specific information, and to use geometry. Thus, we use a different architecture and conditioning signal, and fit our model to examples from the same scenes at training and test time. Other work has learned joint embeddings between vision and touch [28, 36, 56, 58, 61]. 3. Method We collect visual and tactile examples from a scene and register them together with a 3D visual reconstruction to build a TaRF. Specifically, we capture a NeRF F\u03b8 : (x, r) 7\u2192(c, \u03c3) that maps a 3D point x = (x, y, z) and viewing direction r to its corresponding RGB color c and density \u03c3 [39]. We associate to the visual representation a touch model F\u03d5 : vt 7\u2192\u03c4 that generates the tactile signal that one would obtain by touching at the center of the image vt. In the following, we explain how to estimate F\u03b8 and F\u03d5 and put them into the same shared 3D space. 3.1. Capturing vision and touch signals Obtaining a visual 3D reconstruction. We build the visual NeRF, F\u03b8, closely following previous work [12, 55]. A human data collector moves through a scene and records a video, covering as much of the space as possible. We then estimate camera pose using structure from motion [47] and create a NeRF using off-the-shelf packages [53]. Additional details are provided in the supplement. Capturing and registering touch. We simultaneously collect tactile and visual signals by mounting a touch sensor Visual Camera Tactile Sensor Tactile frames Visual frames Visual-Tactile Correspondences Figure 3. Capturing setup. (a) We record paired vision and touch signals using a camera attached to a touch sensor. (b) We estimate the relative pose between the touch sensor and the camera using correspondences between sight and touch. on a camera (Fig. 3), obtaining synchronized touch signals {\u03c4 i}N i=1 and video frames v. We then estimate the pose of the video frames using off-the-shelf structure from motion methods [47], obtaining poses {pv i }N i=1. Finally, we use the calibration of the mount to obtain the poses {pt i}N i=1 of the tactile measurements with respect to the scene\u2019s global reference frame. As a collection device, we mount an iPhone 14 Pro to one end of a camera rod, and a DIGIT [32] touch sensor to the other end. Note that the devices can be replaced with any RGB-D camera and vision-based tactile sensor. Capturing setup calibration. To find the relative pose between the camera and the touch sensor (Fig. 3), we exploit the fact that arbitrary viewpoints can be synthesized from F\u03b8, and that ubiquitous vision-based touch sensors are based on perspective cameras. In these sensors, an elastomer gel is placed on the lens of a commodity camera, which is illuminated by colored lights. When the gel is pressed into an object, it deforms, and the camera records an image of the deformation; this image is used as the tactile signal. This design allows us to estimate the pose of the tactile sensor through multi-view constraints from visualtactile correspondences: pixels in visual images and tactile images that are of the same physical point. We start the calibration process by synthesizing novel views from F\u03b8. The views are generated at the camera location {pv i }N i=1, but rotated 90\u25e6on the x-axis. This is because the camera is approximately orthogonal to the touch sensor (see Fig. 3). Then, we manually annotate corresponding pixels between the touch measurements and the generated frames (Fig. 3). To simplify and standardize this process, we place a braille board in each scene and probe it with the touch sensor. This will generate a distinctive touch signal that is easy to localize [23]. We formulate the problem of estimating the six degrees of freedom relative pose (R, t) between the touch sensor and the generated frames as a resectioning problem [20]. We use the estimated 3D structure from the NeRF F\u03b8 to obtain 3D points {xi}M i=1 for each of the annotated corre3 \fspondences. Each point has a pixel position ui \u2208R2 in the touch measurement. We find (R, t) by minimizing the reprojection error:   \\ min   _ { { \\ma thbf R } , { \\ma t hbf t}} \\frac {1}{M}\\sum _{i=1}^M \\lVert \\pi ({\\mathbf K}[\\mathbf {R}\\,\\,|\\,\\,\\mathbf {t}], \\mathbf {X}_i) \\bu _i \\rVert _1,  (1) where \u03c0 projects a 3D point using a given projection matrix, K are the known intrinsics of the tactile sensor\u2019s camera, and the point Xi is in the coordinate system of the generated vision frames. We perform the optimization on 6-15 annotated correspondences from the braille board. For robustness, we compute correspondences from multiple frames. We represent the rotation matrix using quaternions and optimize using nonlinear least-squares. Once we have (R, t) with respect to the generated frames, we can derive the relative pose between the camera and the touch sensor. 3.2. Imputing the missing touch We use a generative model to estimate the touch signal (represented as an image from a vision-based touch sensor) for other locations within the scene. Specifically, we train a diffusion model p\u03d5(\u03c4 | v, d, b), where v and d are images and depth maps extracted from F\u03b8 (see Fig. 4). We also pass as input to the diffusion model a background image captured by the touch sensor when it is not in contact with anything, denoted as b. Although not essential, we have observed that this additional input empirically improves the model\u2019s performance (e.g., Fig. 1 the background provides the location of defects in the gel, which appear as black dots). We train the model p\u03d5 on our entire vision-touch dataset (Sec. 4). The training of p\u03d5 is divided into two stages. In the first, we pre-train a cross-modal visual-tactile encoder with self-supervised contrastive learning on our dataset. This stage, initially proposed by [23, 57], is equivalent to the self-supervised encoding pre-training that is common for image generation models [45]. We use a ResNet-50 [21] as the backbone for this contrastive model. In the second stage, we use the contrastive model to generate the input for a conditional latent diffusion model, which is built upon Stable Diffusion [45]. A frozen pretrained VQ-GAN [14] is used to obtain the latent representation with a spatial dimension of 64 \u00d7 64. We start training the diffusion model from scratch and pre-train it on the task of unconditional tactile image generation on the YCBSlide dataset [52]. After this stage, we train the conditional generative model p\u03d5 on our spatially aligned visual-tactile dataset, further fine-tuning the contrastive model end-to-end with the generation task. At inference time, given a novel location in the 3D scene, we first render the visual signals \u02c6 v and \u02c6 d from NeRF, and then estimate the touch signal \u02c6 \u03c4 of the position using the diffusion model. Latent Diffusion Gaussian Noise \u001f\u001e\u001e\u001e\u001e\u001d\u001e\u001e\u001e\u001e\u001c Depth RGB Est. Touch NeRF { Figure 4. Touch estimation. We estimate the tactile signal for a given touch sensor pose (R, t). To do this, we synthesize a viewpoint from the NeRF, along with a depth map. We use conditional latent diffusion to predict the tactile signal from these inputs. 4. A 3D Visual-Tactile Dataset In the following, we show the details of the data collection process and statistics of our dataset. 4.1. Data Collection Procedure The data collection procedure is divided into two stages. First, we collect multiple views from the scene, capturing enough frames around the areas we plan to touch. During this stage, we collect approximately 500 frames. Next, we collect synchronized visual and touch data, maximizing the geometry and texture being touched. We then estimate the camera location of the vision frames collected in the previous two stages using off-the-shelf mapping tools [47]. After estimating the camera poses for the vision frames, the touch measurements\u2019 poses can be derived by using the mount calibration matrix. More details about the pose estimation procedure can be found in the supplement. Finally, we associate each touch sensor with a color image by translating the sensor poses upwards by 0.4 meters and querying the NeRF with such poses. The field of view we use when querying the NeRF is 50\u25e6. This provides us with approximately 1,500 temporally aligned vision-touch image pairs per scene. Note that this collection procedure is scalable since it does not require specific expertise or equipment and generates abundant scene-level samples. 4.2. Dataset Statistics We collect our data in 13 ordinary scenes including two offices, a workroom, a conference room, a corridor, a tabletop, a corridor, a lounge, a room with various clothes and four outdoor scenes with interesting materials. Typically, we collect 1k to 2k tactile probes in each scene, resulting in a total of 19.3k image pairs in the dataset. Some representative samples from the collected dataset are shown in Fig. 5. Our data includes a large variety of geometry (edges, surfaces, corners, etc.) and texture (plastic, clothes, snow, wood, etc.) of different materials in the scene. During capturing process, the collector will try to 4 \fFigure 5. Representative examples from the captured dataset. Our dataset is obtained from nine everyday scenes, such as offices, classrooms, and kitchens. We show three such scenes in the figure above, together with samples of spatially aligned visual and tactile data. In each scene, 1k to 2k tactile probes were collected, resulting in a total of 19.3k image pairs. The data encompasses diverse geometries (edges, surfaces, corners, etc.) and textures (plastic, clothes, snow, wood, etc.) of various materials. The collector systematically probed different objects, covering areas with distinct geometry and texture using different sensor poses. thoroughly probe various objects and cover the interesting areas with more distinguishable geometry and texture with different sensor poses. To the best of our knowledge, our dataset is the first dataset that captures full, scene-scale spatially aligned vision-touch image pairs. We provide more details about the dataset in the supplement. 5. Experiments Leveraging the spatially aligned image and touch pairs from our dataset, we first conduct experiments on dense touch estimation. We then show the effectiveness of both the aligned data pairs and the synthesized touch signals by conducting tactile localization and material classification as two downstream tasks. 5.1. Implementation Details NeRF. We use the Nerfacto method from Nerfstudio [53]. For each scene, we utilize approximately 2,000 images as training set, which thoroughly cover the scene from various view points. We train the network with a base learning rate of 1 \u00d7 10\u22122 using Adam [30] optimizer for 200,000 steps on a single NVIDIA RTX 2080 Ti GPU to achieve optimal performance. Visual-tactile contrastive model. Following prior works [27, 57], we leverage contrastive learning methods to train a ResNet-50 [21] as visual encoder. The visual and tactile encoders share the same architecture but have different weights. We encode visual and tactile data into latent vectors in the resulting shared representation space. We set the dimension of the latent vectors to 32. Similar to CLIP [43], the model is trained on InfoNCE loss obtained from the pairwise dot products of the latent vectors. We train the model for 20 epochs by Adam [30] optimizer with a learning rate of 10\u22124 and batch size of 256 on 4 NVIDIA RTX 2080 Ti GPUs. Visual-tactile generative model. Our implementation of the diffusion model closely follows Stable Diffusion [46], with the difference that we use a ResNet-50 to generate the visual encoding from RGB-D images for conditioning. Specifically, we also add the RGB-D images rendered from the tactile sensors\u2019 poses into the conditioning, which we refer to in Sec. 5.2 as multiscale conditioning. The model is optimized for 30 epochs by Adam [30] optimizer with a base learning rate of 10\u22125. The learning rate is scaled by gpu number \u00d7 batch size. We train the model with batch size of 48 on 4 NVIDIA A40 GPUs. At inference time, the model conducts 200 steps of denoising process with a 7.5 guidance scale. Following prior cross-modal synthesis work [44], we use reranking to improve the prediction quality. We obtain 16 samples from the diffusion model for every instance and re-rank the samples with our pretrained contrastive model. The sample with highest similarity is the final prediction. 5.2. Dense Touch Estimation Experimental setup. We now evaluate the diffusion model\u2019s ability to generate touch images. To reduce overlap between the training and test set, we first split the frames into sequences temporally (following previous work [56]). We split them into sequences of 50 touch samples, then divide these sequences into train/validation/test with a ratio of 8/1/1. We evaluate the generated samples on Frechet Inception Distance (FID), a standard evaluation metric for cross-modal generation [56]. We also include Peak Signal to Noise Ratio (PSNR) and Structural Similarity (SSIM), though we note that these metrics are highly sensitive to spatial position of the generated content, and can be optimized by models that minimize simple pixelwise losses [22]. We also include CVTP metric proposed by prior work [57], which measures the similarity between visual and tactile embeddings of a contrastive model, analogous to 5 \fedge Condition VisGel Condition G.T. Ours L1 Ours G.T. L1 VisGel brick rock chair sofa desk wall surface desk carpet Figure 6. Qualitative touch estimation results. Each model is conditioned on the RGB image and depth map rendered from the NeRF (left). The white box indicates the tactile sensor\u2019s approximate field of view (which is much smaller than the full conditional image). The G.T. column shows the ground truth touch images measured from a DIGIT sensor. L1 and VisGel often generate blurry textures and inaccurate geometry. By contrast, our model better captures the features of the tactile image, e.g., the rock\u2019s microgeometry and complex textures and shapes of furniture. The last row shows two failure cases of our model. In both examples, our model generates a touch image that is geometrically misaligned with the ground truth. All of the examples shown here are at least 10cm away from any training sample. CLIP [43] score. We compare against two baselines: VisGel, the approach from Li et. [35], which trains a GAN for touch generation, and L1, a model with the same architecture of VisGel but trained to minimize an L1 loss in pixel space. Results. As is shown in Table 2, our approach performs much better on the high-level metrics, with up to 4x lower FID and 80x higher CVTP. This indicates that our proposed diffusion model captures the distribution and characteristics of the real tactile data more effectively. On the low-level metrics (PSNR and SSIM), all methods are comparable. In particular, the L1 model slightly outperforms the other methods since the loss it is trained on is highly correlated with low-level, pixel-wise metrics. Fig. 6 qualitatively compares samples from the different models. Indeed, our generated samples exhibit enhanced details in micro-geometry of fabrics and richer textures, including snow, wood and carpeting. However, all methods fail on fine details that are barely visible in the image, such as the tree bark. Ablation study. We evaluate the importance of the main components of our proposed touch generation approach (Table 3). Removing the conditioning on the RGB image results in the most prominent performance drop. This is expected since RGB image uniquely determines the fineModel PSNR \u2191 SSIM \u2191 FID \u2193 CVTP \u2191 L1 24.34 0.82 97.05 0.01 VisGel [35] 23.66 0.81 130.22 0.03 Ours 22.84 0.72 28.97 0.80 Table 2. Quantitative results on touch estimation for novel views. While comparable on low-level metrics with the baselines, our approach captures the characteristics of the real tactile data more effectively, resulting in a lower FID score. grained details of a tactile image. Removing depth image or contrastive pretraining has small effect on CVTP but results in a drop on FID. Contrastive re-ranking largely improves CVTP, indicating the necessity of obtaining multiple samples from the diffusion model. We also find that multiscale conditioning provide a small benefit on FID and CVTP. 5.3. Downstream Task I: Tactile Localization To help understand the quality of the captured TaRFs, we evaluate the performance of the contrastive model (used for conditioning our diffusion model) on the task of tactile localization. Given a tactile signal, our goal is to find the corresponding regions in a 2D image or in a 3D scene that are associated with it, i.e., we ask the question: what part of this image/scene feel like this? We perform the following 6 \fQuery Heatmap Query Query Heatmap Heatmap Query Heatmap Figure 7. Tactile localization heatmaps. Given a tactile query image, the heatmap shows the image patches with a higher affinity to this tactile signal, as measured by a contrastive model trained on our dataset. We use a sliding window and compare each extracted patch with the touch signal. In each case, the center patch is the true position. Our model successfully captures the correlation between the two signals. This enables it to localize a variety of touch signals, including fine-grained geometry, e.g., a cable or a keyboard, various types of corners and edges, and large uniform regions, such as a clothing. This ability enables our diffusion model to effectively propagate sparse touch samples to other visually and structurally similar regions of the scene. Model variation PSNR \u2191SSIM \u2191FID \u2193CVTP \u2191 Full 22.84 0.72 28.97 0.80 No RGB conditioning 22.13 0.70 34.31 0.76 No depth conditioning 22.57 0.71 33.16 0.80 No contrastive pretraining 22.82 0.71 32.98 0.79 No re-ranking 22.92 0.72 29.46 0.61 No multiscale 23.19 0.72 30.89 0.77 Table 3. Ablation study. Since the fine-grained details of touch images can be determined from a RGB image, removing conditioning on the latter results in the largest performance drops. Reranking has notable impact on CVTP, indicating the necessity of obtaining multiple samples from the diffusion model. evaluations on the test set of our dataset. Note that we run no task-specific training. 2D Localization. To determine which part of an image are associated with a given tactile measurement, we follow the same setup of SSVTP [28]. We first split the image into patches and compute their embedding. Then, we generate the tactile embedding of the input touch image. Finally, we compute the pairwise similarities between the tactile and visual embeddings, which we plot as a heatmap. As we can see in Fig. 7, our constrastive encoder can successfully capture the correlations between the visual and tactile data. For instance, the tactile embeddings of edges are associated to edges of similar shape in the visual image. Note that the majority of tactile embeddings are highly ambiguous: all edges with a similar geometry feel the same. 3D Localization. In 3D, the association of an image to tactile measurements becomes less ambiguous. Indeed, since tactile-visual samples are rotation-dependent, objects with similar shapes but different orientations will generate different tactile measurements. Lifting the task to 3D still does not remove all ambiguities (for example, each side of a rectangular table cannot be precisely localized). Nonetheless, we believe it to be a good fit for a quantitative evaluation since it\u2019s rare for two ambiguous parts of the scene to be touched with exactly the same orientation. We use the following experimental setup for 3D localization. Given a tactile image as a query, we compute its distance in embedding space to all visual test images from the same scene. Note that all test images are associated with a 3D location. We define as ground-truth correspondences all test images at a distance of at most r from the 3D location of the test sample. We vary r to account for local ambiguities. As typical in the retrieval literature, we benchmark the performance with metric mean Average Precision (mAP). We consider three baselines: (1) chance, which randomly selects corresponding samples; (2) real, which uses the contrastive model trained on our dataset; and (3) real + estimated, which trains the contrastive model on both dataset samples and a set of synthetic samples generated via the scenes\u2019 NeRF and our touch generation model. Specifically, we render a new image and corresponding touch by interpolating the position of two consecutive frames in the training dataset. This results in a training dataset for the contrastive model that is twice as large. 7 \fr(m) Dataset 0.001 0.005 0.01 0.05 0.1 Chance 3.55 6.82 10.25 18.26 21.33 Real 12.10 22.93 32.10 50.30 57.15 Real + Est. 14.92 26.69 36.17 53.62 60.61 Table 4. Quantitative results on 3D tactile localization. We evaluate using mean Average Precision (mAP) as a metric. Training the contrastive model on our dataset of visually aligned real samples together with estimated samples from new locations in the scene results in the highest performance. The results, presented in Table 4, demonstrate the performance benefit of employing both real and synthetic tactile pairs. Combining synthetic tactile images with the original pairs achieves highest performance on all distance thresholds. Overall, this indicates that touch measurements from novel views are not only qualitatively accurate, but also beneficial for this downstream task. 5.4. Downstream Task II: Material Classification We investigate the efficacy of our visual-tactile dataset for understanding material properties, focusing on the task of material classification. We follow the formulation by Yang et al. [56], which consists of three subtasks: (i) material classification, requiring the distinction of materials among 20 possible classes; (ii) softness classification, a binary problem dividing materials as either hard or soft; and (iii) hardness classification, which requires the classification of materials as either rough or smooth. We follow the same experimental procedure of [56]: we pretrain a contrastive model on a dataset and perform linear probing on the sub-tasks\u2019 training set. Our experiments only vary the pretraining dataset, leaving all architectural choices and hyperparameters the same. We compare against four baselines. A random classifier (chance); the ObjectFolder 2.0 dataset [17]; the VisGel dataset [35]; and the Touch and Go dataset [56]. Note that the touch sensor used in the test data (GelSight) differs from the one used in our dataset (DIGIT). Therefore, we use for pretraining a combination of our dataset and Touch and Go. To ensure a fair comparison, we also compare to the combination of each dataset and Touch and Go. The findings from this evaluation, as shown in Table 5, suggest that our data improves the effectiveness of the contrastive pretraining objective, even though our data is from a different distribution. Moreover, we find that adding estimated touch probes for pretraining results in a higher performance on all the three tasks, especially the smoothness classification. This indicates that not only does our dataset covers a wide range of materials but also our diffusion model captures the distinguishable and useful patterns of different materials. Dataset Material Hard/ Soft Rough/ Smooth Chance 18.6 66.1 56.3 ObjectFolder 2.0 [17] 36.2 72.0 69.0 VisGel [35] 39.1 69.4 70.4 Touch and Go [56] 54.7 77.3 79.4 + ObjectFolder 2.0 [17] 54.6 87.3 84.8 + VisGel [35] 53.1 86.7 83.6 + Ours\u2217(Real) 57.6 88.4 81.7 + Ours\u2217(Real + Estimated) 59.0 88.7 86.1 Table 5. Material classification. We show the downstream material recognition accuracy of models pre-trained on different datasets. The final rows show the performance when combining different datasets with Touch and Go [56]. \u2217The task-specific training and testing datasets for this task are collected with a GelSight sensor. We note that our data comes from a different distribution, since it is collected with a DIGIT sensor [32]. 6."
+                }
+            ],
+            "Fengyu Yang": [
+                {
+                    "url": "http://arxiv.org/abs/2401.18084v1",
+                    "title": "Binding Touch to Everything: Learning Unified Multimodal Tactile Representations",
+                    "abstract": "The ability to associate touch with other modalities has huge implications\nfor humans and computational systems. However, multimodal learning with touch\nremains challenging due to the expensive data collection process and\nnon-standardized sensor outputs. We introduce UniTouch, a unified tactile model\nfor vision-based touch sensors connected to multiple modalities, including\nvision, language, and sound. We achieve this by aligning our UniTouch\nembeddings to pretrained image embeddings already associated with a variety of\nother modalities. We further propose learnable sensor-specific tokens, allowing\nthe model to learn from a set of heterogeneous tactile sensors, all at the same\ntime. UniTouch is capable of conducting various touch sensing tasks in the\nzero-shot setting, from robot grasping prediction to touch image question\nanswering. To the best of our knowledge, UniTouch is the first to demonstrate\nsuch capabilities. Project page: https://cfeng16.github.io/UniTouch/",
+                    "authors": "Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, Alex Wong",
+                    "published": "2024-01-31",
+                    "updated": "2024-01-31",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.RO"
+                    ],
+                    "main_content": "Introduction Amongst our five main senses, touch sensing is perhaps the most crucial to human survival, due to its role in perceiving physical contact \u2014 rivaling even vision in its overall importance [46, 73, 79]. Our ability to form cross-modal associations between touch and our other senses [91] thus underlies a great deal of our physical capabilities. For example, we predict from vision how a surface will feel before we touch it, and we predict from touch how an object will sound before we strike it. These cross-modal associations are also a 1 arXiv:2401.18084v1  [cs.CV]  31 Jan 2024 \fkey component of computational systems, such as for robotic manipulation [6, 8, 65, 68, 75, 83\u201385, 107, 114, 116], material and geometry estimation [10, 38, 111, 119], assistive technology [42], and texture recognition [50, 78, 118]. Despite their importance, cross-modal associations between touch and other modalities have received considerably less attention from the multimodal research community than those of other modalities, such as vision, language, and sound. Touch is expensive to acquire [30, 32, 111] as it requires actively probing objects with touch sensors, limiting the scale of data collected for training tactile \u201cfoundation\u201d models. Moreover, touch sensors are not fully standardized, and thus there are large differences between outputs of different sensors [31, 121]. Even amongst the commonly used vision-based sensors, the difference in mechanical design and elastomeric material will lead to divergent artifacts, limiting generalization (Fig. 2). As a result, existing tactile representations are typically constrained to a single sensor. An emerging line of work has addressed the challenges of learning from other low-resource modalities, like sound, point clouds, and depth, by aligning examples with pretrained vision-language embeddings [35, 64, 109]. In this paper, we show that this approach can be adapted to tactile sensing. We align tactile signals to visual signals, thereby linking touch to a variety of other modalities, such as language and sound. Then we can use the representations within off-the-shelf models trained on other modalities (e.g. CLIP [87]), to solve a variety of tactile sensing tasks. To deal with the large variations in different touch sensors, we train a single model with multiple tactile signals at once, and introduce learnable tokens to model sensor-specific properties, such as the calibration and intensity profiles in the touch signal. Our trained model, which we call UniTouch, is a generalpurpose interface for multiple vision-based tactile sensors. Our model unifies many previously studied tactile sensing tasks \u201czero shot\u201d and greatly expands the range of tasks that touch sensing can be applied, as shown in Fig. 1: (i) We apply it to zero-shot touch understanding tasks like material recognition and robotic grasp stability prediction. (ii) We obtain strong performance in cross-modal retrieval with touch by aligning touch with other modalities in a shared latent space. (iii) The learned representation can also support image synthesis tasks, including touch-to-image generation [71, 112] and tactile-driven image stylization [111, 112], by using it within off-the-shelf text-to-image diffusion models. (iv) We combine touch with large language models (LLM), allowing us to perform tasks such as tactile question answering in a variety of tactile domains, including contact localization, grasping stability prediction, and etc. (v) Finally, we perform \u201cX-to-touch\u201d generation, producing touch images from vision, text, and audio. Our experiments suggest our zero-shot model achieves competitive (or even better) performance GelSight from [111] DIGIT from [94] Taxim from [32] GelSlim from [33] TACTO from [30] DIGIT from [57] Figure 2. Tactile images of different sensors and datasets. In contrast to many other modalities, signals from different touch sensing hardware exhibit large amounts of variation. than previously proposed approaches on multiple tasks. 2. Related Work Tactile sensing. Early tactile sensors were chiefly engineered to register fundamental, low-dimensional sensory outputs such as force, pressure, vibration, and temperature [19, 56, 61, 62]. Lately, there has been a growing focus on vision-based tactile sensors. GelSight [54, 117] as one of the representative sensors, features an elastomeric gel with an embedded camera and illumination system. The gel deforms upon contact with an object and creates a highresolution height map using photometric stereo [55], which provides detailed information about the shape and physical properties of touch [66, 97]. One variant, DIGIT [59], has a specially designed silicone-based elastomer gel with a harder surface and a different illumination system. Another variant GelSlim [97] contains a stretchy, loose-weave fabric gel surface. Recent work also turns into the simulation of tactile sensors [1, 17, 36, 53, 90, 101]. Taxim [90] simulates the optical response of a GelSight sensor and TACTO [101] calculates the local contact geometry and the corresponding rendering. We focus on these vision-based sensors as they are widely available in visuo-tactile datasets [30, 32, 33, 94, 100, 108, 111, 117, 118], are commonly used in various applications [9, 11, 12, 34, 41, 45, 51, 60, 67, 68, 95, 115, 127], and all adopt image as the output format. While these visionbased tactile sensors and simulators share similar imaging patterns, the difference in design and calibration results in a significant domain gap, as shown in Fig. 2. Hence, researchers typically study each sensor separately. In our work, we introduce a novel approach to understanding multiple sensors through our unified touch encoder. Representation learning with touch. The initial efforts learn tactile representations for specific tasks [29, 63, 72, 96, 118]. Lee et al. [63] undertook a collaborative training of Convolutional Neural Networks (CNN) for an RGB camera 2 \fand a force sensor to facilitate contact-rich manipulation tasks. Similarly, Yuan et al. [118] employed a comparable methodology to establish a shared latent space between visual and tactile modalities using the Gelsight touch sensor, aimed at precise fabric classification. Recently, researchers have learned general representations of touch through selfsupervision. Yang et al. [111] learned tactile representations for Gelsight sensors with visuo-tactile contrastive multiview coding [98] and Kerr et al. [57] proposed a contrastive pretraining method for the DIGIT sensor. Other works adopted BYOL framework [39] or contrastive predictive coding [120] to learn representations for non vision-based tactile sensors like BioTac. Some work [52] applies masked autoencoders to learn tactile representations directly from tactile inputs. Unlike methods concentrated solely on visuo-tactile learning for a single sensor, our approach aims to learn touch representations that can be applied across various sensors and interconnected with multiple modalities. Multimodal representation learning. The success of vision-language pretraining [23, 77, 86, 88, 106] has demonstrated the ability to bridge the gap between visual content, such as images or videos, and textual descriptions [48, 49, 69]. Furthermore, some researchers have extended the application of CLIP into the 3D domain [37, 122, 123, 128]. Some works learn shared audio-visual representation [2, 13, 25, 27, 44, 80, 82, 93, 105] by leveraging natural correspondence with videos. Some works also study shared audio-language representation [26, 40, 103]. Bender et al. [4] crafted an embedding space for the flavors of wines by leveraging both image and text annotations. Chen et al. [15] learned shared spatial information from binaural sound and vision. Some works learned the association between vision and metadata [14, 102, 126]. Imagebind [35] proposed to learn a joint embedding for six diverse modalities solely through image alignment and emerge zero-shot cross-modal capabilities. In our work, we extend this concept to the sense of touch and bind it to other modalities including text and audio by aligning tactile data with images, encouraging a more comprehensive understanding of cross-modal touch interactions without paired data. 3. Method We aim to learn a unified tactile representation for different touch sensors that captures relationships between touch and different modalities, e.g. vision, text, and audio. First, we present our contrastive visuo-tactile pretraining, inspired by [35], that can emerge interconnections of touch and other modalities. We then introduce our touch encoder design and data sampling strategy that can be used for different tactile sensors at once. Finally, we show how our learned representation can be applied to various downstream tasks. Image  Encoder Touch  Encoder Contrastive loss Binding space L Sensor token Image Touch Frozen Trainable < GelSight > Touch Image Figure 3. Method overview. We align our touch embedding with a pre-trained image embedding derived from large-scale vision language data, using sensor-specific tokens for multi-sensor training. 3.1. Binding touch with images We learn a multimodal tactile representation from touch and vision solely, without the need for paired text and audio data for touch. We achieve that by aligning our touch embedding to a pretrained image embedding using contrastive learning as shown in Fig. 3, where the image embedding is already aligned with modalities like language and audio training from large-scale image-paired datasets [35]. We denote \u2126v as the visual image domain and \u2126t as the tactile image domain. Thus, given B visual and touch pairs in a batch, {(vi, ti)}B i=1, where vi : \u2126v \u2282R2 \u2192R3 and ti : \u2126t \u2282R2 \u2192R3, we align a tactile embedding FT (ti) \u2208RC with the pretrained visual embedding FV (vi) \u2208RC from [35] by maximizing the cosine similarity between corresponding visuo-tactile pairs. We optimize this objective using InfoNCE loss [81] to match touches to correct images: LT \u2192V = \u22121 B B X i=1 log exp(FT (ti) \u00b7 FV (vi)/\u03c4) PB j=1 exp(FT (ti) \u00b7 FV (vj)/\u03c4) , (1) where \u03c4 is a temperature hyperparameter [104] and C is feature dimension. Analogously, we can also match from image vi to touch ti using the loss LV \u2192T . Thus, we minimize the overall loss: L = LT \u2192V + LV \u2192T . (2) Naturally, minimizing the contrastive objective [27, 98, 110, 126] will \u201cpull\u201d a visuo-tactile pair close together and \u201cpush\u201d it away from other pairs, achieving the alignment between touch and visual embedding. As the visual embedding comes from a learned joint space that has already aligned with different modalities, touch that is bound with images will bridge a connection to other modalities, yielding a multi-modal unified tactile representation. 3 \f3.2. Learning from multiple sensors at once We want to learn a generalizable tactile representation that will be suitable for different tactile sensors. Therefore, we designed our touch encoder FT to bridge the domain gap among various vision-based tactile sensors caused by the difference in sensor designs. Specifically, we introduce a set of learnable sensorspecific tokens {sk}K k=1, where sk \u2208RL\u00d7D, to capture specific details for each senor, e.g., calibration and background color in touch images, so that the remaining model capacity can be used to learn common knowledge across different type of touch sensors, such as texture and geometry. Here, K represents the number of sensors we train on, L is the number of sensor-specific tokens for each sensor, and D is the token dimension. For the given touch image ti, and its corresponding tactile sensor tokens sti, we append these sensor-specific tokens as prefixes to touch image patch tokens and then encode them with our touch encoder resulting in the final embedding FT (ti, sti) (Fig. 3). For our contrastive vision-touch pretraining, we optimize: LT \u2192V = \u22121 B B X i=1 log exp(FT (ti, sti) \u00b7 FV (vi)/\u03c4) PB j=1 exp(FT (ti, sti) \u00b7 FV (vj)/\u03c4) , (3) as well as LV \u2192T from the other direction. In-batch data sampling. We found that batch sampling strategy [18] plays an important role when we train with data, acquired by multiple touch sensors, using contrastive learning. The model will under-perform if we randomly sample from each data source [113] which results in a surplus of easy negatives due to the domain gap between different sensors. Therefore, we design a batch sampling strategy to guarantee that \u03c3 percent of training examples in a batch are sampled from the same datasets. Given that our dataset D is the union over N datasets collected with diverse tactile sensors D = S n\u2208{1,2,...,N} Dn, the probability of selecting a given dataset Dn to sample from is defined as: pn = \u2225Dn\u2225 PN m=1 \u2225Dm\u2225 , (4) where \u2225\u00b7 \u2225denotes cardinality. D\u03c3 denotes the selected dataset from which we perform uniform random sampling to yield \u03c3 \u00b7 B examples; the rest (1 \u2212\u03c3) \u00b7 B examples are uniformly sampled from other datasets, i.e., D \\ D\u03c3, where \u03c3 is a hyperparameter range from 0 to 1 representing the portion of the batch. This batch sampling strategy significantly benefits our training as it allows the model to mostly focus on intra-sensor hard negatives but still be exposed to different sensors to enhance inter-sensor discrimination. Inference. To generalize our learned representation to unseen types of sensors during the inference, we retrieve the Dataset Sensor # data Material cls. Robot grasp Train & Eval Touch and Go [111] GelSight 120k \u2713 The Feeling of Success [6] GelSight 9.3k \u2713 YCB-Slide [94] DIGIT 183k \u2713 Object Folder 2.0 [32] Taxim 180k \u2713 \u2713 Eval. Object Folder Real [33] GelSlim 20k \u2713 Object Folder 1.0 [30] TACTO 20k \u2713 \u2713 SSVTP [57] DIGIT 4.6k \u2713 Table 1. Datasets for training and evaluation. nearest neighbor sensor-specific tokens from the learned sensor set {sk}N k=1. Specifically, we first compute a prototype for each sensor, a 1D vector that averages all the raw pixels belonging to the tactile images collected by this sensor, and store these prototypes after training. Then, during the inference stage, we compute the L1 distance of between an input tactile image and all the sensor prototypes and retrieve the sensor with minimum distance. 3.3. Applications By aligning our touch embedding to the joint latent space, we establish a link between touch and other modalities. These alignments allow us to perform various zero-shot and crossmodal applications without any further training. Zero-shot touch understanding. Emergent alignment of touch and text enables zero-shot touch understanding, e.g., material classification and grasp stability prediction. Following CLIP [88], we encode the touch images and text prompts with templates and class names. We compute their similarity score and rank them to achieve the zero-shot classification. Touch-LLM. Using an existing vision-language model [28, 124] with the image embedding [35] that we align our touch embedding with, we can create our touch-language model by switching to our touch encoder. Given the touch image and language inputs, we can obtain a more comprehensive understanding via question-answering. Image synthesis with touch. Binding touch with text also opens up more potential abilities for image synthesis with touch. We leverage the pretrained text-to-image diffusion model [89] and use our touch features to condition the denoising process, achieving zero-shot touch-to-image generation [71, 112] and tactile-driven image stylization. X-to-touch generation. We also connect other modalities to touch using the diffusion model so that we can achieve xto-touch generation, where we imagine the touch by seeing, describing, or listening. We train an image-to-touch diffusion model [112] using the pretrained joint image embedding and then we can generate touch from text and audio as well. 4 \fMethod Pretrain Data In domain Datasets Out-of-domain Datasets Touch and Go ObjectFolder 2.0 YCB-Slide ObjectFolder 1.0 ObjectFolder Real SSVTP Chance \u2013 5.0 14.2 10.0 14.2 14.2 16.6 Linear Probing Supervised ImageNet 47.1 70.3 72.3 37.5 54.8 73.4 VT CMC [111] Single 56.5 74.3 75.2 \u2013 \u2013 \u2013 SSVTP [57] Single 47.6 69.8 74.8 \u2013 \u2013 \u2013 VT CMC [111] All 49.2 70.3 69.5 33.8 48.1 68.5 SSVTP [57] All 43.8 68.9 67.4 35.1 49.7 66.8 Ours All 61.3 85.4 78.1 41.3 61.2 77.4 Zero-Shot Ours All 52.7 43.5 66.4 32.7 33.2 60.9 Table 2. Tactile material classification. We compare our touch features with other methods and ImageNet pretraining. We also report our zero-shot classification performance. The metric is accuracy (%). Method Pretrain Data In domain Out-of-domain Feeling OF 2.0 OF 1.0 Chance 52.3 52.0 50.7 Linear Probing Supervised ImageNet 75.9 70.1 68.9 VT CMC [111] Single 80.1 74.8 SSVTP [57] Single 80.3 74.0 VT CMC [111] All 66.1 65.8 67.2 SSVTP [57] All 65.8 64.2 65.3 Ours All 82.3 78.1 75.8 Zero-Shot Ours All 65.5 64.3 64.7 Table 3. Robotics grasping stability prediction. We compare our touch features with other methods and ImageNet pretraining on grasping stability prediction task. We report our zero-shot results. The metric is accuracy (%). 4. Experiments We evaluate our model on extensive tasks spanning various application domains, including zero-shot touch understanding, cross-modal retrieval, zero-shot image synthesis with touch, Touch-LLM, and X-to-touch generation. Implementations. We base our model on ImageBind [35]. We use the AdamW optimizer [58, 76] with the base learning rate of 1 \u00d7 10\u22125 and cosine decay learning rate scheduler. We train our model with a batch size of 48 on each of the 4 NVIDIA A40 GPUs for 150 epochs. We set the temperature parameter \u03c4 = 0.07. We adopt Vision Transformer (ViT) [24] as the backbone for our touch encoder, which contains 24 multi-head attention blocks with 16 heads on each. The feature dimension C is 1024. We use L = 5 learnable tokens for each sensor type in our pretraining datasets with K = 3 different sensors. For the in-batch sampling, we set \u03c3 = 0.75, meaning that 75% of the data comes from the same dataset, with the remainder sourced from others. Datasets. We train and evaluate our model on four visuotactile datasets collected by three different vision-based tactile sensors (Tab. 1). These include the real-world dataset Touch and Go [111], the robotic dataset Feeling of Success [6], the YCB-Slide [94] dataset featuring DIGIT sensor interactions, and the multimodal dataset ObjectFolder 2.0 [32] which contains simulated visual, tactile, and audio data of daily objects using Taxim tactile simulators. We train our model solely on the naturally paired image and touch data via self-supervision. To test the generalization ability of our model, we also evaluate it with three out-of-domain datasets with two unseen sensors, including ObjectFolder Real [33], ObjectFolder 1.0 [30] and SSVTP [57]. We specifically select objects 101-1000 from ObjectFolder 2.0 to avoid overlap with ObjectFolder 1.0. Also, ObejctFolder Real contains objects distinct from those in ObjectFolder 1.0 and 2.0. Please see Appendix A.1 for more details. 4.1. UniTouch representation First, we evaluate the quality of our learned touch features for downstream tasks: material classification and grasping stability prediction via linear probing. We freeze the learned touch embeddings and train a linear classifier on the downstream tasks for specific datasets. Baselines. We compare our model with two recent visuotactile self-supervised methods for vision-based tactile sensors: VT CMC [111] and SSVTP [57]. We also adopt them to our multi-dataset setup. We use the same architectures to ensure a fair comparison. We also compare with the supervised ImageNet [22] features, which are commonly used to represent tactile images [6, 7, 119]. Following [6, 33, 111], we evaluate models\u2019 performance via accuracy metric for both downstream tasks. Material classification. We evaluate the touch material classification task on three in-domain datasets Touch and Go, ObjectFolder 2.0, and YCB-Slide, and three out-of-domain datasets ObjectFolder 1.0, ObjectFolder Real, and SSVTP. It is worth noting that ObjectFolder Real and ObjectFolder 1.0 contain sensors never seen during the training. Tab. 2 shows results on linear probing. UniTouch outperforms all the baselines by a large margin, implying that our tactile representations benefit from the alignment to a wellstructured embedding space trained on large-scale datasets. In addition, the consistent improvements across all datasets 5 \fTactile-driven Image Stylization Touch-to-Image Generation Vision-from-touch Touch Reference Ours Vision-from-touch Touch Reference Ours Source Figure 4. Zero-shot image synthesis with touch. (Left) We generate an image of a scene given a tactile signal. (Right) We perform tactile-driven image stylization to manipulate an image to match a given touch signal. We compare our method to the state-of-the-art supervised diffusion method [112] trained on Touch and Go. We denote \u201creference\u201d as visual images paired with the input touch in the dataset, which are not seen by the model but only shown for the demonstration purpose. See Appendix A.4 for more examples. Method Retrieved Modality Touch \u2192Vision Touch \u2192Audio Touch \u2192Text Chance 1.0 1.0 1.0 Fully supervised CCA\u2020 8.50 6.18 PLSCA\u2020 6.25 7.11 DSCMR\u2020 4.92 6.15 DAR\u2020 8.80 7.77 CCA 17.8 15.7 16.8 PLSCA 16.8 15.9 18.2 DSCMR 26.5 19.6 22.7 DAR 32.3 27.8 31.9 Zero-shot Ours 41.9 37.9 38.0 Table 4. Cross-modal retrieval from touch. We evaluate the performance using mean Average Precision (mAP) on ObjectFolder 2.0. \u2020 denotes results from [33]. and sensors validate our proposed sensor-specific tokens and in-batch sampling strategy during training \u2013 resulting in insignificant generalization gains across different sensors. Grasping stability prediction. We follow the setting of [6, 33] to predict, from tactile input, whether a robotic gripper can successfully grasp and stably hold an object before it is lifted. Failures occur when the grasped object slips by more than 3cm. We evaluate UniTouch on three datasets: Feeling of Success, ObjectFolder 2.0, and ObjectFolder 1.0, where ObjectFolder 1.0 is an out-of-domain dataset. The linear probing results are shown in Tab. 3. Our performance consistently outperforms existing baselines by a large margin. Thus, we further demonstrate that our model design and training paradigm are useful not only in computer vision but also can be generalized to robotics tasks. 4.2. Zero-shot touch understanding We further evaluate UniTouch with zero-shot classification tasks, enabled by the emergent alignment with text during pretraining. We perform material classification and grasping prediction tasks by computing the cosine similarity between the embeddings of touch and corresponding text prompts. Class predictions are chosen based on highest scores, without training on labeled data. To the best of our knowledge, there are no other baselines that can perform zero-shot touch understanding in our manner. Material classification. We conduct zero-shot material classification by prompting the model with \u201cThis feels like [CLS]\u201d, where [CLS] is the name of the material. We show our zero-shot performance in the last row of Tab. 2. Our zeroshot method shows a comparable performance against several supervised methods, which not only indicates a strong tactile representation that is well-aligned with the text but also shows that off-the-shelf models trained for other modalities can be used to successfully solve touch sensing tasks. Grasping stability prediction. Similarly, we perform the zero-shot grasping stability prediction task by using text prompts like \u201cthe object is lifted in the air\u201d and \u201c\u201dthe object is falling on the ground\u201d. Tab. 3 shows that we are comparable to some of the supervised methods, demonstrating the capabilities of aligning touch and text can be extended to robotics tasks, which may be out of the training scope of the vision language model like CLIP with appropriate prompting. This may come from the fact that we link the touch of the successful grasps to the robot\u2019s action of lifting objects while failed grasps as those falling. We found consistent performance in both in and out-of-distribution datasets, demonstrating the generalization capability of this link. 6 \fNo, the object cannot be  grasped into the air as the  gripper is touching the  object at the edge. Can this object be grasped  into the air? Explain why by  describing where specifically  is the gripper touching on  the object. Input touch Reference The object in the touch  image is a green, grassy  blade. It is a thin, flat piece  of grass that is likely part of  a larger grassy field or lawn. Can you describe material  of the object presented in  touch image? Input touch Reference The touch is located  on the corner of  the table. What part of the table this  touch located? Input touch Reference Figure 5. Touch-LLM. Our Touch-LLM can conduct a series of tactile question-answer tasks such as robot grasping stability prediction, contact localization, and touch image captioning. We also show \u201creference\u201d visual images paired with the input touch, for better demonstration. See Appendix A.4 for more examples. Method Evaluation Metrics CVTP (\u2191) Material (\u2191) FID (\u2193) Supervised Pix2Pix [47] 0.09 0.15 136.4 VisGel [71] 0.10 0.20 128.3 Vision-from-touch [112] 0.35 0.25 81.2 Zero-shot Ours 0.56 0.31 103.11 Table 5. Zero-shot touch-to-image generation on Touch and Go. 4.3. Cross-modal retrieval with touch We conduct cross-modal retrieval to evaluate the alignment of our touch embeddings to those of other modalities. Given a touch image, we aim to identify the corresponding vision, text, and audio describing the same point of contact. Experimental setup. We evaluate on ObjectFolder 2.0 cross-sensory retrieval benchmark [33]. Following [33], we treat points from the same object as positive samples and evaluate using mAP. To evaluate touch-to-text retrieval, we annotated text descriptions that depict the contact point of the object from its visual input, serving as paired ground-truth text. We obtain the retrieval result by ranking the cosine similarity between an input touch and other modalities. Given that our method is not trained with paired audio or text data, we consider its performance in these two modalities as a demonstration of zero-shot learning. Baselines. We compare our method with several established baselines, including Canonical Correlation Analysis (CCA) [43], Partial Least Squares (PLSCA) [21], Deep Aligned Representations (DAR) [3], and Deep Supervised Cross-Modal Retrieval (DSCMR) [125]. Results. UniTouch achieves state-of-the-art performance on all three modalities and outperforms those supervised methods that are trained with paired modalities by a large Method LLM Eval GPT-4 Rating (\u2191) BLIP-2 [70] Vicuna [16] 1.01 InstructBLIP [20] Vicuna [16] 1.93 LLaVA-1.5 [74] Vicuna [16] 2.33 Touch-LLM (ours) LLaMA [99] 3.30 Table 6. Touch image caption evaluation. We evaluate our TouchLLM and three baselines on our test cases from Touch and Go [111]. Each model\u2019s response is rated by GPT-4 on a scale from 1 to 5. margin (Tab. 4). This demonstrates our strong cross-modal ability to align touch with other modalities without the need for explicit paired training data or additional supervision. 4.4. Image synthesis with touch In this part, we demonstrate that we can combine our touch embedding with an off-the-shelf image synthesis model easily to perform the image synthesis tasks conditioning touch images in a zero-shot manner. We perform two tasks: touchto-image generation [71, 112] and tactile-driven image stylization [111, 112]. Following [111, 112], we use three evaluation metrics: Frechet Inception Distance (FID), Contrastive Visuo-Tactile Pre-Training (CVTP), and material classification consistency. See Appendix A.3 for details. Touch-to-image generation. We aim to generate images solely from touch. We use a pretrained text-to-image diffusion model [89], conditioning on our touch features, and guiding the denoising process. Compared to the state-of-theart visuo-tactile diffusion-based model [112], our method generates more realistic objects that have not been previously seen in the dataset (see Fig. 4 (left)). While the images generated by [112] not only include the sensor and the arm holding it but also closely resemble the visual images in the training 7 \fPrompt Datasets Touch and Go OF 2 This is an image of [CLS] 40.7 34.3 This is a touch image of [CLS] 43.8 36.8 This looks like [CLS] 49.3 41.7 This feels like [CLS] 52.7 43.5 Image of [CLS] 48.8 40.3 Touch of [CLS] 51.2 40.9 Table 7. Prompt analysis for touch. We evaluate our prompt designs for zero-shot material classification on Touch and Go and ObjectFolder 2.0 datasets. set. Tab. 5 shows quantitative results, where we compare with Vision-from-touch [112], VisGel [71] and Pix2Pix [47] on Touch and Go [111]. Despite a slightly lower FID score compared to [112], our method outperforms on the CVTP and material consistency metrics. This suggests that while our generated images are out of the distribution of Touch and Go, our approach effectively bridges vision and touch. Tactile-driven image stylization. We also manipulate an image to align with a given touch signal [111, 112] zero shot. We achieve this by mixing the input image embedding with our conditioned touch embedding and feeding it into the pretrained diffusion model. We show qualitative results in Fig. 4 (right), where the input image is out of the distribution of Touch and Go [111]. We observe the supervised state-of-the-art method [112] fails to change the visual style according to the touch images even though these are seen during the training stage. See Appendix A.4 for more details. 4.5. Touch-LLM Interpreting vision-based touch images, crucial for delicate tasks in fields like robotics, is challenging due to human perceptual limitations. To address this, we integrate UniTouch embedding into a large language model (LLM), leveraging its robust understanding and reasoning capabilities for touch image interpretation, and name it as Touch-LLM. TouchLLM is capable of a series of tactile tasks such as grasping stability prediction, touch image interpretation, tactile contact localization and etc., most of which are non-trivial to humans, demonstrating the usefulness of combining touch with LLMs. We show some example tasks in Fig. 5. Quantitatively, we compare our model with three opensource vision-language models (VLMs): BLIP-2 [70], InstructBLIP [20], and LLaVA-1.5 [74] in the touch image captioning task by feeding them the same touch images and text prompts. We manually create captions for 400 randomly sampled RGB images from Touch and Go [111] as the ground truth. Following [5], we use GPT\u20134 to perform automatic evaluation by instructing GPT-4 to rate each model\u2019s generations on a scale of 1 to 5 given the reference response. As shown in Tab. 6, our Touch-LLM outperforms other VLMs by a large margin, indicating that our TouchLLM has much better understanding capabilities for touch images even with a less powerful LLM than Vicuna [16] which used by other models. See Appendix A.3 for more details. 4.6. X-to-touch generation We conduct X-to-touch generation to synthesize realistic tactile images corresponding to the input modality of vision, language, and audio. Fig. 1 shows plausible and consistent tactile images generated from both the visual input and its text captioning. Quantitatively, we evaluate our model on Touch and Go [111], where we measure material classification consistency between touch images generated from vision and its corresponding language captions. Our model achieves 55.3% consistency, illustrating the reliability of the generated results. See Appendices A.3 and A.4 for more examples and details. 4.7. Ablation study Learning from multiple sensors. Tab. 8 ablates the importance of each module design on the zero-shot material classification task with the Touch and Go dataset. The baseline, a vanilla transformer model aligning touch embedding to a fixed vision encoder, drops performance significantly when applied to multiple sensors and datasets, i.e., from 43.1% to 21.4%, indicating the difficulty of the sensor domain gap. We improve the performance by 17% by adding the sensor-specific tokens to it. Similarly, we found a 19% by adding our sampling strategy. With our proposed batch sampling strategy and sensor-specific tokens, our model can achieve strong performance, surpassing the model trained on a single dataset, which emphasizes the significance of our proposed methods for learning a better touch representation from multiple sensors. We argue that this is because sensor-specific embeddings help distinguish hard samples from different sensors while sampling strategy helps identify hard negatives within the same sensor in the training. Combining these, we can tackle inter-sensor and intra-sensor hard samples thus obtaining the performance boost. Language prompting for touch. We explore how language prompting can help with understanding touch, the first endeavor in this domain. Given that vision captures more global and semantic information, and touch focuses on material properties, texture, and microgeometry, directly adopting prompts from vision-language works may not yield satisfactory results. We design touch-specific prompt templates by adopting the common prompts from vision-language works and replacing with words related to haptics, i.e., changing \u201cimage\u201d to \u201ctouch image\u201d and \u201clook like\u201d to \u201cfeel like\u201d (see Tab. 7). We evaluate them using the zero-shot material classification task on Touch and Go and ObjectFolder 2.0. We empirically found that our prompts 8 \fMethod Pretrain Data Eval Touch and Go Chance \u2013 16.7 Baseline Touch and Go 43.1 Baseline All 21.4 Baseline + sensor token All 38.1 Baseline + sample All 40.3 Baseline + sensor token + sample All 52.7 Table 8. Ablation study. We ablate the effectiveness of each of our proposed contributions via the zero-shot material classification. can significantly improve the performance, indicating that language can indeed understand touch. We suspect this phenomenon may be due to the design of visuo-tactile datasets, which feature human or robotic touch actions, thus enabling the model to associate tactile images with these actions. 5. Discussion We introduced UniTouch, a unified multimodal tactile representation for vision-based tactile sensors. To achieve this, we align our touch embedding to a shared multimodal embedding space using contrastive learning. We further introduce sensor-specific tokens that enables learning from different sensors all at once. UniTouch unifies many existing tactile sensing tasks and significantly expands the range of tasks for touch sensing. Nonetheless, the field of multimodal (foundational) model is admittedly still young. Agents, like ourselves, leverage complementary strengths of multi-sensory observations, incorporating all five senses in everyday tasks. With that goal in mind, we see our work as a concrete step towards that direction, opening new avenues for multimodal touch experience beyond vision and touch and integrating tactile sensing into multimodal foundation models. Limitations. As the full range of tactile sensors exhibit differing output formats (e.g. image, barometric signals, force), we limit our scope to vision-based tactile sensors. Scaling up our training strategy is key to further integrate emerging tactile sensors in the future. In addition, like other multimodal foundational models, our representation is \u201cblackbox\u201d, which does not easily for interpretability in the space, where one may benefit from explainability. Acknowledgements. We thank Jiacheng Zhang, Shaokai Wu and Chenyang Ma for the helpful discussions and feedbacks on our manuscript. This work is supported by NSF 2112562 Athena AI Institute and Sony Research."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2309.15117v1",
+                    "title": "Generating Visual Scenes from Touch",
+                    "abstract": "An emerging line of work has sought to generate plausible imagery from touch.\nExisting approaches, however, tackle only narrow aspects of the visuo-tactile\nsynthesis problem, and lag significantly behind the quality of cross-modal\nsynthesis methods in other domains. We draw on recent advances in latent\ndiffusion to create a model for synthesizing images from tactile signals (and\nvice versa) and apply it to a number of visuo-tactile synthesis tasks. Using\nthis model, we significantly outperform prior work on the tactile-driven\nstylization problem, i.e., manipulating an image to match a touch signal, and\nwe are the first to successfully generate images from touch without additional\nsources of information about the scene. We also successfully use our model to\naddress two novel synthesis problems: generating images that do not contain the\ntouch sensor or the hand holding it, and estimating an image's shading from its\nreflectance and touch.",
+                    "authors": "Fengyu Yang, Jiacheng Zhang, Andrew Owens",
+                    "published": "2023-09-26",
+                    "updated": "2023-09-26",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction Humans rely crucially on cross-modal associations between sight and touch to physically interact with the world [58]. For example, our sense of sight tells us how the ground in front of us will feel when we place our feet on it, while our sense of touch conveys the likely visual appearance of an unseen object from a brief contact. Translating between these modalities requires an understanding of physical and material properties. Models trained to solve this problem must learn, for instance, to associate rapid changes in shading with rough microgeometry, and smooth textures with soft surfaces. Touch is arguably the most important sensory modality for humans [48, 43, 40], due to its role in basic survival [40, 9, 23] and physical interaction. Yet touch sensing has received comparably little attention in multimodal learning. An emerging line of work has addressed the problem of translating touch to sight, such as by learning joint embeddings [64, 39], manipulating visual styles to match a tactile signal [64], or adding a plausible imagery of a robotic arm to an existing photo of a scene [38]. While these tasks each capture important parts of the cross-modal prediction problem, each currently requires a separate, specialpurpose method. Existing methods also lag significantly behind those of other areas of multimodal perception, which provide general-purpose methods for cross-modal synthe1 arXiv:2309.15117v1  [cs.CV]  26 Sep 2023 \fsis, and can translate between modalities without the aid of extra conditional information. In this paper, we generate plausible images of natural scenes from touch (and vice versa), drawing on recent advances in diffusion models [51, 12, 21, 22, 45]. We adapt latent diffusion models to a variety of visuo-tactile synthesis problems. Our proposed framework obtains strong results on several novel synthesis problems, and unifies many previously studied visuo-tactile synthesis tasks. First, we study the problem of generating images from touch (and vice versa). We address the task of generating images from touch without any image-based conditioning, where we are the first method to successfully generate images for natural scenes (Fig. 1a). We also address the task of adding an arm to a photo of an existing scene, where we significantly outperform prior work [38]. Second, we address the recently proposed tactile-driven image stylization task, i.e., the problem of manipulating an image to match a given touch signal [64] (Fig. 1b), using an approach based on guided image synthesis [44]. Our approach obtains results that are higher fidelity and that match the tactile signal significantly more closely than those of prior work. It also provides the ability to control the amount of image content preserved from the input image. Finally, we show that we can augment our model with additional conditional information. Taking inspiration from the classic problem of intrinsic image decomposition [41, 3], we perform tactile-driven shading estimation, predicting an image after conditioning on reflectance and touch (Fig. 1c). Since changes in tactile microgeometry often manifest as changes in shading (i.e., the information missing from reflectance), this tests the model\u2019s ability to link the two signals. We also use segmentation masks to create \u201chand-less\u201d images that contain the object being pressed but not the tactile sensor or arm that pressed it. We demonstrate our framework\u2019s effectiveness using natural scenes from the Touch and Go dataset [64], a collection of egocentric videos that capture a wide variety of materials and objects using GelSight [28], and using robotcollected data from VisGel [38]. 2. Related Work Cross-modal synthesis with diffusion models. Diffusion models have recently become a favored generative model family due to their ability to produce high-quality samples. However, one major concern for diffusion models is their slow inference speed due to the iterative generation process on high dimensional data. Recently, latent diffusion [51] addressed this drawback by working on a compressed latent space of lower dimensionality, which allows diffusion models to work on more extensive tasks with accelerating the speed. These models have demonstrated remarkable success in tasks such as image synthesis [12, 21, 22, 45], super-resolution [54], and image editing [57, 44, 8]. Additionally, the advancements in multimodal learning [25, 27, 16] have enabled diffusion models to be utilized for cross-modal synthesis tasks. For visionlanguage generation, diffusion models have been studied for text-to-image synthesis [1, 29, 46, 50, 53], text-to-speech generation [7, 31, 35], text-to-3D generation [42, 55]. In addition, diffusion models also show promising results in audio synthesis including text-to-audio generation [56], waveform generation [32, 18, 6]. In this work, we are the first to employ diffusion model on real-world visual-tactile data, exploring the possibility of utilizing tactile data as a prompt for image synthesis. In concurrent work, Higuera et al. [20] used diffusion to simulate tactile data, which they used to train a braille classifier. Tactile sensing. Early touch sensors recorded simple, low-dimensional sensory signals, such as measures of force, vibration, and temperature [33, 34, 10]. Beginning with GelSight [65, 28], researchers proposed a variety of visionbased tactile sensors, which convert the deformation of an illuminated membrane using a camera, thereby providing detailed information about shape and material properties [59, 36]. We focus on these sensors, particularly using GelSight, since it is widely used applications [38, 4], and available in visuo-tactile datasets [15, 17, 64]. Crucially, these sensors produce images as output, allowing us to use the same network architectures for both images and touch [66]. Other work proposes collocated vision and touch sensors [62, 5]. Cross-modal models for vision and touch. Li et al. [38] used a GAN [24] to translate between tactile signals and images, using a dataset acquired by a robot. In contrast, they require conditioning their touch-to-image model on another photo from the same scene. This is a task that amounts to adding an arm grasping the correct object (given several possible choices), rather than generating an object that could have plausibly led to a touch signal according to its physical properties. It is not straightforward to adapt their method to the other touch-to-image synthesis problems we address without major modifications. Yang et al. [64] proposed a visuo-tactile dataset and used a GAN to restyle images to match a touch signal. Their approach only learns a limited number of visual styles, and cannot be straightforwardly adopt extra conditional information (such as reflectance) or be applied to unconditional cross-modal translation tasks. Other work has learned multimodal visuotactile embeddings [64, 39]. Other work learns to associate touch and sight for servoing and manipulation [5]. 3. Method Our goal is to translate touch to vision (and vision to touch) using a generative model. We will do this using a model based on latent diffusion [51]. We will use this 2 \fDiffusion  Process CVTP Condition Attn. Attn. Input Output \u2130 \ud835\udc9f Hand  Mask Encoder Decoder Touch Image \ud835\udc38\u2205! \ud835\udc38\u2205\" \ud835\udc67\" # \ud835\udc67\" #$% \u0303 \ud835\udc67\" # \ud835\udcae Segmentation Intrinsic Image Decomposition \u2112 \u2131 & Downsample Downsample Reflectance Figure 2: Touch-to-image model. We use a latent diffusion model to generate an image of a scene from touch. The touch signal is represented using multiple frames of video from a GelSight sensor. The model uses a segmentation mask to optionally generate only the scene content containing the pressed object (i.e., without a hand or touch sensor). We also optionally condition on reflectance from a scene, in which case the model\u2019s generation task requires it to estimate shading. model to solve a number of tasks, including: 1) cross-modal visual-tactile synthesis, 2) tactile-driven image stylization, and 3) tactile-driven shading estimation. 3.1. Cross-Modal Synthesis of Vision and Touch We now describe our framework for cross-modal synthesis. First, we describe a contrastive visuo-tactile model, which we use to perform conditional generation. Second, we describe our cross-modal latent diffusion model. 3.1.1 Contrastive Visuo-tactile Pretraining (CVTP) Following other work in cross-modal synthesis [49, 51], we provide conditional information to our generation models through multimodal embeddings via contrastive learning [14, 67, 63, 60]. Our embedding-learning approach resembles that of Yang et al. [64] and contrastive multiview coding [60]. A key difference is that we incorporate temporal information into our visual and tactile representations. Touching an object is a dynamic process, and the information we obtain varies over time, from the moment when the tactile sensor begins touching the object, to the point when the sensor has reached it maximum deformation. Adding temporal cues provides information about material properties that may be hard to perceive from a single sample, such as the hardness or softness of a surface [66, 26]. Given the visual and tactile datasets XI and XT , which consist of N synchronized visual-tactile frames {xi I, xi T }N i=1, we denote the video clip sampled at time i with the window size w = 2C + 1, vi I = {xi\u2212C I , ..., xi I, ..., xi+C I } and the corresponding tactile clip vt I = {xi\u2212C I , ..., xi I, ..., xi+C I }. We denote examples taken from the same visual-tactile recording {vi I, vi T } as positives, and samples from different visual-tactile video pair {\u03c5i I, \u03c5j T } as negatives. Our goal is to jointly learn temporal visual zI = E\u03d5I(vI) and tactile zT = E\u03d5T (vT ) encoder. We use a 2D ResNet as the architecture for both encoders. For easy comparison to static models, we incorporate temporal information into the model via early fusion (concatenating channel-wise). Then we maximize the probability of finding the corresponding visuo-tactile video pair in a memory bank containing K samples using InfoNCE [47] loss: LVI,VT i = \u2212log exp(E\u03d5I(vi I) \u00b7 E\u03d5T (vi T )/\u03c4) PK j=1 exp(E\u03d5I(vi I) \u00b7 E\u03d5T (vj T )/\u03c4) (1) where \u03c4 is a small constant. Analogously, we get a symmetric objective LVT ,VI and minimize: LCVTP = LVI,VT + LVT ,VI. (2) 3.1.2 Touch-conditioned Image Generation We now describe the tactile-to-image generation model (an image-to-touch model can be formulated in an analogous way). Our approach follows Rombach et al. [51], which translates language to images, but with a variety of extensions specific to the visuo-tactile synthesis problem. Given a visuo-tactile image pair {xI, xT } \u2208RH\u00d7W \u00d73, our goal is to generate an image e xI from tactile input xT . We encode the input x into a latent representation z = E(x) \u2208 Rh\u00d7w\u00d73. A decoder D will reconstruct the image \u02c6 x = D(z) 3 \fVision Touch Vision Touch Touch and Go [64] VisGel [38] Figure 3: Visuo-tactile datasets. For our experiments, we evaluate our model on natural scenes from Touch and Go [64] and robot-collected data from VisGel [38]. from the code. The latent dimension h \u00d7 w is smaller than the image dimension H \u00d7 W. Training. We train a touch-to-vision diffusion generation in the latent space zI = E(xI). Diffusion models learn to generate images by recursively denoising from a normal distribution to the desired data distribution. Specifically, given our latent representation zI, we uniformly sample a diffusion step t \u2208{1, ..., T} and obtain the corresponding noisy image zt I by iteratively adding Gaussian noise with a variance schedule. We use a U-Net [52] network \u03f5\u03b8 as our denoising model, which is conditioned on the tactile representation encoded through the tactile encoder E\u03d5T trained in Section 3.1.1. We minimize: L(\u03b8, \u03d5) = EzI,c,\u03f5,t \u0002 \u2225\u03f5t \u2212\u03f5\u03b8(zt I, t, E\u03d5T (vT ))\u22252 2 \u0003 , (3) where \u03f5t is the added noise at time t, and vT is the tactile example. The denoising network \u03f5\u03b8 and the tactile encoder E\u03d5T are jointly trained. Inference. At test time, we first sample noise e zT I \u223c N(0, 1) at time T, and then use the trained diffusion model to iteratively predict the noise e \u03f5t, resulting in a denoised latent representation e zt I = e zt+1 I \u2212e \u03f5t+1 from t \u2208{T \u22121, ..., 0}. Following [51, 12], we use classifier-free guidance to trade off between sample quality and diversity in the conditional generation, computing the noise as: e \u03f5t = \u03f5\u03b8(e zt I, t, \u2205) + s \u00b7 \u0000\u03f5\u03b8(e zt I, t, E\u03d5T (vT )) \u2212\u03f5\u03b8(e zt I, t, \u2205) \u0001 , (4) where \u2205denotes a zero-filled conditional example (for unconditional generation), and s is the guidance scale. Finally, we convert the latent representation e z0 I to an image e xI = D(e z0 I) \u2208RH\u00d7W \u00d73. 3.2. Visuo-Tactile Synthesis Models So far, we have presented models for translating between touch and images (and vice versa). We now describe several visuo-tactile synthesis models that we build on this diffusion framework. 3.2.1 Generating realistic images without hands One of the challenges of dealing with visuo-tactile data is that the tactile sensor typically occludes the object that is being touched (Fig. 3). Generated images will therefore contain the sensor, and potentially the arm that held it. This is not always desirable, as a major goal of touch sensing is to generate images of objects or materials that could have plausibly led to a given touch signal. We address this problem for the natural scenes from the Touch and Go dataset [64], which contain visible human hands and GelSight sensors [65]. To generate images containing only objects that yield a given tactile signal (without hands or touch sensors), we only compute the loss for pixels that do not overlap with hands during the training, thereby depriving the model of supervision for hand pixels. We first generate hand segmentation masks for the visual image mI = S(xI) and obtain the downsampled mask zm of the same spatial dimension of the image latent representation. For this, we use the off-the-shelf hand segmentation model from Darkhalil et al. [11], which is a modified model from PointRend [30] instance segmentation designed specifically for segmenting hands. We then mask the diffusion loss (Eq. 6) to be: Ezm,zI,c,\u03f5,t \u0002 \u2225zm \u2299 \u0000\u03f5t \u2212\u03f5\u03b8(zt I, t, E\u03d5T (vT )) \u0001 \u22252 2 \u0003 , (5) where zm indicates whether a pixel overlaps with a hand, and \u2299denotes pointwise multiplication. 3.2.2 Tactile-driven Image Stylization Tactile-driven image stylization [64] aims to manipulate the visual appearance of an object so that it looks more consistent with a given touch signal. Previous work posed the problem of editing the visual style of an image while preserving its structure [64, 37]. Given an input image xI and a desired tactile signal x\u2032 T (obtained from a different scene), our goal is to manipulate xI so that it appears to \u201cfeels\u201d more like x\u2032 T . We adapt the approach of Meng et al. [44]. We first compute the noisy latent representation zN I at time 0 \u2264N \u2264T, where T denotes the total number of denoising steps. We then conduct the denoising process for zN I from time step N to 0 conditioned on x\u2032 T . This allows for fine-grained control over the amount of content preserved from the input image, via the parameter N. We analyze the choice of N at Sec. 4.6. 3.2.3 Tactile-driven Shading Estimation Touch conveys a great deal of information about a surface\u2019s microgeometry [28]. Much of this information can also be perceived through shading cues: intensity variations due to light interacting with surface orientation for objects with Lambertian material properties. Following classic work in intrinsic image decomposition [2, 19, 3], we assume that the image can be factorized into reflectance and shading for each pixel, i.e., we can write our image xI = xR \u2299xS where the two terms in the product are the per-pixel reflectance and shading. 4 \fOurs Input Yang et al. Ours Input Yang et al. Touch Reference Reference Touch Input Output Input Output Input Output Input Output Reference Touch Figure 4: Tactile-driven Image Stylization. (Top) We restyle the input image using the given touch signal (reference image from scene provided for clarity). We compare our approach to Yang et al. [64]. Our approach generates images with higher quality matching more closely to the given tactile signal. (Bottom) We show more examples of the manipulated images. Please see supplement for more examples. We propose a model that deals with inferring shading from touch. Given an image\u2019s estimated reflectance map xR, along with a touch signal xT , we reconstruct the original image xI. This is a task that requires inferring the shading, since it is the component that is missing from the input. By formulating the problem so that we predict the original image, we can easily reuse the latent encoder/decoder from natural images. We address this task by modifying our network so that it also takes reflectance as input (Eq. 6). We first estimate reflectance using the intrinsic image decomposition model of Liu et al. [41] and downsample it to the same dimensions as the latent space. We then concatenate the downsampled reflectance zR to the noisy representation zt I as the input for each denoising step. Thus we modify the loss function (Eq. 6) as the following: L(\u03b8, \u03d5) = EzI,c,\u03f5,t \u0002 \u2225\u03f5t \u2212\u03f5\u03b8(zt I \u2297zR, t, E\u03d5T (vT ))\u22252 2 \u0003 , (6) where \u2297denotes concatenation. 4. Results We evaluate our cross-modal synthesis models through qualitative and quantitative experiments on natural scenes and robot-collected data. 4.1. Implementation details Contrastive visuo-tactile model. Following [64], we use ResNet-18 as the backbone of contrastive model, and train on Touch and Go [64]. This model is trained using SGD for 240 epochs with the learning rate of 0.1 and weight decay of 10\u22124. The ResNet takes 5 reference frames as input using early fusion (concatenated channel-wise) and we take the feature embedding from the last layer of the feature and map it to 512 dimensions. Following prior work [60], we use \u03c4 = 0.07 and use a memory bank with 16,385 examples. Visuo-tactile diffusion model. We base our latent diffusion model on Stable Diffusion [51]. We use the Adam optimizer with the base learning rate of 2 \u00d7 10\u22126. Models are all trained with 30 iterations using the above learning rate policy. We train our model with the batch size of 96 on 4 RTX A40 GPUs. The conditional model is finetuned along with the diffusion model. We use the frozen, pretrained VQGAN [13] to obtain our latent representation, with the spatial dimension of 64\u00d764. During the inference, we conduct denoising process for 200 steps and set the guidance scale s = 7.5. 4.2. Experimental Setup Dataset. We conduct our experiments on two real-world visuo-tactile datasets: \u2022 Touch and Go dataset. The Touch and Go dataset is a recent, real-world visuo-tactile dataset in which humans probe a variety of objects in both indoor and outdoor scenes. There are 13,900 touches from roughly 4000 different object instances and 20 material categories. Since this is the only available dataset with zoomed-in images and clearly visible materials, we use it for all three tasks. \u2022 VisGel dataset. The VisGel dataset contains synchronized videos of a robot arm equipped with a GelSight sensor interacting with 195 household objects. The dataset includes 195 objects from a wide range of indoor 5 \fReal Image VisGel Ours Condition VisGel Ours Condition Real Image Condition Real Image Ours w/o hands Ours with hands Figure 5: Visuo-tactile Cross Generation on Touch and Go dataset. (Top) We compare our approach to state-of-the-art method Visgel [38]. (Bottom) We show more results of our generated images with and without hands. In both case our approach is able to generate realsitic images with high fidelity. Table 1: Evalutation of cross-modal generation on Touch and Go. Method Touch \u2192Image Image \u2192Touch CVTP (\u2191)Material(\u2191)FID(\u2193)SSIM(\u2191)PSNR(\u2191) Pix2Pix [24] 0.08 0.15 136.4 0.43 14.3 VisGel [38] 0.07 0.15 128.3 0.45 15.0 Ours w/ hands 0.12 0.22 48.7 0.50 15.4 Ours w/o hands 0.12 0.24 81.5 0.50 15.4 scenes of food items, tools, kitchen items, to fabrics and stationery. In total, the dataset contains 12k touches and around 3M frames. Evaluation metrics. We use several quantitative metrics to evaluate the quality of our generated images or tactile signals. We use Frechet Inception Distance (FID), which compares the distribution of real and generated image activations using trained network. Following Yang et al. [64] and CLIP [49], we take the cosine similarity between our learned visual and tactile embeddings for the generated images and conditioned tactile signals, a metric we call Contrastive Visuo-Tactile Pre-Training (CVTP). A higher Table 2: Evaluation of cross-modal generation on VisGel (and conditioning on another photo from the scene). Method Touch \u2192Image Image \u2192Touch SSIM(\u2191) PSNR(\u2191) SSIM (\u2191) PSNR (\u2191) Pix2Pix [24] 0.50 15.1 0.71 20.7 VisGel [38] 0.59 17.9 0.76 26.2 Ours 0.76 21.5 0.85 27.6 score indicates a better correlation between touch and images. It is worth noting that the CVTP metric only takes one frame of touch input. Following [64], we measure Material Classification Consistency: we use the material classifier from Yang et al. [64] to categorize the predicted and ground truth images, and measure the rate at which they agree. Finally, following [16], we evaluate standard Structural Similarity Index Measure (SSIM) and Peak Signal to Noise Ratio (PSNR) [61] metrics. 4.3. Cross-modal Generation We perform cross-modal generation, i.e., generating an image from touch and vice versa, on both in-the-wild Touch 6 \fVisGel Ours Input Ground  Truth Reference VisGel Ours Input Ground  Truth Reference Ground truth pix2pix pix2pix w/ temporal Ours w/o temporal Ours w/o rebalancing Ours pix2pix pix2pix w/ temporal Ours w/o temporal Ours Supervised prediction Ground truth Touch Reference Output Touch Output Reference Touch Output Touch Output Ground truth pix2pix w/ temporal Ours w/o temporal Ours w/o rebalancing Ours pix2pix w/ temporal Ours w/o temporal Ours Supervised prediction Ground truth Figure 6: Visuo-tactile Cross Generation on VisGel dataset. (Top) We compare our approach to state-of-the-art method VisGel [38]. (Bottom) Our approach is able to generate robotic hands touching reasonable locations of objects given the same reference image but different tactile signals. and Go dataset and robot-collected dataset VisGel. For straightforward comparison to prior work [38], on VisGel we provide a reference photo of the scene as an input to the model. Thus, successfully predicting the ground truth image amounts to inserting imagery of the robotic arm to the correct location in the scene. For Touch and Go, we do not condition the model on a visual input: instead, we simply translate one modality to the other. For evaluation metrics, we use CVTP, material classification consistency, and FID score for touch-to-image generation and SSIM and PSNR for image-to-touch generation. For VisGel dataset we leverage SSIM and PSNR as the evaluation metric for both tasks. We only use CVTP, material classification consistency and FID only on touch-to-image generation task on Touch and Go, since these evaluation metrics rely on a pretrained neural network from datasets of natural images, which may not generalize well on a different modality or to robot-collected data. We compare our model to the prior state-of-the-art visuotactile generation method [38], which is adapted from pix2pix [24] and is specifically designed to bridge the large domain gap between modalities by adding a reference image and temporal condition. As it is not possible to find a reference image in the natural image dataset, we remove the reference image while keeping everything else the same. We show quantitative results for both tasks on Touch and Go and VisGel in Table 1 and Table 2 respectively. Our methods outperform existing state-of-the-art methods by a large margin for all evaluation metrics. We note that the Table 3: Quantitative results of of tactile-driven image stylization. Method Evaluation Metrics CVTP (\u2191) Material (\u2191) FID (\u2193) Cycle GAN [68] 0.09 0.15 24.6 Yang et al. [64] 0.10 0.20 22.5 Ours 0.13 0.22 15.8 variation of our model that removes hands from images obtains a worse FID score compared to those with hands, due to the discrepancy of hands between the original dataset and our generated images. Interestingly, the presence of hands does not does not affect the performance of CVTP and material classification consistency. We provide qualitative results from both models in Figure 5 (bottom). 4.4. Tactile-Driven Image Stylization Following [64], we evaluate the performance of tactiledriven image stylization on Touch and Go [64] using CVTP and material classification metrics. We also calculate the FID score between the set of generated images and the set of real images associated with the given tactile signals, which measures the fidelity of the output. We compare our model to a modified version of CycleGAN [68] and the state-ofthe-art method of Yang et al. [64]. From the quantitative comparisons in Table 3, our method demonstrates a significant improvement over existing methods. We also show qualitative comparisons in Figure 3, where the generated images more closely match the tactile signal, and we are 7 \fRef. Only Ground Truth Touch Ref.+Touch Reflectance (Input) Figure 7: Tactile-driven shading estimation. We compare our approach to a model without a tactile signal (only reflectance), finding that the tactile-driven model better captures subtle material properties, such as roughness. Table 4: Quantitative results for tactile-driven shading estimation. Method Reflectance \u2192Image SSIM(\u2191) PSNR(\u2191) FID(\u2193) Touch Only 0.27 11.6 48.7 Reflectance Only 0.46 14.5 40.7 Reflectance + Touch 0.48 15.4 36.9 able to generate styles that existing methods fail to capture. 4.5. Tactile-driven Shading Estimation We hypothesize that the tactile signal conveys information about the microgeometry of an image, and thus allows a model to produce more accurate images than a reflectanceto-image model that does not have access to touch. We evaluated both models on Touch and Go (Table 4) and found that adding touch indeed improves performance on all evaluation metrics. We also show qualitative comparisons in Figure 7. We found that tactile signals are especially informative for predicting roughness and smoothness of Lambertian surfaces, such as bricks. 4.6. Analysis Importance of temporal information. We first study the effect of adding multiple GelSight frames to the contrastive visuo-tactile embedding (Figure 9). We compare our method with the unconditional generation and material class conditional generation on Touch and Go. We found that conditioned generation provides a large improvement in performance compared to the unconditional generation. We also observed that the generation conditioned on the pretrained model is significantly better than that without pretraining. Interestingly, the model conditioned on the material class outperforms the variation of the model that only observes a single GelSight frame, suggesting that perceiving a touch signal from only a single moment in time may Tactile Condition Input N = T/4 N = T/2 N = 3T/4 Image Reference Figure 8: Controlling the amount of preserved image content. Manipulated images of tactile-driven image stylization using different values of N. 75 80 85 90 95 100 105 FID 101.2 97.5 85.7 88.9 87.1 84.2 81.5 FID score of different conditions Unconditional Channel-wise concat. Material label CVTP w/o pretrain CVTP 1 frame CVTP 3 frames CVTP 5 frames Figure 9: Effect of different types of tactile conditioning. be less informative than the material category. Providing the model with additional frames significantly improves the model, with the 5-frame model obtaining the overall best performance. Controllable Image Stylization Our method allows us to control over the amount of image content that is preserved from the original image by changing the denoising staring point N (Sec. 3.2.2) [44]. From Figure 8, we observe that if we select the larger N, the generated image will be changed more drastically where the visual appearance will be changed completely to match the tactile signal while ruining the original image structure. In extreme case, where N = T the manipulated result will be equal to the touch-to-image generation result, while small N will result in little overall change. We empirically found that selecting N = T/2 obtains a good trade-off between these factors. 5."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2212.03435v1",
+                    "title": "Improve Bilingual TTS Using Dynamic Language and Phonology Embedding",
+                    "abstract": "In most cases, bilingual TTS needs to handle three types of input scripts:\nfirst language only, second language only, and second language embedded in the\nfirst language. In the latter two situations, the pronunciation and intonation\nof the second language are usually quite different due to the influence of the\nfirst language. Therefore, it is a big challenge to accurately model the\npronunciation and intonation of the second language in different contexts\nwithout mutual interference. This paper builds a Mandarin-English TTS system to\nacquire more standard spoken English speech from a monolingual Chinese speaker.\nWe introduce phonology embedding to capture the English differences between\ndifferent phonology. Embedding mask is applied to language embedding for\ndistinguishing information between different languages and to phonology\nembedding for focusing on English expression. We specially design an embedding\nstrength modulator to capture the dynamic strength of language and phonology.\nExperiments show that our approach can produce significantly more natural and\nstandard spoken English speech of the monolingual Chinese speaker. From\nanalysis, we find that suitable phonology control contributes to better\nperformance in different scenarios.",
+                    "authors": "Fengyu Yang, Jian Luan, Yujun Wang",
+                    "published": "2022-12-07",
+                    "updated": "2022-12-07",
+                    "primary_cat": "cs.SD",
+                    "cats": [
+                        "cs.SD",
+                        "cs.CL",
+                        "eess.AS"
+                    ],
+                    "main_content": "INTRODUCTION Nowadays, a bilingual text-to-speech(TTS) system is necessary for many application scenarios like voice assistant. For example, the names of English songs and movies are often directly embedded in Chinese responses. A straightforward way to build a bilingual TTS system is by collecting speech data from bilingual speakers. [1] proposed a shared hidden Markov model (HMM)-based bilingual TTS system, using a Mandarin-English corpus recorded by a bilingual speaker. [2] presented a TTS system using a speaker and language factorized deep neural network(DNN) with a corpus of three bilingual speakers. However, mixed-lingual corpora are scarce while a large number of monolingual corpora are easily accessible. Another way is to leverage monolingual speech data from different speakers [3, 4, 5, 6, 7, 8]. [5] proposes a polyglot synthesis method adapting the shared HMM states to the target speaker, trained on monolingual corpora. [6] proposes to factorize speaker and language based on an HMM-based parametric TTS system. [7] utilizes a combined phonetic space in two languages to build a codeswitched TTS system based on HMM. [8] maps the senones between two monolingual corpora in two languages with a speakerindependent DNN ASR output based on HMM TTS. End-to-end TTS systems also extend to multilingual tasks using monolingual speech[9, 10, 11, 12, 13, 14, 15, 16]. [13] used Unicode bytes as a uni\ufb01ed new language representation for multilingual TTS. 125 hours of speech were used and their system can read codeswitching text, despite the problem of speaker inconsistency when cross-language. [14] trained with designed loss terms preserving the speaker\u2019s identity in multiple languages based on the VoiceLoop architecture [17]. The trained speech is recorded by 410 monolingual speakers speech from English, Spanish and German. [15] used an adversarial loss term to disentangle speaker identity from the speech content, which trained with 550 hours of speech from 92 monolingual speakers. Limited by corpus size, [16] proposed tone embedding and tone classi\ufb01er for tone preservation to generate utterances in a proper prosodic accent of the target language. Generally, each speaker speaks only one language, leading to speaker and language characteristics being highly correlated. Using only monolingual corpora for bilingual or multilingual TTS easily leads to heavy accent carry-over in synthesized speech or inconsistent voice between languages. Actually, bilingual corpus helps deal with the problem. [18] trained a TTS system transforming speaker embedding between languages from a bilingual speaker to other monolingual speakers for a high degree of naturalness. In this paper, we expect to utilize scarce bilingual corpora to acquire more standard spoken English from a monolingual speaker, which is highly correlated with phonology learning. For example, in mixed-lingual utterances, the pronunciation of English by a non-native speaker, like Chinese, is strongly in\ufb02uenced by their native language and is most often different from the standard English pronunciation[19]. Mandarin derives pronunciation directly from the spellings of the word with different tones, which have a high grapheme-to-phoneme(g2p) correlation. In contrast, English is an alphabetic and highly non-phonemic language. In Consequence, native phonemic language speakers, whose pronunciation is in\ufb02uenced by the spelling of the word, often pronounce English words differently from standard English speakers[20]. In mixed-lingual utterances, these speakers, despite quali\ufb01ed bilingual speakers, generally replace some English phonemes with the closest phoneme in their native language, resulting in mispronunciation and differences in phonology like articulation change and intonation variation[21]. Given these challenges, building a state-of-the-art bilingual TTS system requires special designs handling the English differences in phonology between mixed-lingual and monolingual utterances. In this paper, our contributions include: (1) introducing phonology embedding to capture the English differences between mixed-lingual and monolingual utterances; (2) proposing embedding mask to language embedding for distinguishing information between different languages and phonology embedding for focusing on expression between different phonology of English; (3) designing embedding strength modulator(ESM) to capture the dynamic information of language and phonology, which helps to generate more standard spoken English speech; (4) experiments showing that static and dynamic components in ESM can control different attributes of phonology. Phonology decomposition and control can make a contribution to more standard spoken English expression and better performance in different scenarios. arXiv:2212.03435v1  [cs.SD]  7 Dec 2022 \fText sequence Decoder Mel Spectrogram GMM Attention Text Encoder Gradient  Reversal Speaker  Classifier Language  Embedding Speaker  Embedding Concat Phonology Embedding Embedding Strength  Modulator Embedding Strength  Modulator Add Add Adversarial Loss Fig. 1. Overview of the proposed bilingual architecture with specially designed modules marked in yellow color. 2. MODEL STRUCTURE Fig. 1 illustrates the proposed bilingual TTS architecture. The encoder-attention-decoder backbone with speaker and language embedding will be described in Sec. 2.1. The phonology embedding and specially designed masks for language and phonology embedding respectively will be described in Sec. 2.2. The embedding strength modulator will be described in Sec. 2.3. 2.1. Baseline Our baseline system adopted from [22] is a popular Tacotron2[22]based multilingual TTS architecture. It uses attention to bridge encoder and decoder. Language and speaker information are embedded in separate look-up tables. They are combined with the encoder output to distinguish different languages and speakers. Besides, an adversarially-trained speaker classi\ufb01er is employed to disentangle text encoder output from speaker information. Mel-Lpcnet adopted from [23] is used as a vocoder to reconstruct waveform from given mel-spectrogram. The architecture takes phoneme sequences as inputs for both English and Mandarin. Their phoneme sets are simply concatenated and no phoneme is shared across. Tone or stress tokens are inserted into the phoneme sequence at the end of each syllable. For Mandarin, there are 4 lexicon tones and one neutral tone. Instead, there are 4 stress types for English includes the sentence, primary, secondary, and none. Moreover, prosodic break tokens are inserted into the input sequence as well. Finally, the expanded phoneme set contains: 73 Mandarin phonemes, 39 English phonemes, 5 Mandarin tones, 4 English stresses, Mandarin character boundary, English syllable boundary, English liaison symbol and 4 shared prosodic break types, i.e. prosodic word (PW), prosodic phrase (PPH), intonation phrase (IPH) and silence at the beginning or end. 2.2. Embedding mask Fig. 2 shows an example of embedding mask in language and phonology embedding. Instead of broadcasting language embedding to all the tokens of the input sequence, the proposed method applies language embedding only to the token types shared across languages, i.e. PW, PPH, IPH and /sil/. Because other token types are language-speci\ufb01c already and need no additional information to distinguish language. On the other hand, to capture the English differences between the mixed-lingual and monolingual utterances, a special phonology embedding is designed. To focus on English expression, it is applied sil DH AH 6 #1 K AO 7 L #2 z an 4 sil Language embedding Phonology embedding The call\u8d5e\u3002 Front-end Encoder Fig. 2. An illustration of how to mask embedding. Language and phonology embedding only applied to the highlighted position of encoder outputs. The symbols #1, #2, #3 and /sil/ denote 4 shared prosodic break types. The numbers 1-5 denote tones of the previous Chinese syllable. The numbers 6-9 denote stresses of the previous English syllable. to all English-speci\ufb01c tokens, including 4 types of stresses, syllable boundary and liaison symbol. 2.3. Embedding strength modulator Even though the language and phonology embedding have been limited to only part of input tokens by masks, we think their strength should vary for different contexts. To capture the dynamic strength of languages and phonology, we propose an attention-based embedding strength modulator, whose framework is similar to [24]. The structure of the ESM is shown in Fig. 3. There are two subnetworks in ESM: multi-head attention and a feed-forward network. The layer normalization and residual connection are applied to both of the sub-networks. Formally, from the encoder output with scaled positional encoding Eo, and the language or phonology embedding LP, the \ufb01rst sub-network Mo and the second sub-network Fo are calculated as: Mo = MH(Eo, LN(LP), LN(LP)) + LP, (1) Fo = FFN(LN(Mo)) + Mo. (2) where MH(query, key, value), FFN(\u00b7) and LN(\u00b7) are multi-head attention, feed-forward network and layer normalization respectively. Since the attention key and value (LP) have only one item, the energy need not be normalized by softmax operation. Instead, each head in multi-head attention is computed by: headh = \u03b1h\u00b7Vh = Qh\u00b7Kh \u2225Qh\u2225\u2225Kh\u2225\u00b7Vh, (3) where \u2225\u00b7\u2225is the L2 norm of the last dimension, {Q, K, V } represent query, key and value through linear transformation respectively and the strength \u03b1 is a scaled cosine similarity between the query and key to be in the range of [-1, 1]. In particular, there are two components in Fig. 3 marked in yellow color. The original embedding learned for each language and phonology is regarded as a static component. While the output of multi-head attention, the static embedding multiplied by a dynamic weight, is regarded as a dynamic component. We will analyze the roles each component of language and phonology embedding play in Sec. 3.3. 3. EXPERIMENTS 3.1. Basic setups Models are trained with proprietary datasets composing three kinds of high-quality speech: (1) bilingual corpus from two Chinese \fLanguage or  Phonology  Embedding MultiHead  Attention LN LN FeedForward  Module Add Add Static Component Dynamic Component Encoder Outputs Scaled  Positional  Encoding Fig. 3. The structure of the embedding strength modulator of both language and phonology embedding. speakers, 45000 and 25000 Mandarin utterances for female and male speakers respectively, 9000 mixed-lingual utterances and 9000 English utterances for both speakers; (2) English corpus from two American speakers, 9000 and 25000 English utterances for female and male speakers respectively; (3) Mandarin corpus from a female Chinese speaker, 9000 Mandarin utterances for cross-lingual experiments. Labels of the above corpora in language and phonology embedding are described in Tab. 1. Mandarin utterances are all from Chinese speakers. Their language is labeled Mandarin and no English phonology label is required. For plain English utterances, the corpora recorded by both American and Chinese speakers are labeled as English for language and Standard-English for phonology. These Chinese speakers pronounce English well and the corpora recording by American speakers are used as supplementary English datasets and are bene\ufb01cial to speaker learning. Particularly, since English parts in mixed-lingual utterances are in a small amount and are mostly words and abbreviations with heavy Chinese phonology, we label them Mandarin for language and Chinese-English for phonology, treated as Mandarin utterances. The additional inputs of the learned speaker (64-dim), language and phonology embedding (both 512-dim same with the dimensions of encoder output) are injected into the backbone. In ESM, the \ufb01rst sub-network includes 8-head multi-head attention and the feed-forward sub-network consists of two convolution networks with 2048 and 512 hidden units. Linguistic inputs have been introduced in Sec. 2.2 and for acoustic features, we use an 80-band mel-spectrogram extracted from 16kHz waveforms. We built the following systems for comparison: \u2022 BASE: Baseline system with senteneial language embedding as described in Sec. 2.1; \u2022 EM: Baseline system with specially designed language and phonology embedding as described in Sec. 2.2; \u2022 ESM: Baseline system with specially designed language and phonology embedding through ESM as described in Sec. 2.3. 3.2. Subjective evaluation We conduct Mean Opinion Score (MOS) evaluations of speech naturalness and speaker similarity via subjective listening tests. 20 speakers are asked to listen to the generated 20 English utterances and 10 mixed-lingual utterances. MOS results are reported in Tab. 2. Except for parts of samples in listing tests, generated Mandarin demos of this monolingual speaker are also shown in demo pages1. We can \ufb01nd that the EM system with masked embedding brings better performance on both speech naturalness and speaker similarity than the conventional BASE system. It indicates that masked embedding captures features that better represent language and phonol1Samples can be found from: https://fyyang1996.github.io/esm/ ogy. For the further proposed embedding strength modulator, we \ufb01nd that by capturing the dynamic strength of language and phonology system ESM achieves signi\ufb01cantly better performance than the EM system. It demonstrates that the dynamic strength of language and phonology is bene\ufb01cial to speech naturalness and speaker similarity of generated speech. 3.3. ESM component analysis As mentioned above, the output of ESM may be regarded as the combination of a static component and a dynamic component. One simple method of analyzing the contribution of each component is to condition the model on only one component at each time. In the generation phase, we replace the static or dynamic component from Mandarin label to English label for language embedding or from Chinese-English phonology label to Standard-English phonology label for phonology embedding respectively. Fig. 4 shows the spectrogram and F0 contour, extracted by parselmouth[25], of the same sentence synthesized with six kinds of label combinations as described below: (a) Base combination: using Mandarin and Chinese-English phonology labels both in dynamic and static components; (b) Reference combination: using English and Standard-English phonology labels both in dynamic and static components; (c) Based on (a), replacing dynamic phonology embedding from Chinese-English to Standard-English phonology. (d) Based on (a), replacing static phonology embedding from Chinese-English to Standard-English phonology; (e) Based on (a), replacing dynamic language embedding from Mandarin to English; (f) Based on (a), replacing static language embedding from Mandarin to English; Empirically, we \ufb01nd that each component represents articulation, intonation, speaking rate and pause duration changes respectively, which in\ufb02uence phonology collectively. Listing to the samples of (a) and (c) in the demo page, we can easily hear about articulation changes between them, which is dif\ufb01cult to be caught sight of. Perceptually, the trend of F0 values in Fig. 4(d) is different from that in Fig. 4(a), showing that static phonology embedding major affects intonation. Fig. 4(e) shows that replacing the dynamic language embedding from Mandarin to English causes a gradual compression of the spectrogram and F0 values in the time domain. We believe that the dynamic language embedding encodes the information correlated with speaking rate variation. Besides, syllables in Fig. 4(f) have distinct intervals compared with that in Fig. 4(a), which demonstrates that static language embedding represents the average duration of pauses. More demos and be found in the demo page. \fTable 1. Labels of trained corpora in language and phonology embedding. Corpus Language embedding Phonology embedding (1) Chinese speaker Train Mandarin Mandarin None Mixed-lingual Mandarin Chinese-English English English Standard-English (2) American speaker Train English English Standard-English (3) Chinese speaker Test Mandarin Mandarin None Table 2. The MOS of different systems with con\ufb01dence intervals of 95%. Model BASE EM ESM Naturalness 3.81\u00b10.12 4.03\u00b10.10 4.39\u00b10.08 Similarity 3.79\u00b10.12 3.91\u00b10.11 4.04\u00b10.10 0 50 100 150 200 250 300 Frame 0 10 20 30 40 50 60 70 80 Mel channel 150 175 200 225 250 275 300 325 350 Fundamental frequency [Hz] (a) Base combination 0 50 100 150 200 250 300 Frame 0 10 20 30 40 50 60 70 80 Mel channel 150 175 200 225 250 275 300 325 350 Fundamental frequency [Hz] (b) Reference combination 0 50 100 150 200 250 300 Frame 0 10 20 30 40 50 60 70 80 Mel channel 150 175 200 225 250 275 300 325 350 Fundamental frequency [Hz] (c) Dynamic phonology embedding 0 50 100 150 200 250 300 Frame 0 10 20 30 40 50 60 70 80 Mel channel 150 175 200 225 250 275 300 325 350 Fundamental frequency [Hz] (d) Static phonology embedding 0 50 100 150 200 250 300 Frame 0 10 20 30 40 50 60 70 80 Mel channel 150 175 200 225 250 275 300 325 350 Fundamental frequency [Hz] (e) Dynamic language embedding 0 50 100 150 200 250 300 Frame 0 10 20 30 40 50 60 70 80 Mel channel 150 175 200 225 250 275 300 325 350 Fundamental frequency [Hz] (f) Static language embedding Fig. 4. Spectrogram and F0 of a test sentence generated by different combinations, which refers to 2.1 in demo page. 3.4. Control To validate the above analysis, we conduct MOS evaluations of speech naturalness and speaker similarity via subjective listening tests. 20 speakers are asked to listen to the generated 15 English utterances for enhancing English expressiveness and 15 mixed-lingual utterances for smooth mixed-lingual transition. Demos can be found in 3 and 4 on the demo pages. Enhance expressiveness To enhance the expressiveness of a plain English text, we double the dynamic components of both language and phonology embedding while remaining their static components. The \u201ddouble\u201d herein means that the \ufb01nal vector has a double distance of the reference vector to the base vector. For language embedding, the reference is English and the base is Mandarin. While for phonology embedding, the reference is Standard English and the base is Chinese-English. Fig. 5 shows the results of MOS evalua27% 36% 26% 4% 47% 60% Similarity Naturalness Proposed No Preference Controlled Fig. 5. A/B preference results for control in enhancing expressiveness or not with con\ufb01dence intervals of 95% and p-value<0.0001 from the t-test. 17% 22% 30% 27% 53% 51% Similarity Naturalness Proposed No Preference Controlled Fig. 6. A/B preference results for control in smooth transition or not with con\ufb01dence intervals of 95% and p-value<0.0001 from the t-test. tions. We \ufb01nd that by the \u201ddouble\u201d operation herein system ESM achieves signi\ufb01cantly better performance than the ESM system on speech naturalness. It indicates that the control operation enhances English expressiveness signi\ufb01cantly. Smooth transition When synthesizing a mixed-lingual text, we modify the language labels of embedded English words from Mandarin to English while phonology labels of them from ChineseEnglish to Standard-English. Particularly, their static component of phonology embedding remains Chinese-English. In this way, the English words will have standard-English articulation but more compatible intonation with the context of Chinese words. Fig. 6 shows the results of MOS evaluations. It can be found that the controlled ESM system brings better performance on both speech naturalness and speaker similarity than the proposed ESM system. It demonstrates that the control operation is bene\ufb01cial to smooth mixed-lingual transition. 4."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2211.12498v2",
+                    "title": "Touch and Go: Learning from Human-Collected Vision and Touch",
+                    "abstract": "The ability to associate touch with sight is essential for tasks that require\nphysically interacting with objects in the world. We propose a dataset with\npaired visual and tactile data called Touch and Go, in which human data\ncollectors probe objects in natural environments using tactile sensors, while\nsimultaneously recording egocentric video. In contrast to previous efforts,\nwhich have largely been confined to lab settings or simulated environments, our\ndataset spans a large number of \"in the wild\" objects and scenes. To\ndemonstrate our dataset's effectiveness, we successfully apply it to a variety\nof tasks: 1) self-supervised visuo-tactile feature learning, 2) tactile-driven\nimage stylization, i.e., making the visual appearance of an object more\nconsistent with a given tactile signal, and 3) predicting future frames of a\ntactile signal from visuo-tactile inputs.",
+                    "authors": "Fengyu Yang, Chenyang Ma, Jiacheng Zhang, Jing Zhu, Wenzhen Yuan, Andrew Owens",
+                    "published": "2022-11-22",
+                    "updated": "2022-11-29",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction As humans, our ability to correlate touch with sight is an essential component of understanding the physical properties of the objects around us. While recent advances in other areas of multimodal learning have been fueled by large datasets, the dif\ufb01culty of collecting high-quality data has made it challenging for the community to develop similarly effective visuo-tactile models. An intuitively appealing solution is to of\ufb02oad data collection to robots [8, 9, 42], which can acquire enormous amounts of data by repeatedly probing objects around them. However, this approach captures only a narrow, \u201crobot-centric\u201d slice of the visuo-tactile world. The data typically is limited to a speci\ufb01c environment (e.g., a robotics lab), and it fundamentally suffers from a chicken-and-egg problem, as the robot must already be capable of touching and manipulating the objects it acquires data from. In practice, this often amounts to recording data from tabletops, typically with small objects the robots can safely grasp. Recent work has also turned to simulation [21, 22], such as by modeling special cases where tactile interactions can be accurately simulated (e.g., rigid objects). Yet this approach, too, is highly limited. Real objects squish, deform, and bend in complex ways, and their seemingly simple surfaces can hide complicated microgeometry, such as weaves of fabric and tiny pores. Obtaining a full understanding of vision and touch, beyond simple robotic manipulation tasks, requires modeling these subtle visuo-tactile properties. We argue that many aspects of the visuo-tactile world are currently best learned by observing physical interactions performed by humans. Humans can easily access a wide range of spaces and objects that would be very challenging for robots. By capturing data from objects in situ, the recorded sensory signals more closely match how the objects would be encountered in the wild. Inspired by this idea, we present a dataset, called Touch and Go, in which human data collectors walk through a variety of environments, probing objects with tactile sensors and simultaneously recording their actions on video. Our dataset spans a wide range of indoor and outdoor environments, such as classrooms, gyms, streets, and hiking trails. The objects and \u201cstuff\u201d [1] they contain are thus signi\ufb01cantly more diverse than those of existing datasets, making it well-suited to self-supervised learning, and to tasks that require an understanding of material properties, such as visual synthesis tasks. We apply our dataset to a variety of multimodal learning tasks. First, we learn tactile features through self-supervised learning, by training a model to associate images with touch. We \ufb01nd that the learned features signi\ufb01cantly outperform supervised ImageNet [15] features on a robotic manipulation task, and on recognizing materials in our dataset (Fig. 1b). Second, we propose a novel task of tactile-driven image stylization: making an image \u201cfeel more like\u201d a given tactile input. To solve this problem, we adapt the recent method of Li et al. [41] to generate an image whose structure matches an input image but whose style is likely to co-occur with the given tactile information. This task evaluates the ability to learn cross-modal associations \u2013 i.e., how an object feels from how it looks and vice versa. The resulting model can successfully change the texture of an input image, such as by adding bumps to a smooth surface to match the tactile information recorded from a rock (Fig. 1c). Finally, we study multimodal models for future touch prediction: predicting future frames of a touch sensor\u2019s recording, given both visual and tactile signals. We show that visual information improves these predictions over touch alone (Fig. 1d). 2 Related Work Simulated vision and touch. A variety of methods have simulated low-dimensional Biotac [59] and fabric-based tactile sensing [46]. Other work has proposed to simulate high-dimensional tactile data, based on visual tactile sensors such as GelSight [73, 32, 31]. Wang et al. [65] simulated visual and tactile data for robotic grasps of rigid objects. Gao et al. [21, 22] proposed a dataset of simulated visual, tactile, and audio data, derived from CAD models with only rigid deformation. In contrast to these works, we collect our data from real objects and scenes, which contain non-rigid deformation, microgeometry, and wider variations in visual appearance. Robotic vision and touch. Researchers have proposed a variety of methods that use visual and touch signals for robotic applications [9, 8, 13, 34, 37, 48, 43, 12, 70, 11, 18]. Several of these have proposed visuo-tactile datasets. Calandra et al. created a dataset for multimodal grasping [9] and regrasping [8] with a robotic arm. Li et al. [42] collected data from a robotic arm synthesis, and proposed a model based on generative adversarial networks [24] for cross-modal translation. Murali 2 \fet al. [48] proposed a dataset for tactile-only grasping. These datasets have largely been con\ufb01ned to speci\ufb01c environments (e.g., a lab space containing the robots), and only contain objects provided to them by humans that the robots are capable of interacting with. Consequently, they contain a small number of object instances (each less than 200). Human-collected multimodal data. We take inspiration from work that collects data by having humans physically interact with objects in situ. Song et al. [60] proposed a human-collected grasping dataset. In contrast, our focus is on having humans collect rich multimodal sensory data that is well-suited to self-supervised learning. Our approach is similar to Owens et al. [51], which learns audio-visual associations by probing objects with a drumstick [51]. In contrast, we collect touch instead of sound, and record data in an approximately egocentric manner as they move from object to object. Later work predicts the trajectory of a bounced ball [55]. Sundaram et al. [61] proposed a glove that records tactile signals, and collected a dataset of human grasps for 26 objects in a lab setting. Other work predicts hand pose from touch [75]. Burka et al. combined several haptic sensors [5, 6]. They then demonstrated the resulting sensor by collecting a preliminary (currently unreleased) dataset of 357 real-world surfaces, and training a model to predict human ratings of 4 surface properties from touch. By contrast, we have signi\ufb01cantly larger and more diverse data from indoor and outdoor scenes (rather than \ufb02at, largely indoor surfaces), use a rich vision-based tactile sensor (GelSight), and demonstrate our dataset on cross-modal prediction tasks. Multimodal feature learning. In seminal work, de Sa [14] proposed to learn from correlating sight from sound. A variety of methods of been proposed for training deep networks to learn features from audio-visual correlations [49, 51, 52, 2, 35, 50, 47], from images and depth [63], and from vision and language [33, 45, 17, 56], and matching images and touch [38]. We adapt the contrastive model of Tian et al. [63] to visuo-tactile learning. Multimodal image prediction. A variety of methods have been proposed for predicting images from another modality, such as by using text or labels [58, 30, 4, 57] or sound [39, 10, 41]. Li et al. [41] proposed a model for audio-driven stylization, i.e. learning to restyle images to better match an audio signal. We adapt this model to tactile-driven stylization, creating a model that is conditioned on tactile inputs instead of sound. We also take inspirations from work on future video prediction [16, 23, 54, 66, 20, 71, 67, 3, 64]. In particular, Tian et al. [62] trains an action-conditioned video prediction method to estimate future tactile signals, using a GelSight sensor controlled by a CNC machine. In contrast, we predict future tactile signals from natural objects, and show that visual information can improve the prediction quality. Cross-modal image stylization. Many areas of multimodal perception have used cross-modal image stylization to evaluate whether models can capture associations between modalities (e.g. Textto-image stylization, Audio-visual stylization). We adapt the cross-modal stylization method of Li et al. [41] to tactile-driven stylization, a task that requires learning visual-tactile associations \u2013 i.e., how an object feels from how it looks and vice versa. This direction is also related to work in computer graphics that synthesizes images that have speci\ufb01c material properties [40, 25, 26, 77]. In contrast to these works, we synthesize images with material properties that are captured implicitly from a touch signal. 3 The Touch and Go Dataset We collect a dataset of natural vision-and-touch signals. Our dataset contains multimodal data recorded by humans, who probe objects in their natural locations with a tactile sensor. To more easily train and analyze models on this dataset, we also collect material labels and identify the frames within the press. 3.1 Collecting a natural visuo-tactile dataset To acquire our dataset, human data collectors (the authors) walked through a variety of environments, probing the objects with a tactile sensor. To obtain images that show clear, zoomed-in images of the objects being touched, two people collected data at once: one who presses the tactile sensor onto an object, and another who records an \u201capproximately egocentric\u201d video (see supplement for a visualization) of their hand. The two data collectors moved from object to object in the space as part of a single, continuously recording video, touching the objects around them. We show examples from our dataset in Fig. 2. The captured data varies heavily in material properties (e.g., soft/hard, smooth/rough), geometries, and semantics. 3 \fLeather Paper Synthetic Fabric  Natural Fabric Materials Brick Grass Plastic Rock Plants Metal Tree Sand Wood Concrete Soil Tile Figure 2: The Touch and Go Dataset. Human data collectors record paired visual and tactile information by probing objects in a variety of indoor and outdoor spaces. We show a selection of images, paired with the corresponding frame recorded by the GelSight tactile sensor. We show 16 representative categories (out of 20), and provide the distribution of material and scene types.2 Capturing procedure. To ensure that our dataset captures the natural variation of real-world vision and touch, we collect both rigid and deformable objects in indoor and outdoor scenes. These scenes include rooms in university buildings, such as classrooms and hallways, apartments, hiking trails, playgrounds, and streets. We show example footage from our model in Fig. 1, and in the supplement. The data collectors selected a variety of objects in each scene to press including chairs, walls, ground, sofa, table, etc. in indoor scenes, and grass, rock, tree, sand etc. in outdoor scenes, pressing each one approximately 3 times. Each press lasted for 0.7 sec on average. If an object contains multiple materials (e.g., a chair with a cushion and a plastic arm), collectors generally directed their presses to each one. The data collectors also aimed to touch object parts with complex geometry, rather than \ufb02at surfaces. To avoid capturing human faces, in public spaces the captures point the camera toward the ground when moving between objects. Since the GelSight may provide information about force implicitly [73], and explicit force readings are not required for many visuo-tactile tasks, we do not use a separate force sensor. Hardware. For the tactile sensor, we use GelSight [32], the variant designed for robotic manipulation [73]. This is a vision-based tactile sensor, approximately 1.5cm in diameter, in which a camera observes the deformation of a curved elastomer gel illuminated by multiple colored light sources, with markers embedded inside it. When the sensor is pressed against an object, the gel deforms, which results in changes to the re\ufb02ected illumination. The color conveys the surface normal of the object being touched, similar to photometric stereo. These black dots are \u201cmarkers\" that are physically embedded within the GelSight\u2019s elastomer gel. Thus, GelSight records a video in which surface orientation, depth, and shear can be estimated by analyzing the appearance of each video frame. The tactile sensor\u2019s recordings are recorded concurrently with visual images from an ordinary webcam (both at approximately 27 Hz). Both videos (tactile and visual) are recorded on a single computer. 3.2 Annotating the dataset To make it easier to analyze results and train models, we provide annotations for material categories and frames within the press. Detecting the press. As the data collectors move between objects, the tactile sensor does not make contact with anything. Thus, for convenience, we provide the subset of frames within the touch, so that applications can use trimmed videos (Sec. 4). To obtain these timings, we train a detector for press detection. We (the authors) hand-label 10k frames and train a binary ResNet-18 classi\ufb01er [27] that operates on GelSight images. On our test set of 2k hand-labeled frames, our classi\ufb01er obtains 97% accuracy (chance is around 66%). To further ensure high accuracy, we hire workers to review the frames within the touch and correct errors. 2Following common practice, the tactile images are enhanced for visualization purpose. Contrast and sharpness are increased by 30% and saturation is increased by 20%. 4 \fTable 1: Tactile datasets. We compare attributes of our dataset with several previously proposed datasets. Object inst. Touches Source Real-world Environment Sensor More Than a Feeling [8] 65 6.5k Robot \u2713 Tabletop GelSight [72] The Feeling of Success [9] 106 9.3k Robot \u2713 Tabletop GelSight [72] VisGel [42] 195 12k Robot \u2713 Tabletop GelSight [72] ObjectFolder 1.0 [21] 100 Synthetic % DIGIT [36] ObjectFolder 2.0 [22] 1000 Synthetic % GelSight [72] Burka et al. [7] 357 1.1k Human \u2713 Mostly Indoor Multiple Sensors Touch and Go (Ours) 3971 13.9k Human \u2713 Indoor GelSight [72] Labeling materials. We label the material category for all (visual) video frames that our detector predicts within the press, using a labeling scheme similar to [51]. Online workers assign a label from a list of categories. If an object is not in this category list, the workers will label it other material. To ensure accuracy, we have 5 workers label each image. We show the distribution of labels in Fig. 2. 3.3 Dataset analysis We analyze the contents of our dataset and compare it to other works. Data distribution. In Fig. 2, we show statistics of labeled materials and scene types, and provide qualitative results from the dataset. It contains approximately 13.9k detected touches and approximately 3971 individual object instances. Since we do not explicitly label instances, we obtain the latter number by exhaustively counting the objects in 10% of the videos and extrapolating. Our dataset is relatively balanced between indoor (52.2%) and outdoor (47.8%) scenes. We found that several categories, namely synthetic fabric, tile, paper, and leather, are only present in our indoor scenes, while tree, grass, plant, and sand are only present in outdoor scenes. The remaining materials exist in both scenes. We provide more details in the supplement. Comparison to other datasets. In Table 1, we compare our dataset to several previously proposed visuo-tactile datasets collected by robots, by humans, or through simulation. Our dataset contains approximately 4\u00d7 as many object instances as the second-largest dataset, the simulation-based ObjectFolder 2.0 [22], and 11\u00d7 larger compared with the human-collected dataset by Burka et al. [7]. Compared to the existing robot-collected datasets, ours contains more touches (e.g., 1.15\u00d7 more than VisGel [42] and 1.5\u00d7 more than The Feeling of Success [9]). Our dataset also contains data from more diverse scenes than prior work, with a mixture of natural indoor and outdoor scenes. In contrast, robot-collected datasets [8, 9, 42] are con\ufb01ned to a single lab space containing the robot. Qualitative comparison to other datasets We show qualitative examples of data from other datasets in Fig. 3 to help understand the differences between our dataset and those of previous work: Object Folder 2.0 [22], which contains virtual objects, and two robotic datasets: Feeling of Success [9], and VisGel [42]. We show examples from indoor scenes, since the other datasets do not contain outdoor scenes, and with rigid materials (since the virtual scenes do not contain deformable materials). Each row illustrates objects which are composed of similar materials, along with their corresponding GelSight images. As can be seen, the robot-centric datasets [42, 9] are con\ufb01ned to a \ufb01xed space. Their objects are also smaller than those in our dataset, since they must be capable of being grasped by the robot\u2019s gripper. Synthetic datasets [21, 22] do not contain complex microgeometry, and their rigid objects do not deform when pressed. 4 Applications To evaluate the effectiveness of our dataset, we perform tasks that are designed to span a variety of application domains, including representation learning, image synthesis, and future prediction. 4.1 Multimodal self-supervised representation learning We ask, \ufb01rst, whether we can use the multimodal data to learn representations for the tactile modality by associating touch with sight. We then ask how well the learned representations convey material properties from our dataset, and whether they are useful for robotic learning tasks whose data has been collected in other works [8, 9]. The latter task requires a signi\ufb01cant amount of generalization, since the objects manipulated by robots while our dataset is collected by humans. 5 \fTouch and Go (Ours) Object Folder: 5, 60, ?, 475 Feeling of Success: 1, 179, 193, 214 VisGel: 2_2004_133, 4_4244_178, 8_8003_190, 9_9003_197 Object Folder 2.0 Feeling of Success VisGel Object Folder 2.0 [22] Feeling of Success [9] VisGel [42] Ours Figure 3: Visuo-tactile data from other datasets. We provide qualitative examples of visual and tactile data from other datasets (left), along with examples from similar material taken from our dataset (right). Our goal is to learn a representation that captures the information and correspondences between visual and tactile images, which can be useful for downstream tasks. Given the visual and tactile datasets, XI and XT , we aim to extract the corresponding visual-tactile pairs, {xi I, xi T } and mismatched pairs {xi I, xj T } using the Contrastive Multiview Coding (CMC) model proposed by Tian et al. [63]. The detailed procedure is shown below. For each visual-tactile image pair, we \ufb01rst encode visual images and tactile inputs as L2-normalized embeddings using two networks, where zI = f\u03b8I(xI) for visual images and zT = f\u03b8T (xT ) for tactile images. Recall that our goal is to \ufb01nd the corresponding sample of the other modality, given a set S that contains both the corresponding example and K \u22121 random examples. When matching a visual example xi I to tactile, the loss for matching visual images to tactile images is: LXI,XT contrast = \u2212log exp(f\u03b8I(xi I) \u00b7 f\u03b8T (xi T )/\u03c4) PK j=1 exp(f\u03b8I(xi I) \u00b7 f\u03b8T (xj T )/\u03c4) (1) where \u03c4 = 0.07 is a constant and j indexes the tactile examples in S. Analogously, we can obtain a loss in which tactile examples are matched to images, LXT ,XI contrast. We minimize both losses: L(XI, XT ) = LXI,XT contrast + LXT ,XI contrast (2) The training details and hyperparameters are provided in the supplementary material. 4.2 Tactile-driven image stylization Touch provides complementary information that may not be easily conveyed through other modalities that are commonly used to drive image stylization, such as language and sound. For example, touch can precisely de\ufb01ne how smooth/rough a surface ought to be, and express the subtle shape of its microgeometry. A model that can successfully predict these properties from visuo-tactile data therefore ought to be able to translate between modalities. Inspired by the audio-driven image stylization of Li et al. [41], we propose the task of tactile-driven image stylization: making an image look as though it \u201cfeels like\u201d a given touch signal. Following Li et al. [41], we base our approach on contrastive unpaired translation (CUT) [53]. Given an input image xI and a tactile example xT , our goal is to manipulate the image via a generator such that the manipulated pair \u02c6 xI = G(xI, xT ) is more likely to co-occur in the dataset. Our model consists of an image translation network that is conditioned on a tactile example (a GelSight image). The loss function encourages the model to preserve the image structure, which is enforced using an image-based contrastive loss [53], while adjusting the image style such that the resulting textures are more likely to co-occur with the tactile signal, which is measured using a visuo-tactile discriminator. Making sight consistent with touch. We train the model with a discriminator D that distinguishes between real (and fake) visuo-tactile pairs. During training time, we shuf\ufb02e the dataset to generate the 6 \fset Sn containing mismatched image-tactile pairs {xI, x\u2032 T } \u2208Sn. Likewise, we de\ufb01ne the original dataset as Sm which contains matched image-tactile pairs {xI, xT } \u2208Sm. In formal terms, the visual-tactile adversarial loss can be written as: LVT = E{xI,xT }\u223cSmlogD(xI, xT ) + E{xI,x\u2032 T }\u223cSnlog(1 \u2212D(G(xI, x\u2032 T ), x\u2032 T )) (3) Loss. We combine our visuo-tactile discriminator with the structure-preserving loss used in Li et al. [42], which was originally proposed by CUT [53]. This loss, which we call LCUT, works by training a contrastive learning model that puts patches in the input and predicted images into correspondence, such that patches at the same positions are close in an embedding space. Please see the supplement for more details. The overall loss is: LTDIS = LVT + LCUT. (4) 4.3 Multimodal future touch prediction Inspired by the challenges of action-conditional future touch prediction [62], we use our dataset to ask whether visual data can improve our estimates of future tactile signals: i.e., what will this object feel like in a moment? Visual information can help touch prediction in a number of ways, such as by conveying material properties (e.g., deformation) and and geometry. It can also provide information about which action is being performed. One may thus also consider it as an implicit form of action conditioning [19], since the visual information provides information analogous to actions. We adapt the video prediction architecture of Geng et al. [23] to this task. This model is based on residual networks [27, 29] with 3D space-time convolutions (please see the supplement for architecture details). We predict multiple frames by autoregressively feeding our output images back to the original model. Given a series of paired visual and tactile images from times 1 to t, {(x1 I, x1 T ), ..., (xt I, xt T )}, our goal is to predict the subsequent tactile image, \u02c6 x(t+1) T . We train the model with L1 and perceptual loss, following [23]. 5 Experiments 5.1 Self-supervised feature learning We train a self-supervised model that learns to associate images with touch. We evaluate this learned representation on two downstream tasks: robot grasping and material understanding tasks. . Robotic grasping task. For the robot grasping task, we use the experimental setup and dataset of Calandra et al. [9]. Thus, the task requires generalizing to data recorded in a very different environment, and with different GelSight sensors. The goal of this task is to predict whether a robotic arm will successfully grasp an object, based on inputs from two GelSight images recorded before and after grasping. Since there is no standard training/test split from [9], we split their objects randomly into training/test. The resulting dataset contains 68 objects and 5921 touches for training, 16 objects and 1204 touches for validation, and 21 objects and 1204 touches for testing. Similar to the tactile-only model from Calandra et al. [9], we compute features for each of the 4 tactile images (before/after images for 2 tactile sensors) using our self-supervised model. We concatenate these features together and train a linear classi\ufb01er to solve this binary classi\ufb01cation task. Material understanding tasks. We evaluate whether the learned features convey material properties. Given tactile features, we recognize: 1) material categories, 2) hard vs. soft surfaces, and 3) smooth vs. rough surfaces. Following [51], we re-categorize material categories to generate soft/hard labels and hire online workers to label the smooth/rough according to the visual image. Since the smoothness and roughness may vary within a material category, we hire online workers to label the smooth vs. rough according to the visual image. To avoid providing our self-supervised learning model with object instances that appear in the linear probing experiment, we split the dataset into an unlabeled set containing 5172 touches (51.7%), a labeled training set of 3921 touches (39.2%), and a labeled test set of 923 touches (9.1%). We split the dataset by video (rather than by frame) to avoid having the same (or nearly same) object appear in both training and test. 7 \fTable 2: Comparison of pretrained models on ImageNet and other tactile datasets on different downstream tasks. We also evaluate variations of our model trained only on subsets of the material classes. Dataset Method Grasping Acc(%) Material Acc(%) Hard/Soft Acc (%) Rough/Smooth Acc (%) Chance 56.1 18.6 66.1 56.3 ImageNet [15] Supervised 73.0 46.9 72.3 76.3 Object Folder 2.0 [22] Visuo-tactile CMC 69.4 36.2 72.0 69.0 VisGel [42] Visuo-tactile CMC 75.6 39.1 69.4 70.4 Ours 25% Classes Visuo-tactile CMC 62.3 25.7 67.3 65.3 Ours 50% Classes Visuo-tactile CMC 66.5 35.9 71.2 66.8 Ours 75% Classes Visuo-tactile CMC 70.8 48.4 73.3 74.7 Ours 100% Classes Visuo-tactile CMC 78.1 54.7 77.3 79.4 Implementation details. We train our model for 240 epochs, using the optimization parameters from CMC [63], after adjusting the learning rate schedule to compensate for longer training. We set the weight decay to be 10\u22124. We train our model with the batch size of 128 on 4 Nvidia 2080-Ti GPUs. For the downstream classi\ufb01cation tasks, we froze our network weights and obtained visual features by performing global average pooling on the \ufb01nal convolutional layer. We follow the approach of [63] for learning the linear classi\ufb01er. Comparison to other feature sets. We show downstream classi\ufb01cation results in Table 2. To evaluate the effectiveness of our dataset, we compare our learned features to those of several other approaches. These include using supervised ImageNet [15] features, which are commonly used to represent GelSight images [74, 9, 8], and visual CMC [63] features trained on ImageNet, which treats the L and ab color channels of an image as different modalities. We see that our model obtains signi\ufb01cantly better performance than these models on both tasks, due to its ability to learn from real tactile data. These results suggest that our dataset provides a useful signal for training self-supervised tactile representations. We also show that increasing the material categories leading to much better downstream performance. 5.2 Tactile-driven image stylization We use our model to modify the style of an image to match a touch signal. Implementation details. Following Li et al. [41], during training we sample a random image from the dataset, along with a second visuo-tactile pair, and use the pair to restyle the image, using the loss in Eq. 4. We provide architectural details in the supplement. For the discriminator we adopt the PatchGAN architecture [28]. The architecture of discriminator follows [41], which concatenates the two input images channel-wise, and passes the combined images to the discriminator. We train our model on 4 Nvidia 2080-Ti GPUs for 100 epochs with a batch size of 8 and the learning rate of 0.0002. We augment the vision images with random cropping and horizontal \ufb02ipping. Table 3: Quantitative results for tactiledriven image stylization. Method CMC Material Baseline 0.165 0.107 CycleGAN [78] 0.178 0.129 Ours 0.197 0.142 Experimental setup. Following [41], we evaluate our model by restyling images in our dataset with random tactile inputs, both of which are taken from the test set. We use evaluation metrics that measure the consistency of the manipulated image with the example used for conditioning. First, we measure the similarity of the manipulated image and the tactile example used for conditioning. Similar to [41], we use our trained CMC model, by taking the dot product between visual and tactile embeddings. Second, we compare material prediction consistency between the manipulated image with the (held out) conditional image. We use our material classi\ufb01er to categorize the predicted and conditioning images, and measure the rate at which they agree. Since this is a novel task, we create a second variation of our model for comparison, following [78]. This model performs the stylization using CycleGAN, rather than CUT (see supplement for details). Quantitative results. Quantitative results are shown in Table 3. Here, the \u201cbaseline\u201d indicates results from original image before stylization. We can see that the CUT-based method obtains higher 8 \fFigure 4: Qualitative results of our model on tactile-driven image stylization. For each row, we show an input image (left) and the manipulated image (to its right) obtained by stylizing with a given tactile input (right side). For reference, we also show the image that corresponds to the tactile example at rightmost (not used by the model). The manipulated images convey physical properties of the tactile signal, such as its roughness (e.g., \ufb01rst three rows) or smoothness (e.g., row 10). Other inputs result in images that combine the properties of two inputs (e.g., by adding grass, as in row 8). We also show failure cases in the last row. Zoom in for better view. CMC similarity than a CycleGAN-based method [78]. In terms of material classi\ufb01cation consistency, our model consistently outperforms CycleGAN-based method [78]. Qualitative results. In Fig. 7, we show results from our model. Our model successfully manipulates images to match tactile inputs, such as by making surfaces rougher or smoother, or by creating \u201chybrid\u201d materials (e.g., adding grass to a surface). These results are obtained without having access to the tactile example\u2019s corresponding image, suggesting that the model has learned which physical properties are shared between sight and touch. 5.3 Multimodal video prediction We evaluate our model for predicting future tactile signals. We compare a tactile-only model to a multimodal visuo-tactile model, and show that the latter obtains better performance. Experimental setting. We evaluate the effectiveness of multimodal inputs using three context frames to predict the next frame under two different time horizons: skipping 3 and 5 frames between 9 \fTable 4: Quantitative results for video prediction. Time horizon Method Modality L1 \u2193 SSIM \u2191 LPIPS \u2193 Skip 3 frames SVG [16] Touch only 0.782 0.572 0.391 Geng et al. [23] Touch only 0.628 0.708 0.103 SVG [16] Touch + vision 0.757 0.602 0.368 Geng et al. [23] Touch + vision 0.617 0.719 0.091 Skip 5 frames SVG [16] Touch only 0.807 0.513 0.412 Geng et al. [23] Touch only 0.691 0.698 0.279 SVG [16] Touch + vision 0.762 0.546 0.397 Geng et al. [23] Touch + vision 0.663 0.713 0.265 Time Tactile Inputs Ground Truth Multimodal Prediction Single-modal Prediction Video Inputs Figure 5: Future touch prediction. We show the results of tactile-only and visuo-tactile models. contexts. Following [23], we adopt three evaluation metrics: MAE, SSIM [69] and LPIPS [76]. We provide training hyperparameters in the supplement. Multimodal vs. single-modal prediction. We show the quantitative results in Table 4. Here, we adopt two video prediction baselines [16] and [23]. We can see that, by incorporating our dataset\u2019s visual signal, the models gain a constant performance increase under different evaluation metrics for both model, under both experimental settings. The gap becomes larger for longer time horizon, suggesting that visual information may be more helpful in this case. 6 Discussion We proposed Touch and Go, a human-collected visuo-tactile dataset. Our dataset comes from realworld objects, and is signi\ufb01cantly more diverse than prior datasets. We demonstrate its effectiveness on a variety of applications that involve robotic manipulation, material understanding, and image synthesis. In the tradition of previous work [51], we see our work as a step toward human-collected multimodal data collection, in which humans equipped with multiple sensors collect diverse dataset by recording themselves physically interacting with the world. We hope this data will enable researchers to study diverse visuo-tactile learning applications, beyond the \u201crobotics-centric\u201d domains that are often the focus of previous efforts. Limitations. Collecting diverse tactile data is an ongoing challenge, since it requires physically being present in the locations where data is collected. While adding human collectors improves diversity in many ways, our dataset was mainly collected in one geographic location (near University of Michigan\u2019s campus). Consequently, the data we recorded may not generalize to all spaces. The use of humans in the data collection process also potentially introduces bias, which differs from \u201crobotic\u201d or \u201cvirtual data\u201d bias. For example, humans may choose unrepresentative parts of the objects to probe, and do not perform actions with consistent force. The humans who recorded the dataset may also not be representative of the general population, which may introduce bias (e.g., in skin tone). Acknowledgements. We thank Xiaofeng Guo and Yufan Zhang for the extensive help with the GelSight sensor, and thank Daniel Geng, Yuexi Du and Zhaoying Pan for the helpful discussions. This work was supported in part by Cisco Systems and Wang Chu Chien-Wen Research Scholarship. 10"
+                },
+                {
+                    "url": "http://arxiv.org/abs/2110.09780v2",
+                    "title": "Improving Emotional Speech Synthesis by Using SUS-Constrained VAE and Text Encoder Aggregation",
+                    "abstract": "Learning emotion embedding from reference audio is a straightforward approach\nfor multi-emotion speech synthesis in encoder-decoder systems. But how to get\nbetter emotion embedding and how to inject it into TTS acoustic model more\neffectively are still under investigation. In this paper, we propose an\ninnovative constraint to help VAE extract emotion embedding with better cluster\ncohesion. Besides, the obtained emotion embedding is used as query to aggregate\nlatent representations of all encoder layers via attention. Moreover, the\nqueries from encoder layers themselves are also helpful. Experiments prove the\nproposed methods can enhance the encoding of comprehensive syntactic and\nsemantic information and produce more expressive emotional speech.",
+                    "authors": "Fengyu Yang, Jian Luan, Yujun Wang",
+                    "published": "2021-10-19",
+                    "updated": "2022-01-28",
+                    "primary_cat": "cs.SD",
+                    "cats": [
+                        "cs.SD",
+                        "eess.AS"
+                    ],
+                    "main_content": "INTRODUCTION Emotional speech synthesis is widely applied in various scenarios, such as voice assistant, audio customer service, audio book and etc. Generally, a group of emotion types are de\ufb01ned based on product requirements. For each emotion type, hundreds of sentences are designed and recording lines are collected accordingly. Due to the limited data of each type, usually a multi-emotion TTS model training is implemented by leveraging all the data. To distinguish the data from different emotion types, two questions emerge:1) how to get the emotion embedding from category label; 2) how to inject the emotion embedding into TTS acoustic model. Based on the popular encoder-decoder TTS frameworks, many attempts have been reported to address these two questions. For the emotion embedding, the basic idea is to learn an embedding vector for each emotion type [1, 2], which is usually called as one-hot embedding or look-up table. However, emotion expression always has slight vibration in intensity or status among utterances, even if the voice talent is guided to keep consistent. Global Style Token(GST) [3] is then introduced to get utterance-level embedding. At the same time, an emotion classi\ufb01er can be added to restrict embedding vectors to be clustered by category [4]. During inference, the average embedding vector of target emotion type may be employed [5], while [6] even tries to control the emotion intensity and inter-emotion transition by interpolating the embedding. Also, GST can generate \ufb01ne-grained embedding at phoneme level [7]. In recent studies, Variational AutoEncoder(VAE) [8] shows stronger capabilities in disentanglement, scaling and interpolation for expression modeling [9] and style control[10]. However, VAE training is not robust and usually suffers from posterior collapse. For the injection of emotion embedding, mostly popular method is to concatenate the embedding vector into decoder input [1, 3, 5, 6, 4, 7, 9, 10], while some studies inject emotion embedding to both attention and decoder RNN layers of Tacotron framework [2]. Both of them in\ufb02uence the decoder merely, not considering the effects of emotion embedding to the textual emotion. In this paper, we propose a framework with innovative solutions for both of the above questions. Firstly, VAE is applied for the emotion embedding. Instead of conventional KL-divergence regularizer, the new constraint expects the means of the embedding vectors are on the surface of the unit sphere while all dimensions have a uniform standard deviation. Secondly for the injection of emotion embedding, this paper uses it as one resource of queries in the attention-based text encoder aggregation to enable emotion-speci\ufb01c sentential information encoding. Another resource of queries are from text encoder themselves as we did in previous study [11]. In summary, the multi-query attention is designed to capture the syntactic and semantic information better for emotional speech generation. 2. RELATED WORK Some previous works also employ VAE embedding as query to attend encoder output. In BVAE-TTS [12], reference audio is encoded by VAE to get frame-level latent variables. These variables work as query to attend text encoder output for better decoder input. In VARA-TTS [13], however, the output of VAE intermediate layers (called hierarchical latent variables) are all used as query. Different from them, the proposed method encodes reference audio into one vector, i.e. a utterance level emotion embedding, rather than a frame level sequence. Moreover, both BVAE-TTS and VARA-TTS only arXiv:2110.09780v2  [cs.SD]  28 Jan 2022 \fuse the output from the last layer of text encoder as attention memory and the attention is implemented along the time axis. In our framework, the outputs from all intermediate encoder layers are leveraged and the attention is implemented along the stacked layers. In our previous work[11], we utilize the contexts extracted form the stacked layers to do self-learned multi-query attention over an expressive corpus. In this paper, we propose to introduce acoustic emotion information to the multi-query for better emotion modeling over a multi-emotion corpus. 3. PROPOSED MODEL Encoder Pre-Net Character Embeddings Scaled  Positional  Encoding Self-attention Layer 1 Self-attention Layer l Self-attention Layer L ... ... Decoder Mel Spectrogram GMM Attention 3.3. Text Encoder Aggregation Add & Norm Add & Norm FeedForward Network Multi-Query Attention Reference  Encoder FC FC z \u00b5  \u03c3 3.1.  Emotion Encoder Classifier 3.2. Text Encoder 3.4.  Combined  Multi-Query Fig. 1. Proposed architecture with multi-query attention. Figure1 illustrates our proposed approach with text encoder aggregation on exploiting emotional contexts for emotional speech synthesis. It contains a self-attention-based text encoder, an RNN-based auto-regressive decoder, a GMMbased attention[14] bridging them, a VAE-based emotion encoder and an emotion classi\ufb01er. WaveRNN[15] is adopted to convert mel spectrogram to waveforms. The augmented encoder with a context aggregation module will be described in detail. 3.1. SUS-constrained VAE VAE does not generate the latent vector directly. Instead, it generates Gaussian distributions each represented by a mean and a standard deviation. During inference, a latent vector is sampled from these distributions. If no additional constraints are employed, the standard deviation will trend to be 0 and sampled latent vectors will always be the mean. Therefore, the desired sampling mechanism becomes invalid. However if a Kullback-Leibler (KL) divergence regularizer is added, it is usually found the generated distributions become normal Gaussian distribution independently of the inputs. Both these two phenomena can be regarded as posterior collapse. Many attempts have been made to address this puzzle, such as annealing strategy in [10]. The critical problem here, we think, is the distances between the means should be in the similar order with their standard deviations. If the distance between means is much bigger than their standard deviation, latent vectors will collapse to the means. Conversely, if the distance between means is much smaller, latent vectors will collapse to be independent on the input. Inspired by it, this paper restricts the means approaching to the Surface of the Unit Sphere (SUS) while set the standard deviations to be an appropriate constant for all dimensions, such as 1. In this way, the distributions of latent vectors will \ufb01nally have appropriate overlapping proportions, which guarantees VAE\u2019s advantages of disentanglement, scaling and interpolation. Formally, sampling z from distribution N(\u00b5, \u03c32I) is decomposed to \ufb01rst sampling \u03f5 \u2208(0, I) and then computing z = \u00b5 + \u03c3 \u00b7 \u03f5, where \u00b7 represents an element-wise product. We restrict the means \u00b5 approaching to the surface of the unit sphere through L2 distance: lossSUS = ( qX (\u00b52) \u22121)2. (1) Meanwhile, we set the standard deviations \u03c3 as a constant. 3.2. Self-attention based Encoder Self-attention based sequence-to-sequence framework has been successfully applied to speech synthesis, such as TransformerTTS[16] and Fastspeech[17]. We also adopt self-attention networks(SAN) as our based text encoder following [16]. Formally, from the previous self-attention block output Hl\u22121, the multi-head attention Cl and the followed feed forward network Hl can be computed by: Cl = LN(MH(headl 1, . . . , headl H) + Hl\u22121), (2) Hl = LN(FFN(Cl) + Cl), (3) where MH(\u00b7), FFN(\u00b7) and LN(\u00b7) represent multi-head attention, feed forward network and layer normalization respectively. And in multi-head attention, each head split from the previous self-attention block is calculated as: headh = softmax(QhKT h \u221a d )\u00b7Vh, (4) where {Q, K, V } are queries, keys and values, d represents the hidden state\u2019s dimension. 3.3. Weighted Aggregation As different SAN layers extract different levels of prosodicrelated sentential context information[18], we propose a text encoder aggregation module, aggregating them to learn a comprehensive sentence representation to enhance the emotion of the \ufb01nal generated speech. In detail, we utilize a multi-query attention to learn the contribution of each block across the stacked layers. Formally, given a sequence X of \fN elements, the multi-query calculates the correlation of individual sentential contexts {H0, . . . , HL} in phoneme level, which are transposed as {headg 1, . . . , headg N} for keys and values. We modify Eq. (2) to obtain the weighted contexts: Cg = LN(MH(headg 1, . . . , headg N) + HL), (5) Hg = LN(FFN(Cg) + Cg). (6) There are several choices for the multi-query. The \ufb01rst is as our previous work does[11], utilizing {H0, . . . , HL} to obtain the textual multi-query Qt. This self-learned weighted aggregation module leverages the textual contexts information to learn the combination relationship across layers. 3.4. Combined multi-query For the textual multi-query, the sententail contexts are totally extracted from the stacked textual encoder layers, which does not consider the proved important information from emotion embedding. Commonly the emotion embedding is directly concatenated to the encoder output, in\ufb02uencing the decoder merely. Assuming emotion embedding affects the textual emotion signi\ufb01cantly, we propose to introduce contexts extracted from emotion embedding to out multi-query on the basis of direct concatenation. In details, we employ the output of VAE vae to learn a weighted matrix Qa: Qa = DNN(vae), (7) where DNN(\u00b7) represents a nonlinear transformation with tanh. Then, we combine the contexts information from text and emotion embedding with a learned coef\ufb01cient to investigate the effectiveness of a comprehensive multi-query: Qc = Qt + Sigmiod(w)Qa, (8) where Sigmoid(\u00b7) is a activation function. 4. EXPERIMENTS 4.1. Basic setups To investigate the effectiveness of modeling emotion, we carried out experiments on a Mandarin corpora from a male speaker with 7 emotion categories (neutral, happy, sad, angry, shy, concerned and surprised), which contains about 4.4 hours and a total of 4371 utterances. Except for the neutral, each emotion categories has nearly 500 utterance with consistent emotional strength, separated to non-overlapping training and testing sets (with data ratio 9:1) respectively. For linguistic inputs, we use phones, tones, character segments and three levels of prosodic segments: prosodic word (PW), phonological phrase (PPH) and intonation phrase (IPH). And 80-band mel-spectrogram is extracted from 16KHz waveforms as acoustic targets. For objective evaluation, we conduct mel cepstral distortion (MCD) on test set. And we conduct A/B preference test on 30 randomly selected test set samples with 20 native Chinese listeners as subjective evaluation. Table 1. MCD scores of different systems for parallel transfer. Emotion BASE BASE-SUS SA-WA SA-WAC Neutral 5.5 5.15 4.44 4.22 4.2. Model details The decoder structure in Tacotron2[19] is used as our baseline. But the CBHG encoder and GMMv2 attention are adopted instead for superior naturalness and stability[14], where the output of VAE[10] is added to the encoder output. For the encoder using SAN, the input text embeddings with positional information are pushed to a 3-layer CNN \ufb01rstly. Then each self-attention block includes an 8-head self-attention and a feed forward sub-network. In the text encoder, there are totally 6 self-attention blocks. As for the aggregation module, HL is double fed for the convenience of implementation. In VAE module [10], the reference encoder consists of six 2D-convolution layers and a GRU layer. Further, the plugged emotion classi\ufb01er in all systems has a fully connected (FC) layer with ReLu activation and a 7-unit output layer. In our proposed SUS-constrained VAE, we set the standard deviation to 1. WaveRNN is used as vocoder totally following [15], trained using the neutral set about 16 hours with the same speaker. For comparison, we built the following different systems: \u2022 BASE: Baseline system following VAE-Tacotron2 [10] with CBHG as text encoder and slightly modi\ufb01ed GMMv2 attention. \u2022 BASE-SUS: Baseline system with SUS constraint instead of KL divergence constraint for VAE training described in Section 3.1. \u2022 SA-WA: SAN based encoder with the aggregation module using textual multi-query described in Section 3.3. \u2022 SA-WAC: SAN based encoder with the aggregation module using combined multi-query described in Section 3.4. (The learned coef\ufb01cient for combination over the multi-emotion corpus is 0.47.) 4.3. Objective Evaluation The MCD results of different systems with parallel transfer are showed in Table 1. It shows that SUS-constrained VAE has lower MCD than KL constrained VAE, with better ability in generating more similar emotional speech for ground-truth. Meanwhile, it demonstrates that SAN based encoder with text encoder aggregation can improve the performance than the RNN based encoder, where combined multi-query is a better way than textual multi-query in text encoder aggregation to \fextract emotional contexts. With the help of the injection of the emotion embedding to the textual multi-query, the synthesized speech samples turn into more similar ones to the real speech samples. 4.4. Subjective Evaluation As Figure 2 shows, we conduct AB preference tests on the emotional test sets with non-parallel transfer, which include 7 emotion categories. The listeners are asked to select preferred audio according to the overall impression on the expressiveness of emotion in the testing samples1. Comparing the BASE system, we \ufb01nd that with the SUS-constrained VAE, our proposed BASE-SUS system obtains much more preferred, due to more expressive emotional speech. Meanwhile, both of the two systems using self-attention based encoder and text encoder aggregation achieve higher preference scores than the RNN based encoder one, which demonstrates that selfattention based aggregation module is a better strategy for generating more expressive emotional speech. Further, combined multi-query attention brings extra performance gain than simple textual multi-query attention according to the A/B preference test. 27% 13% 12% 15% 30% 13% 46% 29% 43% 74% 42% 56% SA-WA vs SA-WAC BASE vs SA-WAC BASE vs SA-WA BASE vs BASE-SUS Left No Preference Right Fig. 2. A/B preference results for non-parallel transfer with con\ufb01dence intervals of 95% and p-value<0.0001 from t-test. 4.5. Analysis Emotion Distortion We visualize the two systems with parallel transfer for seven emotion categories in emotion embedding space by t-distributed stochastic neighbor embedding (tSNE) plots[20]. Figure 3 shows that both the BASE system and the BASE-SUS system appear clear cluster separation, which demonstrates that both two systems have high classi\ufb01cation accuracy. But the cluster cohesion of the BASE-SUS system is much better than that of the BASE system. It means that the proposed SUS-constrained VAE can extract emotion information more robustly with less disturbance, which \ufb01nally helps TTS to generate emotional speech more accurately and expressively. They are also certi\ufb01ed in above objective and subjective evaluations. Prosody Correlation To further estimate the emotion with parallel transfer for statistical signi\ufb01cance, phonemelevel intensity, rhythm and intonation of audio are selected. 1Samples can be found from: https://fyyang1996.github.io/emotion (a) The BASE system (b) The BASE-SUS system Fig. 3. Visualization of two systems using t-SNE for seven emotion categories. Table 2. Correlation in relative energy, duration and F0 within a phoneme computed from different systems for parallel transfer. BASE BASE-SUS SA-WA SA-WAC E 0.542 0.56 0.582 0.595 Dur. 0.806 0.811 0.820 0.824 F0 0.322 0.338 0.403 0.422 We extract three acoustic features commonly associated with emotion: relative energy within each phoneme (E), duration in ms (Dur.) and fundamental frequency in Hertz (F0) according to [21, 11]. Additional alignments are done to catch the three prosody attributes in phoneme level. The Pearson correlation coef\ufb01cient between each system and the ground truth is calculated to evaluate these statistics, using 100 random samples in the test set. The higher Pearson correlation coef\ufb01cient value demonstrates the higher accuracy of the predicted prosody attribute. From Table 2 we know that out proposed BASE-SUS achieves higher correlation scores than baseline in all three prosody attributes, which demonstrates that our SUS-constrained VAE has better reconstruction performance in phoneme-level intensity, rhythm and intonation. Meanwhile, in all three prosody attributes, our proposed both SA-WA system and SA-WAC system obtain higher correlation scores than baseline, and SA-WAC system acquires the highest scores. Consequently, we believe that the combined multi-query attention in text encoder aggregation has strong ability in modeling all the three emotional associated attributes. 5."
+                },
+                {
+                    "url": "http://arxiv.org/abs/2008.00613v1",
+                    "title": "Exploiting Deep Sentential Context for Expressive End-to-End Speech Synthesis",
+                    "abstract": "Attention-based seq2seq text-to-speech systems, especially those use\nself-attention networks (SAN), have achieved state-of-art performance. But an\nexpressive corpus with rich prosody is still challenging to model as 1)\nprosodic aspects, which span across different sentential granularities and\nmainly determine acoustic expressiveness, are difficult to quantize and label\nand 2) the current seq2seq framework extracts prosodic information solely from\na text encoder, which is easily collapsed to an averaged expression for\nexpressive contents. In this paper, we propose a context extractor, which is\nbuilt upon SAN-based text encoder, to sufficiently exploit the sentential\ncontext over an expressive corpus for seq2seq-based TTS. Our context extractor\nfirst collects prosodic-related sentential context information from different\nSAN layers and then aggregates them to learn a comprehensive sentence\nrepresentation to enhance the expressiveness of the final generated speech.\nSpecifically, we investigate two methods of context aggregation: 1) direct\naggregation which directly concatenates the outputs of different SAN layers,\nand 2) weighted aggregation which uses multi-head attention to automatically\nlearn contributions for different SAN layers. Experiments on two expressive\ncorpora show that our approach can produce more natural speech with much richer\nprosodic variations, and weighted aggregation is more superior in modeling\nexpressivity.",
+                    "authors": "Fengyu Yang, Shan Yang, Qinghua Wu, Yujun Wang, Lei Xie",
+                    "published": "2020-08-03",
+                    "updated": "2020-08-03",
+                    "primary_cat": "eess.AS",
+                    "cats": [
+                        "eess.AS",
+                        "cs.SD"
+                    ],
+                    "main_content": "Introduction Recently, the naturalness of corpus-based text-to-speech (TTS) has been signi\ufb01cantly improved with the use of attention-based sequence-to-sequence (seq2seq) mapping framework [1, 2]. Such so-called end-to-end (E2E) systems directly employ a text encoder network to learn linguistic, syntactic and semantic information from simple character or phoneme sequences. The sequence of aggregated textual representation is further attended by an acoustic decoder network through some attention mechanism, producing predicted speech representations (e.g., mel-spectrogram) that are subsequently transformed to waveforms via a neural vocoder. Sentential context [3] mainly involves the latent syntactic and semantic information embedded in the text, recently proved to be important in natural language processing (NLP) tasks [3, 4]. It might be essential to the naturalness of speech First author performed part of this work at Xiaomi. Lei Xie is the corresponding author. This work was partially supported by the National Key Research and Development Program of China (No.2017YFB1002102). synthesis as well, especially for a system built upon an expressive corpus with rich prosodic variations. The seq2seq framework extracts prosodic information solely from the text encoder in an unsupervised way, which is easily collapsed to an averaged expression for expressive contents. To better make use of the sentential context in an E2E framework, one way is feature engineering as the previous generation of TTS does. For example, recent study has shown that exploiting syntactic features in a parsed tree is bene\ufb01cial to the richness of the prosodic outcomes, leading to more natural synthesized speech [5]. However, modeling expressiveness in text-to-speech is still challenging as it refers to different levels of syntactic and semantic information re\ufb02ected in intensity, rhythm, intonation and other prosody related factors. However, it is dif\ufb01cult to de\ufb01ne the relations explicitly between the syntactic/semantic factors and the prosodic factors. To model expressivity, the global style tokens (GST) family [6, 7] learns style embeddings from a reference audio in an unsupervised way, which lets the synthesized speech imitate the style of reference audio. Although the style embeddings from a reference audio is helpful to control the style of synthesized speech, it is hard to choose an appropriate reference audio for each input sentence. Likewise, the variational autoencoder (VAE) models styles or expressivity in a similar way [8]. Recent studies have revealed that self-attention based networks (SAN) [9, 10, 11, 12] have strong ability in capturing global prosodic information, leading to more natural synthesized speech. And unveiled by recent NLP tasks, different SAN encoder layers can capture latent syntactic and semantic properties of the input sentence at different levels [3, 4]. But current SAN-based TTS systems only leverage the highly aggregated latent text representation, usually the outputs of the text encoder, from the simple textual input, to guide the speech generation process. Although the highly aggregated representation can be treated as a global description of the sentential context, it is not enough to generate expressive content according to our experiments as it may disperse the contribution of sentential context embedded in the intermediate SAN layers [13]. In this paper, to excavate the sentential context for expressive speech synthesis, we propose a context extractor to suf\ufb01ciently exploit sentential context over an expressive corpus for seq2seq-based TTS. Speci\ufb01cally, we utilize different levels of representations from the SAN-based text encoder to build a context extractor, which is helpful to extract different levels of syntactic and semantic information [14]. In details, our context extractor \ufb01rst collects the prosodic-related sentential context information from different SAN-based encoder layers, and then aggregates them to learn a comprehensive sentence representation to enhance the expressiveness of the \ufb01nal generated arXiv:2008.00613v1  [eess.AS]  3 Aug 2020 \fEncoder Pre-Net Character Embeddings Scaled  Positional  Encoding Self-attention Layer 1 Self-attention Layer l Self-attention Layer L ... ... Decoder Mel Spectrogram GMM Attention Context Aggregation Add & Norm Add & Norm FeedForward Aggregation Add & Norm Multihead Attention Aggregation Add & Norm FFN Self-attention Multihead  attention or Direct Aggregation thru Concatenation Weighted  Aggregation thru MultiHead  Attention Linear Linear Linear Linear Linear Linear Linear Linear Linear V K Q Scaled Dot-Product Attention Scaled Dot-Product Attention Scaled Dot-Product Attention Concat Linear h Figure 1: Proposed architecture with context aggregation based on Tacotron2 and SAN encoder. speech. Speci\ufb01cally, we investigate two methods of context aggregation: 1) direct aggregation which directly concatenates the outputs of different SAN layers, and 2) weighted aggregation which uses multi-head attention to automatically learn contributions for different SAN layers. Experiments on two expressive corpora show that our approach can produce more natural speech with richer prosodic variations, and weighted aggregation is more superior in modeling expressivity. 2. Proposed Model Figure 1 illustrates our proposed approach on exploiting deep sentential contexts for expressive speech synthesis. It contains a modi\ufb01ed self-attention based text encoder, an auto-regressive decoder and a GMM-based attention [15] to bridge the encoder and the decoder. WaveGlow [16] is adopted to reconstruct waveforms from mel spectrogram. We augment the encoder with a context aggregation module, which will be described in detail. 2.1. Self-attention based Encoder Self-attention based sequence-to-sequence framework has been successfully applied to speech synthesis [9, 10, 17]. In the basic SAN-based text encoder, there is a stack of L blocks, each of which has two sub-networks: a multi-head attention and a feed forward network. The residual connection and layer normalization are applied to both of the sub-networks. Formally, from the previous encoder block output Hl\u22121, the \ufb01rst sub-network Cl and the second sub-network Hl are calculated as: Cl = LN(MultiHead(headl 1, . . . , headl H) + Hl\u22121), (1) Hl = LN(FFN(Cl) + Cl). (2) where MultiHead(\u00b7), FFN(\u00b7) and LN(\u00b7) are multi-head attention, feed forward network and layer normalization respectively. And each head in multi-head attention split from the previous encoder block output is computed by: headh = \u03b1\u00b7V = softmax(QKT \u221a d )\u00b7V, (3) where {Q, K, V } represent queries, keys and values, d is the dimension of the hidden state and \u03b1 represents the weight matrix for each head. 2.2. Direct Aggregation Although the SANs have the ability of directly capturing global dependencies among whole input sequence [18], it may not appropriately exploit the sentential context because it calculates the relevance between the characters without considering the contextual information [3, 14]. Besides, the weighted sum option from the lower layers in SANs has only aggregated the global contextual information, which may weaken the contribution of sequential context extracted in each block. To fully make use of the contexts extracted from each block, we propose a context extractor to aggregate the different levels of contexts to form a comprehensive sentence representation. For the lth self-attention block, we extract the intermediate context from the output Hl through: gl = g(Hl) = MeanPool(Conv1d(Hl)), (4) where Conv1d means 1d-convolution, MeanPool represents mean pooling[19], g(\u00b7) denotes the function to summarize the outputs of self-attention layers, and gl represents the sentential context from lth block. A straightforward and intuitive choice to aggregate the different levels of contexts is through a concatenation operation, with residual connection and layer normalization [9]: Cg = LN(Concat(g0, . . . , gL) + gL), (5) where g0 represents the inputs of the \ufb01rst self-attention layer through Eq. (4). To further integrate the information concatenated from all sentential contexts, we use another round of feedforward network and layer normalization as the \ufb01nal aggregation function [14][20]: g = LN(FFN(Cg) + Cg). (6) Here, g is the \ufb01nal sentential context. 2.3. Weighted Aggregation With direct aggregation, the sequential contexts of each block are simply concatenated to guide the auto-regressive generation, which does not consider the varying importance of each gl. Assuming the sequential contexts in each block may have different contribution to the expressiveness of the synthesized speech, we utilize a self-learned weighted aggregation module across layers to catch the different levels of contribution. In detail, we employ a multi-head attention to learn the contribution of each block. The individual sentential contexts {g0, g1, . . . , gL} are treated as attention memory for the attention based weighted aggregation. Speci\ufb01cally, we transpose the dimension of sequential length with the number of heads in the multi-head attention to combine the contextual information across layers. Therefore, we modify Eq. (5) to obtain the weighted contexts: Cg = LN(MultiHead(g0, . . . , gL) + gL), (7) where the modi\ufb01ed Cg offers a more precise control of aggregation for each gl. 3. Experiments 3.1. Basic setups To investigate the effectiveness of modeling expressiveness, we carried out experiments on two expressive Mandarin corpora \u2013 the publicly-available Blizzard Challenge 2019 corpus [21] from a male talk-show speaker and an internal voice assistant corpus from a female speaker. The talk-show (TS) corpus contains about 8 hours speech of, and the voice assistant (VA) corpus contains about 40 hours of speech. Both corpora are \fTable 1: MCD scores over the two expressive corpora. Corpus BASE SA SA-DA SA-WA TS 8.01 7.48 7.42 7.32 VA 7.60 7.37 7.32 7.23 expressive in prosodic delivery, separated to non-overlapping training and testing sets (with data ratio 9:1). Besides, we also utilize a publicly-available standard reading-style corpus [22] with less expressivity to see how our approach perform. The corpus, named DB1, contains 10 hours of female speech with consistent reading style. Again, we separate the corpus to training and testing with ratio 9:1. Linguistic inputs include phones, tones, character segments and three levels of prosodic segments: prosodic word (PW), phonological phrase (PPH) and intonation phrase (IPH). We use 80-band mel-spectrogram extracted from 22.05KHz waveforms (for TS and VA) and 16KHz waveforms (for DB1) as acoustic targets. We calculate mel cepstral distortion (MCD) on test set for objective evaluation. As for subjective evaluation, we conduct mean opinion score (MOS) and A/B preference test on 30 randomly selected test set samples and 20 native Chinese listeners participated in the tests. 3.2. Model details We use the standard encoder-decoder structure in Tacotron2 [2] as the baseline, but GMM attention is adopted instead because it can bring superior naturalness and stability [15]. For networks using SAN based encoder, a 3-layer CNN is \ufb01rstly applied to the input text embeddings with positional information. Each self-attention block includes an 8-head self-attention and a feed forward sub-network consisting of two linear transformations with 2048 and 512 hidden units. Residual connection and layer normalization are applied to these two sub-networks. There are totally 6 self-attention blocks. In the aggregation module, we double feed gL into aggregation attention function for the convenience of implementation, where the number of heads in multi-head attention are length and the dimension of weighted matrix are [batch, length, 8, 8]. For the remaining part, we adopt the auto-regressive decoder described in [2]. We use WaveGlow as vocoder which follows the structure in [16], trained using the same training set. We built the following systems for comparison: \u2022 Base: Baseline system following Tacotron2 [2] with slightly modi\ufb01ed GMMv2 attention [15]. \u2022 SA: Another baseline system with SAN based encoder described in Section 2.1. \u2022 SA-DA: SAN based encoder with the direct aggregation module fusing all sentential contexts described in Section 2.2. \u2022 SA-WA: SAN based encoder with the weighted aggregation module fusing all sentential contexts described in Section 2.3. 3.3. Objective Evaluation Table 1 shows the MCD results of different systems. It demonstrates that SAN based encoder has lower MCD than the RNN based encoder for both expressive corpora. It also shows that modeling sentential context can further improve the performance of SAN based encoder. Besides, weighted aggregation is a better way than direct aggregation to extract the deep sentential context. With the help of deep sentential context, the SA-WA system achieves the lowest MCD on both expressive Table 2: The MOS over the two expressive corpora with con\ufb01dence intervals of 95%. Corpus BASE SA SA-DA TS 3.84\u00b10.05 3.97\u00b10.06 4.04\u00b10.06 VA 4.11\u00b10.06 4.20\u00b10.06 4.24\u00b10.06 Corpus SA-WA GT TS 4.11\u00b10.06 4.54\u00b10.05 VA 4.36\u00b10.07 4.63\u00b10.05 57.5% 35.0% 44.6% 28.0% 27.3% 54.6% 20.4% 25.0% 81.7% BASE BASE SA SA SA-DA SA-DA SA-WA SA-WA No Preference No Preference No Preference No Preference 13.3% 5.0% 7.5% Figure 2: AB Preference results on TS with con\ufb01dence intervals of 95% and p-value < 0.0001 from t-test. corpora, which indicates that the synthesized speech samples are the most similar ones to the real speech samples. 3.4. Subjective Evaluation We conduct AB preference tests and MOS tests on the two expressive test sets which include a large number of modal particles, interrogatives and exclamations. The listeners are asked to select preferred audio according to the overall impression on the expressiveness of the testing samples1. The AB preference results are shown in Figure 2 and 3 for TS and VA, respectively. MOS results are reported in Table 2. For baseline systems, we can \ufb01nd that the SA system with SAN based encoder brings better performance on expressiveness than the conventional BASE system. It indicates that using self-attention layers as text encoder may capture features that better represents expressiveness, in accordance with our previous \ufb01ndings [12]. For the proposed context extractor, we \ufb01nd that, by introducing direct aggregation across all the selfattention layers, system SA-DA achieves signi\ufb01cantly better performance than the solely self-attention based encoder system SA. This is con\ufb01rmed by both AB preference and MOS test on the two tested corpora. By further replacing simple concatenation operation with multi-head attention aggregation (i.e., weighted aggregation), system SA-WA brings extra performance gain over system SA-DA. Listeners particularly give the SA-WA system more preferences according to the AB preference test. In summary, the results unveil that the deep sentential context encoder achieves signi\ufb01cantly better performance than the baseline systems, showing that modeling different levels of latent syntax and semantic information through a deep encoder is effective for generating expressive speech. This conclusion is consistently con\ufb01rmed on two expressive corpora. 3.5. Performance on less-expressive corpus We also quickly examine the performance of our approach on a less-expressive reading-style corpus \u2013 DB1 [22], to see how our sentential context extractor perform. Here, we only compare the above best-performed SA-WA system with the BASE system. The MCD scores for BASE and SA-WA are 5.78 and 5.72, respectively. The AB preference is illustrated in Figure 4. In1Samples can be found from: https://fyyang1996.github.io/context/ \f52.7% 19.3% 28.0% 41.9% 33.8% 57.1% 21.7% 84.7% 11.3% BASE BASE SA SA SA-DA SA-DA SA-WA SA-WA No Preference No Preference No Preference No Preference 4.0% 21.2% 24.3% Figure 3: AB Preference results on VA with con\ufb01dence intervals of 95% and p-value < 0.0001 from t-test. 33.1% 27.6% BASE SA-WA No Preference 39.3% Figure 4: AB Preference results on DB1 with con\ufb01dence intervals of 95% and p-value < 0.0001 from t-test. terestingly, the effectiveness of our sentential context extractor is not salient on this less-expressive corpus, which is proved by close MCD and AB preference scores between the two systems. In other words, our sentential context extractor works better on expressive datasets, which is further con\ufb01rmed in the following. 3.6. Analysis Prosody Correlation To further evaluate the expressiveness for statistical signi\ufb01cance, we extract the acoustic features commonly associated with prosody: relative energy within each phoneme (E), duration in ms (Dur.) and fundamental frequency in Hertz (F0), which represent phoneme-level intensity, rhythm and intonation of audio, respectively. According to [23], we measure the three prosody attributes for each phoneme throughout additional alignments. The ratio of the average signal magnitude within a phoneme with the average magnitude of the entire utterance is used as the relative energy of a phoneme. We calculate the number of frames within a phoneme as the duration of the phoneme. And the mean value of F0 within a phoneme is regarded as a prosody attribute. To estimate these statistics, we synthesized 100 random samples in the test set to calculate the Pearson correlation coef\ufb01cient between all systems and the ground truth. The higher Pearson correlation coef\ufb01cient value the model produces, the higher accuracy of the predicted prosody attribute the model can achieve. Table 3 shows that our proposed SA-WA system obtains highest correlation scores in all three prosody attributes, which demonstrates that our approach has the best reconstruction performance in phoneme-level intensity, rhythm and intonation. Additionally, [24] reveals that the order of prosody attributes being captured is always found to be energy, duration, and F0. Energy is the amplitude of the signal directly related to the reconstruction loss and is easier to be captured, but F0 is the most dif\ufb01cult to be captured as it is modeled implicitly. However, our SA-WA system achieves approximately 20% gains than the BASE system in F0, which is far more than the promotion of approximately 6% in energy and duration. Based on these, we believe that the proposed approach has strong ability in modeling F0 that is most dif\ufb01cult one to be captured in the three prosody attributes. Figure 5 shows the F0 trajectories for a synthesized test utterance. The sentence begins with a modal particle (heng1), which reveals sense of disgust emotion in Mandarin. In this case, high raise of F0, where the SA-WA system produces, can deliver more disgust mood to listeners. This example shows that the proposed sequential context extractor can model better expressive patterns compared to the baseline systems. Table 3: Correlation in relative energy, duration and F0 within a phoneme computed from different models on TS. BASE SA SA-DA SA-WA E 0.755 0.776 0.781 0.799 Dur. 0.617 0.638 0.641 0.654 F0 0.42 0.426 0.437 0.501 BASE SA SA-DA SA-WA \u54fc\uff0c\u60f3\u4f60\u6709\u4ec0\u4e48\u7528\uff0c\u4f60\u53c8\u4e0d\u6765\u966a\u6211\u73a9\u3002 Figure 5: F0 values of a test utterance generated by different systems. Audios can be found in Section 1.1 of the demo page. Prosody Diversity An expressive TTS system should be able to generate speech with a large prosody diversity. Consequently, we measure the standard deviation of three prosody attributes at phoneme level as well according to [23]. And the average standard deviation across all 100 utterances is reported for statistical signi\ufb01cance. Table 4 demonstrate that the SAWA system has the highest diversity in phoneme-level intensity, rhythm and intonation among all systems, which is the closest to the ground-truth (GT). We believe that the SA-WA system has better ability in modeling prosody variations on expressive datasets. Table 4: Diversity values using average standard deviation computed across 100 samples on TS. BASE SA SA-DA SA-WA GT E 0.238 0.277 0.285 0.304 0.321 Dur. 33.374 34.337 34.955 37.003 41.866 F0 32.362 33.405 35.161 35.766 36.824 4."
+                }
+            ],
+            "Andrew Owens": [
+                {
+                    "url": "http://arxiv.org/abs/1512.08512v2",
+                    "title": "Visually Indicated Sounds",
+                    "abstract": "Objects make distinctive sounds when they are hit or scratched. These sounds\nreveal aspects of an object's material properties, as well as the actions that\nproduced them. In this paper, we propose the task of predicting what sound an\nobject makes when struck as a way of studying physical interactions within a\nvisual scene. We present an algorithm that synthesizes sound from silent videos\nof people hitting and scratching objects with a drumstick. This algorithm uses\na recurrent neural network to predict sound features from videos and then\nproduces a waveform from these features with an example-based synthesis\nprocedure. We show that the sounds predicted by our model are realistic enough\nto fool participants in a \"real or fake\" psychophysical experiment, and that\nthey convey significant information about material properties and physical\ninteractions.",
+                    "authors": "Andrew Owens, Phillip Isola, Josh McDermott, Antonio Torralba, Edward H. Adelson, William T. Freeman",
+                    "published": "2015-12-28",
+                    "updated": "2016-04-30",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV",
+                        "cs.LG",
+                        "cs.SD"
+                    ],
+                    "main_content": "Introduction From the clink of a ceramic mug placed onto a saucer, to the squelch of a shoe pressed into mud, our days are \ufb01lled with visual experiences accompanied by predictable sounds. On many occasions, these sounds are not just statistically associated with the content of the images \u2013 the way, for example, that the sounds of unseen seagulls are associated with a view of a beach \u2013 but instead are directly caused by the physical interaction being depicted: you see what is making the sound. We call these events visually indicated sounds, and we propose the task of predicting sound from videos as a way to study physical interactions within a visual scene (Figure 1). To accurately predict a video\u2019s held-out soundtrack, an algorithm has to know about the physical properties of what it is seeing and the actions that are being performed. This task implicitly requires material recognition, but unlike traditional work on this problem [4, 38], we never explicitly tell the algorithm about materials. Instead, it learns about them by identifying statistical regularities in the raw audiovisual signal. We take inspiration from the way infants explore the physical properties of a scene by poking and prodding at the objects in front of them [36, 3], a process that may help them learn an intuitive theory of physics [3]. Recent work suggests that the sounds objects make in response to these interactions may play a role in this process [39, 43]. We introduce a dataset that mimics this exploration process, containing hundreds of videos of people hitting, scratching, and prodding objects with a drumstick. To synthesize sound from these videos, we present an algorithm that uses a recurrent neural network to map videos to audio features. It then converts these audio features to a waveform, either by matching them to exemplars in a database 1 arXiv:1512.08512v2  [cs.CV]  30 Apr 2016 \fGlass Dirt Scattering Deformation Splash Materials Actions Reactions Wood Plastic bag Plastic Rock Paper Gravel Grass Leaf Metal Ceramic Cloth Carpet Water hit scratch other 0 0.2 0.4 0.6 0.8 deform splash static rigid-motion scatter other 0 0.2 0.4 0.6 0.8 Figure 2: Greatest Hits: Volume 1 dataset. What do these materials sound like when they are struck? We collected 977 videos in which people explore a scene by hitting and scratching materials with a drumstick, comprising 46,577 total actions. Human annotators labeled the actions with material category labels, the location of impact, an action type label (hit vs. scratch), and a reaction label (shown on right). These labels were used only for analyzing what our sound prediction model learned, not for training it. We show images from a selection of videos from our dataset for a subset of the material categories (here we show examples where it is easy to see the material in question). and transferring their corresponding sounds, or by parametrically inverting the features. We evaluate the quality of our predicted sounds using a psychophysical study, and we also analyze what our method learned about actions and materials through the task of learning to predict sound. 2. Related work Our work closely relates to research in sound and material perception, and to representation learning. Foley The idea of adding sound effects to silent movies goes back at least to the 1920s, when Jack Foley and collaborators discovered that they could create convincing sound effects by crumpling paper, snapping lettuce, and shaking cellophane in their studio1, a method now known as Foley. Our algorithm performs a kind of automatic Foley, synthesizing plausible sound effects without a human in the loop. Sound and materials In the classic mathematical work of [26], Kac showed that the shape of a drum could be partially recovered from the sound it makes. Material properties, such as stiffness and density [37, 31, 14], can likewise be determined from impact sounds. Recent work has used these principles to estimate material properties by measuring tiny vibrations in rods and cloth [8], and similar methods have been used to recover sound from high-speed video of a vibrating membrane [9]. Rather than using a camera as an instrument for measuring vibrations, we infer a plausible sound for an action by recognizing what kind of sound this action would normally make in the visually observed scene. Impact sounds have been used in other work to recognize objects and materials. Arnab et al. [2] recently presented a semantic segmentation model that incorporates audio from impact sounds, and showed that audio information could 1To our delight, Foley artists really do knock two coconuts together to fake the sound of horses galloping [6]. help recognize objects and materials that were ambiguous from visual cues alone. Other work recognizes objects using audio produced by robotic interaction [41, 29]. Sound synthesis Our technical approach resembles speech synthesis methods that use neural networks to predict sound features from pre-tokenized text features and then generate a waveform from those features [30]. There are also methods, such as the FoleyAutomatic system, for synthesizing impact sounds from physical simulations [45]. Work in psychology has studied low-dimensional representations for impact sounds [7], and recent work in neuroimaging has shown that silent videos of impact events activate the auditory cortex [19]. Learning visual representations from natural signals Previous work has explored the idea of learning visual representations by predicting one aspect of a raw sensory signal from another. For example, [11, 22] learned image representations by predicting the spatial relationship between image patches, and [1, 23] by predicting the egocentric motion between video frames. Several methods have also used temporal proximity as a supervisory signal [33, 17, 47, 46]. Unlike in these approaches, we learn to predict one sensory modality (sound) from another (vision). There has also been work that trains neural networks from multiple modalities. For example, [34] learned a joint model of audio and video. However, while they study speech using an autoencoder, we focus on material interaction, and we use a recurrent neural network to predict sound features from video. A central goal of other methods has been to use a proxy signal (e.g., temporal proximity) to learn a generically useful representation of the world. In our case, we predict a signal \u2013 sound \u2013 known to be a useful representation for many tasks [14, 37], and we show that the output (i.e. the predicted sound itself, rather than some internal representation \fin the model) is predictive of material and action classes. 3. The Greatest Hits dataset In order to study visually indicated sounds, we collected a dataset containing videos of humans (the authors) probing environments with a drumstick \u2013 hitting, scratching, and poking different objects in the scene (Figure 2). We chose to use a drumstick so that we would have a consistent way of generating the sounds. Moreover, since the drumstick does not occlude much of a scene, we can also observe what happens to the object after it is struck. This motion, which we call a reaction, can be important for inferring material properties \u2013 a soft cushion, for example, will deform more than a \ufb01rm one, and the sound it produces will vary with it. Similarly, individual pieces of gravel will scatter when they are hit, and their sound varies with this motion (Figure 2, right). Unlike traditional objector scene-centric datasets, such as ImageNet [10] or Places [48], where the focus of the image is a full scene, our dataset contains close-up views of a small number of objects. These images re\ufb02ect the viewpoint of an observer who is focused on the interaction taking place (similar to an egocentric viewpoint). They contain enough detail to see \ufb01ne-grained texture and the reaction that occurs after the interaction. In some cases, only part of an object is visible, and neither its identity nor other high-level aspects of the scene are easily discernible. Our dataset is also related to robotic manipulation datasets [41, 35, 15]. While one advantage of using a robot is that its actions are highly consistent, having a human collect the data allows us to rapidly (and inexpensively) capture a large number of physical interactions in real-world scenes. We captured 977 videos from indoor (64%) and outdoor scenes (36%). The outdoor scenes often contain materials that scatter and deform, such as grass and leaves, while the indoor scenes contain a variety of hard and soft materials, such as metal, plastic, cloth, and plastic bags. Each video, on average, contains 48 actions (approximately 69% hits and 31% scratches) and lasts 35 seconds. We recorded sound using a shotgun microphone attached to the top of the camera and used a wind cover for outdoor scenes. We used a separate audio recorder, without auto-gain, and we applied a denoising algorithm [20] to each recording. Detecting impact onsets We detected amplitude peaks in the denoised audio, which largely correspond to the onset of impact sounds. We thresholded the amplitude gradient to \ufb01nd an initial set of peaks, then merged nearby peaks with the mean-shift algorithm [13], treating the amplitude as a density and \ufb01nding the nearest mode for each peak. Finally, we used non-maximal suppression to ensure that onsets were at least 0.25 seconds apart. This is a simple onsetdetection method that most often corresponds to drumstick impacts when the impacts are short and contain a single Deformation Scattering Rock Cloth Wood Dirt Time Frequency 0.25 0.00 0.50 0.2 0.0 (a) Mean cochleagrams (b) Sound confusion matrix Figure 3: (a) Cochleagrams for selected classes. We extracted audio centered on each impact sound in the dataset, computed our subband-envelope representation, and then estimated the mean for each class. (b) Confusion matrix derived by classifying sound features. Rows correspond to confusions made for a single category. The row ordering was determined automatically, by similarity in material confusions (see Section A1.2). peak2. In many of our experiments, we use short video clips that are centered on these amplitude peaks. Semantic annotations We also collected annotations for a sample of impacts (approximately 62%) using online workers from Amazon Mechanical Turk. These include material labels, action labels (hit vs. scratch), reaction labels, and the pixel location of each impact site. To reduce the number of erroneous labels, we manually removed annotations for material categories that we could not \ufb01nd in the scene. During material labeling, workers chose from \ufb01nergrained categories. We then merged similar, frequently confused categories (please see Section A2 for details). Note that these annotations are used only for analysis: we train our models on raw audio and video. Examples of several material and action classes are shown in Figure 2. 4. Sound representation Following work in sound synthesis [42, 32], we compute our sound features by decomposing the waveform into subband envelopes \u2013 a simple representation obtained by \ufb01ltering the waveform and applying a nonlinearity. We apply a bank of 40 band-pass \ufb01lters spaced on an equivalent rectangular bandwidth (ERB) scale [16] (plus a lowand high-pass \ufb01lter) and take the Hilbert envelope of the responses. We then downsample these envelopes to 90Hz (approximately 3 samples per frame) and compress them. More speci\ufb01cally, we compute envelope sn(t) from a wave2Scratches and hits usually satisfy this assumption, but splash sounds often do not \u2013 a problem that could be addressed with more sophisticated onset-detection methods [5]. \fform w(t) and a \ufb01lter fn by taking: sn = D(|(w \u2217fn) + jH(w \u2217fn)|)c, (1) where H is the Hilbert transform, D denotes downsampling, and the compression constant c = 0.3. In Section A1.2, we study how performance varies with the number of frequency channels. The resulting representation is known as a cochleagram. In Figure 3(a), we visualize the mean cochleagram for a selection of material and reaction classes. This reveals, for example, that cloth sounds tend to have more low-frequency energy than those of rock. How well do impact sounds capture material properties in general? To measure this empirically, we trained a linear SVM to predict material class for the sounds in our database, using the subband envelopes as our feature vectors. We resampled our training set so that each class contained an equal number of impacts (260 per class). The resulting material classi\ufb01er has 45.8% (chance = 5.9%) classaveraged accuracy (i.e., the mean of per-class accuracy values), and its confusion matrix is shown in Figure 3(b). These results suggest that impact sounds convey signi\ufb01cant information about materials, and thus if an algorithm could learn to accurately predict these sounds from images, it would have implicit knowledge of material categories. 5. Predicting visually indicated sounds We formulate our task as a regression problem \u2013 one where the goal is to map a sequence of video frames to a sequence of audio features. We solve this problem using a recurrent neural network that takes color and motion information as input and predicts the subband envelopes of an audio waveform. Finally, we generate a waveform from these sound features. Our neural network and synthesis procedure are shown in Figure 4. 5.1. Regressing sound features Given a sequence of input images I1, I2, ..., IN, we would like to estimate a corresponding sequence of sound features \u20d7 s1,\u20d7 s2, ...,\u20d7 sT , where \u20d7 st \u2208R42. These sound features correspond to blocks of the cochleagram shown in Figure 4. We solve this regression problem using a recurrent neural network (RNN) that takes image features computed with a convolutional neural network (CNN) as input. Image representation We found it helpful to represent motion information explicitly in our model using a twostream approach [12, 40]. While two-stream models often use optical \ufb02ow, it is challenging to obtain accurate \ufb02ow estimates due to the presence of fast, non-rigid motion. Instead, we compute spacetime images for each frame \u2013 images whose three channels are grayscale versions of the previous, current, and next frames. This image representation \u2026 \u2026 Video CNN LSTM \u21e2 ( Cochleagram Time Waveform Example-based  synthesis Figure 4: We train a neural network to map video sequences to sound features. These sound features are subsequently converted into a waveform using either parametric or example-based synthesis. We represent the images using a convolutional network, and the time series using a recurrent neural network. We show a subsequence of images corresponding to one impact. is closely related to 3D video CNNs [24, 27], as derivatives across channels correspond to temporal derivatives. For each frame t, we construct an input feature vector xt by concatenating CNN features for the spacetime image at frame t and the color image from the \ufb01rst frame3: xt = [\u03c6(Ft), \u03c6(I1)], (2) where \u03c6 are CNN features obtained from layer fc7 of the AlexNet architecture [28] (its penultimate layer), and Ft is the spacetime image at time t. In our experiments (Section 6), we either initialized the CNN from scratch and trained it jointly with the RNN, or we initialized the CNN with weights from a network trained for ImageNet classi\ufb01cation. When we used pretraining, we precomputed the features from the convolutional layers and \ufb01ne-tuned only the fully connected layers. Sound prediction model We use a recurrent neural network (RNN) with long short-term memory units (LSTM) [18] that takes CNN features as input. To compensate for the difference between the video and audio sampling rates, we replicate each CNN feature vector k times, where k = \u230aT/N\u230b(we use k = 3). This results in a sequence of CNN features x1, x2, ..., xT that is the same length as the sequence of audio features. At each timestep of the RNN, we use the current image feature vector xt to update the 3We use only the \ufb01rst color image to reduce the computational cost. \fvector of hidden variables ht4. We then compute sound features by an af\ufb01ne transformation of the hidden variables: \u20d7 st = Wht + b ht = L(xt, ht\u22121), (3) where L is a function that updates the hidden state [18]. During training, we minimize the difference between the predicted and ground-truth predictions at each timestep: E({\u20d7 st}) = T X t=1 \u03c1(\u2225\u20d7 st \u2212\u02dc \u20d7 st\u22252), (4) where \u02dc \u20d7 st and \u20d7 st are the true and predicted sound features at time t, and \u03c1(r) = log(\u03f5 + r2) is a robust loss that bounds the error at each timestep (we use \u03f5 = 1/252). We also increase robustness of the loss by predicting the square root of the subband envelopes, rather than the envelope values themselves. To make the learning problem easier, we use PCA to project the 42-dimensional feature vector at each timestep down to a 10-dimensional space, and we predict this lower-dimensional vector. When we evaluate the network, we invert the PCA transformation to obtain sound features. We train the RNN and CNN jointly using stochastic gradient descent with Caffe [25, 12]. We found it helpful for convergence to remove dropout [44] and to clip large gradients. When training from scratch, we augmented the data by applying cropping and mirroring transformations to the videos. We also use multiple LSTM layers (the number depends on the task; please see Section A1.1). 5.2. Generating a waveform We consider two methods for generating a waveform from the predicted sound features. The \ufb01rst is the simple parametric synthesis approach of [42, 32], which iteratively imposes the subband envelopes on a sample of white noise (we used just one iteration). This method is useful for examining what information is captured by the audio features, since it represents a fairly direct conversion from features to sound. However, for the task of generating plausible sounds to a human ear, we \ufb01nd it more effective to impose a strong natural sound prior during conversion from features to waveform. Therefore, we also consider an examplebased synthesis method that snaps a window of sound features to the closest exemplar in the training set. We form a query vector by concatenating the predicted sound features \u20d7 s1, ...,\u20d7 sT (or a subsequence of them), searching for its nearest neighbor in the training set as measured by L1 distance, and transferring the corresponding waveform. 6. Experiments We applied our sound-prediction model to several tasks, and evaluated it with a combination of human studies and automated metrics. 4To simplify the presentation, we have omitted the LSTM\u2019s hidden cell state, which is also updated at each timestep. 6.1. Sound prediction tasks In order to study the problem of detection \u2013 that is, the task of determining when and whether an action that produces a sound has occurred \u2013 separately from the task of sound prediction, we consider two kinds of videos. First, we focus on the prediction problem and consider only videos centered on audio amplitude peaks, which often correspond to impact onsets (Section 3). We train our model to predict sound for 15-frame sequences (0.5 sec.) around each peak. For the second task, which we call the detection problem, we train our model on longer sequences (approximately 2 sec. long) sampled from the training videos with a 0.5second stride, and we subsequently evaluate this model on full-length videos. Since it can be dif\ufb01cult to discern the precise timing of an impact, we allow the predicted features to undergo small shifts before they are compared to the ground truth. We also introduce a two-frame lag in the RNN output, which allows the model to observe future frames before outputting sound features. Finally, before querying sound features, we apply a coloring procedure to account for statistical differences between the predicted and real sound features (e.g., under-prediction of amplitude), using the silent videos in the test set to estimate the empirical mean and covariance of the network\u2019s predictions. For these implementation details, please see Section A1.1. For both tasks, we split the full-length videos into training and test sets (75% training and 25% testing). Models For the prediction task, we compared our model to image-based nearest neighbor search. We computed fc7 features from a CNN pretrained on ImageNet [28] for the center frame of each sequence, which by construction is the frame where the impact sound occurs. We then searched the training set for the best match and transferred its corresponding sound. We considered variations where the CNN features were computed on an RGB image, on (three-frame) spacetime images, and on the concatenation of both features. To understand the in\ufb02uence of different design decisions, we also considered several variations of our model. We included models with and without ImageNet pretraining; with and without spacetime images; and with examplebased versus parametric waveform generation. Finally, we included a model where the RNN connections were broken (the hidden state was set to zero between timesteps). For the RNN models that do example-based waveform generation (Section 5.2), we used the centered impacts in the training set as the exemplar database. For the prediction task, we performed the query using the sound features for the entire 15-frame sequence. For the detection task, this is not possible, since the videos may contain multiple, overlapping impacts. Instead, we detected amplitude peaks of the parametrically inverted waveform, and matched the sound features in small (8-frame) windows around each \fPsychophysical study Loudness Centroid Algorithm Labeled real Err. r Err. r Full system 40.01% \u00b1 1.66 0.21 0.44 3.85 0.47 Trained from scratch 36.46% \u00b1 1.68 0.24 0.36 4.73 0.33 No spacetime 37.88% \u00b1 1.67 0.22 0.37 4.30 0.37 Parametric synthesis 34.66% \u00b1 1.62 0.21 0.44 3.85 0.47 No RNN 29.96% \u00b1 1.55 1.24 0.04 7.92 0.28 Image match 32.98% \u00b1 1.59 0.37 0.16 8.39 0.18 Spacetime match 31.92% \u00b1 1.56 0.41 0.14 7.19 0.21 Image + spacetime 33.77% \u00b1 1.58 0.37 0.18 7.74 0.20 Random impact sound 19.77% \u00b1 1.34 0.44 0.00 9.32 0.02 0.25 0.00 0.50 (a) Model evaluation (b) Predicted sound confusions (c) CNN feature confusions Figure 5: (a) We measured the rate at which subjects chose an algorithm\u2019s synthesized sound over the actual sound. Our full system, which was pretrained from ImageNet and used example-based synthesis to generate a waveform, signi\ufb01cantly outperformed models based on image matching. For the neural network models, we computed the auditory metrics for the sound features that were predicted by the network, rather than those of the inverted sounds or transferred exemplars. (b) What sounds like what, according to our algorithm? We applied a classi\ufb01er trained on real sounds to the sounds produced by our algorithm, resulting in a confusion matrix (c.f. Fig. 3(b), which shows a confusion matrix for real sounds). It obtained 22.7% class-averaged accuracy. (c) Confusions made by a classi\ufb01er trained on fc7 features (30.2% class-averaged accuracy). For both confusion matrices, we used the variation of our model that was trained from scratch (see Fig. A1(b) for the sound confusions made with pretraining). peak (starting the window one frame before the peak). 6.2. Evaluating the sound predictions We assessed the quality of the sounds using psychophysical experiments and measurements of acoustic properties. Psychophysical study To test whether the sounds produced by our model varied appropriately with different actions and materials, we conducted a psychophysical study on Amazon Mechanical Turk. We used a two-alternative forced choice test in which participants were asked to distinguish real and fake sounds. We showed them two videos of an impact event \u2013 one playing the recorded sound, the other playing a synthesized sound. We then asked them to choose the one that played the real sound. The sound-prediction algorithm was chosen randomly on a per-video basis. We randomly sampled 15 impact-centered sequences from each full-length video, showing each participant at most one impact from each one. At the start of the experiment, we revealed the correct answer to \ufb01ve practice videos. We measured the rate at which participants mistook our model\u2019s result for the ground-truth sound (Figure 5(a)), \ufb01nding that our full system \u2013 with RGB and spacetime input, RNN connections, ImageNet pretraining, and examplebased waveform generation \u2013 signi\ufb01cantly outperformed the image-matching methods. It also outperformed a baseline that sampled a random (centered) sound from the training set (p < 0.001 with a two-sided t-test). We found that the version of our model that was trained from scratch outperformed the best image-matching method (p = 0.02). Finally, for this task, we did not \ufb01nd the difference between our full and RGB-only models to be signi\ufb01cant (p = 0.08). We show results broken down by semantic category in Algorithm Labeled real Full sys. + mat. 41.82% \u00b1 1.46 Full sys. 39.64% \u00b1 1.46 fc7 NN + mat. 38.20% \u00b1 1.47 fc7 NN 32.83% \u00b1 1.41 Random + mat. 35.36% \u00b1 1.42 Random 20.64% \u00b1 1.22 Real sound match 46.90% \u00b1 1.49 Features Avg. Acc. Audio-supervised CNN 30.4% ImageNet CNN 42.0% Sound 45.8% ImageNet + sound 48.2% ImageNet crop 52.9% Crop + sound 59.4% Figure 6: (a) We ran variations of the full system and an imagematching method (RGB + spacetime). For each model, we include an oracle model (labeled with \u201c+ mat\u201d) that draws its sound examples from videos with the same material label. (b) Class-averaged material recognition accuracy obtained by training an SVM with different image and sound features. Figure 7. For some categories, such as grass and leaf, participants were frequently fooled by our results. Often when a participant was fooled, it was because the sound prediction was simple and prototypical (e.g., a simple thud noise), while the actual sound was complex and atypical. True leaf sounds, for example, are highly varied and may not be fully predictable from a silent video. When they are struck, we hear a combination of the leaves themselves, along with rocks, dirt, and whatever else is underneath them. In contrast, the sounds predicted by our model tend to be closer to prototypical grass/dirt/leaf noises. Participants also sometimes made mistakes when the onset detection failed, or when multiple impacts overlapped, since this may have de\ufb01ed their expectation of hearing a single impact. We found that the model in which the RNN connections were broken was often unable to detect the timing of the hit, and that it under-predicted the amplitude of the sounds. As a \fscatter rigid-motion static splash deform Mean accuracy 0 20 40 60 80 100 hit scratch Mean accuracy 0 20 40 60 80 100 drywall glass leaf dirt tile grass gravel wood carpet rock metal paper plastic water cloth ceramic plastic bag Mean accuracy 0 20 40 60 80 100 Material Reaction % synthesized labeled as real Ours Image+spacetime match Action 20 0 40 60 80 100 Figure 7: Semantic analysis of psychophysical study. We show the rate that our algorithm fooled human participants for each material, action, and reaction class. The error bars show 95% con\ufb01dence intervals. Our approach signi\ufb01cantly outperforms the highest-performing image-matching method (RGB + spacetime). result, it performed poorly on automated metrics and failed to \ufb01nd good matches. The performance of our model with parametric waveform generation varied widely between categories. It did well on materials such as leaf and dirt that are suited to the relatively noisy sounds that the method produces but poorly on hard materials such as wood and metal (e.g., a confusion rate of 62% \u00b1 6% for dirt and 18% \u00b1 5% for metal). On the other hand, the example-based approach was not effective at matching textural sounds, such as those produced by splashing water (Fig. 7). Auditory metrics We measured quantitative properties of sounds for the prediction task. We chose metrics that were not sensitive to precise timing. First, we measured the loudness of the sound, which we took to be the maximum energy (L2 norm) of the compressed subband envelopes over all timesteps. Second, we compared the sounds\u2019 spectral centroids, which we measured by taking the center of mass of the frequency channels for a one-frame (approx. 0.03 sec.) window around the center of the impact. We found that on both metrics, the network was more accurate than the image-matching methods, both in terms of mean squared error and correlation coef\ufb01cients (Figure 5(a)). Oracle results How helpful is material category information? We conducted a second study that controlled for material-recognition accuracy. Using the subset of the data with material annotations, we created a model that chose a random sound from the same class as the input video. We also created a number of oracle models that used these material labels (Table 6(a)). For the best-performing imagematching model (RGB + spacetime), we restricted the pool of matches to be those with the same label as the input (and similarly for the example-based synthesis method). We also considered a model that matched the ground-truth sound to the training set and transferred the best match. We found that, while knowing the material was helpful for each method, it was not suf\ufb01cient, as the oracle models did not outperform our model. In particular, our model signi\ufb01cantly outperformed the random-sampling oracle (p < 10\u22124). Impact detection We also used our methods to produce sounds for long, uncentered videos, a problem setting that allows us to evaluate their ability to detect impact events. We provide qualitative examples in Figure 8 and on our webpage (vis.csail.mit.edu). To quantitatively evaluate its detection accuracy, we used the parametric synthesis method to produce a waveform, applied a large gain to that waveform, and then detected amplitude peaks (Section 3). We then compared the timing of these peaks to those of the ground truth, considering an impact to be detected if a predicted spike occurred within 0.1 seconds of it. Using the predicted amplitude as a measure of con\ufb01dence, we computed average precision. We compared our model to an RGB-only model, \ufb01nding that the spacetime images significantly improve the result, with APs of 43.6% and 21.6% respectively. Both models were pretrained with ImageNet. 6.3. Learning about material and action by predicting sounds By learning to predict sounds, did the network also learn something about material and physical interactions? To assess this, we tested whether the network\u2019s output sounds were informative about material and action class. We applied the same SVM that was trained to predict material/action class on real sound features (Sec. 4) to the sounds predicted by the model. Under this evaluation regime, it is not enough for the network\u2019s sounds to merely be distinguishable by class: they must be close enough to real sounds so as to be classi\ufb01ed correctly by an SVM that has never seen a predicted sound. To avoid the in\ufb02uence of pretraining, we used a network that was trained from scratch. We note that this evaluation method is different from that of recent unsupervised learning models [11, 1, 47] that train a classi\ufb01er on the network\u2019s feature activations, rather than on a ground-truth version of the output. Using this idea, we classi\ufb01ed the material category from predicted sound features. The classi\ufb01er had class-averaged accuracy of 22.7%, and its confusion matrix is shown in Fig. 5(b). This accuracy indicates that our model learned an output representation that was informative about material, even though it was only trained to predict sound. We applied a similar methodology to classify action categories from predicted sounds, obtaining 68.6% class-averaged accuracy (chance = 50%), and 53.5% for classifying reaction categories (chance = 20%). We found that material and reaction recognition accuracy improved with ImageNet pretraining (to 28.8% and to 55.2%, respectively), but that there was a slight decrease for action classi\ufb01cation (to 66.5%). We also tested whether the predicted sound features convey information about the hardness of a surface. We grouped the material classes into superordinate hard and soft classes, and trained a classi\ufb01er on real sound features (sampling 1300 examples per class), \ufb01nding that it obtained \fFrame from input video Real vs. synthesized cochleagram Frame from input video Real vs. synthesized cochleagram Synthesized Real Synthesized Real Synthesized Real Synthesized Real 0.5 0 0.5 0 0.5 0 0.5 0 Time (seconds) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Frequency subband 5 10 15 20 25 30 35 40 Time (seconds) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Frequency subband 5 10 15 20 25 30 35 40 Time (seconds) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Frequency subband 5 10 15 20 25 30 35 40 Time (seconds) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Frequency subband 5 10 15 20 25 30 35 40 Time (seconds) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Frequency subband 5 10 15 20 25 30 35 40 Time (seconds) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Frequency subband 5 10 15 20 25 30 35 40 Time (seconds) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Frequency subband 5 10 15 20 25 30 35 40 Time (seconds) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 Frequency subband 5 10 15 20 25 30 35 40 Figure 8: Automatic sound prediction results. We show cochleagrams for a representative selection of video sequences, with a sample frame from each sequence on the left. The frame is sampled from the location indicated by the black triangle on the x-axis of each cochleagram. Notice that the algorithm\u2019s synthesized cochleagrams match the general structure of the ground truth cochleagrams. Dark lines in the cochleagrams indicate hits, which the algorithm often detects. The algorithm captures aspects of both the temporal and spectral structure of sounds. It correctly predicts staccato taps in rock example and longer waveforms for rustling ivy. Furthermore, it tends to predict lower pitched thuds for a soft couch and higher pitched clicks when the drumstick hits a hard wooden railing (although the spectral differences may appear small in these visualizations, we evaluate this with objective metrics in Section 6). A common failure mode is that the algorithm misses a hit (railing example) or hallucinates false hits (cushion example). This frequently happens when the drumstick moves erratically. Please see our video for qualitative results. 66.8% class-averaged accuracy (chance = 50%). Here we have de\ufb01ned soft materials to be {leaf, grass, cloth, plastic bag, carpet} and hard materials to be {gravel, rock, tile, wood, ceramic, plastic, drywall, glass, metal}. We also considered the problem of directly predicting material class from visual features. In Table 6(b), we trained a classi\ufb01er using fc7 features \u2013 both those of the model trained from scratch, and of a model trained on ImageNet [28]. We concatenated color and spacetime image features, since we found that this improved performance. We also considered an oracle model that cropped a high-resolution (256 \u00d7 256) patch from the impact location using human annotations, and concatenated its features with those of the full image (we used color images). To avoid occlusions from the arm or drumstick, we cropped the patch from the \ufb01nal frame of the video. We found that performing these crops signi\ufb01cantly increased the accuracy, suggesting that localizing the impact is important for classi\ufb01cation. We also tried concatenating vision and sound features (similar to [2]), \ufb01nding that this signi\ufb01cantly improved the accuracy. The kinds of mistakes that the visual classi\ufb01er (video \u2192 material) made were often different from those of the sound classi\ufb01er (sound \u2192material). For instance, the visual classi\ufb01er was able to distinguish classes that have a very different appearance, such as paper and cloth. These classes both make low-pitched sounds (e.g., cardboard and cushions), and were sometimes are confused by the sound classi\ufb01er. On the other hand, the visual classi\ufb01er was more likely to confuse materials from outdoor scenes, such as rocks and leaves \u2013 materials that sound very different but which frequently co-occur in a scene. When we analyzed our model by classifying its sound predictions (video \u2192sound \u2192material), the resulting confusion matrix (Fig. 5(b)) contains both kinds of error: there are visual analysis errors when it misidenti\ufb01es the material that was struck, and sound synthesis errors when it produces a sound that was not a convincing replica of the real sound. 7. Discussion In this work, we proposed the problem of synthesizing visually indicated sounds \u2013 a problem that requires an algorithm to learn about material properties and physical interactions. We introduced a dataset for studying this task, which contains videos of a person probing materials in the world with a drumstick, and an algorithm based on recurrent neural networks. We evaluated the quality of our approach with psychophysical experiments and automated metrics, showing that the performance of our algorithm was signi\ufb01cantly better than baselines. We see our work as opening two possible directions for future research. The \ufb01rst is producing realistic sounds from videos, treating sound production as an end in itself. The second direction is to use sound and material interactions as steps toward physical scene understanding. Acknowledgments. This work was supported by NSF grants 6924450 and 6926677, by Shell, and by a Microsoft Ph.D. Fellowship to A.O. We thank Carl Vondrick and Rui Li for the helpful discussions, and the workers at Middlesex Fells, Arnold Arboretum, and Mt. Auburn Cemetery for not asking too many questions while we were collecting the Greatest Hits dataset."
+                }
+            ],
+            "Jiacheng Zhang": [
+                {
+                    "url": "http://arxiv.org/abs/2307.08095v1",
+                    "title": "Semi-DETR: Semi-Supervised Object Detection with Detection Transformers",
+                    "abstract": "We analyze the DETR-based framework on semi-supervised object detection\n(SSOD) and observe that (1) the one-to-one assignment strategy generates\nincorrect matching when the pseudo ground-truth bounding box is inaccurate,\nleading to training inefficiency; (2) DETR-based detectors lack deterministic\ncorrespondence between the input query and its prediction output, which hinders\nthe applicability of the consistency-based regularization widely used in\ncurrent SSOD methods. We present Semi-DETR, the first transformer-based\nend-to-end semi-supervised object detector, to tackle these problems.\nSpecifically, we propose a Stage-wise Hybrid Matching strategy that combines\nthe one-to-many assignment and one-to-one assignment strategies to improve the\ntraining efficiency of the first stage and thus provide high-quality pseudo\nlabels for the training of the second stage. Besides, we introduce a Crossview\nQuery Consistency method to learn the semantic feature invariance of object\nqueries from different views while avoiding the need to find deterministic\nquery correspondence. Furthermore, we propose a Cost-based Pseudo Label Mining\nmodule to dynamically mine more pseudo boxes based on the matching cost of\npseudo ground truth bounding boxes for consistency training. Extensive\nexperiments on all SSOD settings of both COCO and Pascal VOC benchmark datasets\nshow that our Semi-DETR method outperforms all state-of-the-art methods by\nclear margins. The PaddlePaddle version code1 is at\nhttps://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/semi_det/semi_detr.",
+                    "authors": "Jiacheng Zhang, Xiangru Lin, Wei Zhang, Kuo Wang, Xiao Tan, Junyu Han, Errui Ding, Jingdong Wang, Guanbin Li",
+                    "published": "2023-07-16",
+                    "updated": "2023-07-16",
+                    "primary_cat": "cs.CV",
+                    "cats": [
+                        "cs.CV"
+                    ],
+                    "main_content": "Introduction Supervision Weak Aug Strong Aug Bipartite Matching Stage-wise Hybrid Matching EMA \u2026 \u2026 \u2026 \u2026 DETR (a) DETR-SSOD vanilla (b) Semi-DETR DETR \u2026 \u2026 Teacher DETR \u2026 \u2026 Teacher DETR Student Cross-view query Consistency Weak Aug Strong Aug EMA Supervision Student Figure 1. Comparisons between the vanilla DETR-SSOD framework based on the Teacher-Student architecture and our proposed Semi-DETR framework. Semi-DETR consists of Stage-wise Hybrid Matching, Cross-view Query Consistency powered by a costbased pseudo label mining strategy. Semi-supervised object detection (SSOD) aims to boost the performance of a fully-supervised object detector by exploiting a large amount of unlabeled data. Current state-ofthe-art SSOD methods are primarily based on object detectors with many hand-crafted components, e.g., rule-based label assigner [10, 27, 28, 33] and non-maximum suppression (NMS) [2] post-processing. We term this type of object detector as a traditional object detector. Recently, DETR [3], a simple transformer-based end-to-end object detector, has received growing attention. Generally, the DETR-based framework builds upon transformer [34] encoder-decoder architecture and generates unique predictions by enforcing a set-based global loss via bipartite matching during training. It eliminates the need for various hand-crafted components, achieving state-of-the-art performance in fully-supervised object detection. Although the performance is desirable, how to design a feasible DETR-based SSOD framework remains under-explored. There are still no systematic ways to fulfill this research gap. arXiv:2307.08095v1  [cs.CV]  16 Jul 2023 \f\u0000\u0001 \u0002\u0001 \u0000\u0003\u0001 \u0000\u0003\u0003\u0001 \u0000\u0001\u0002\u0003\u0002\u0001\u0004\u0005\u0002\u0006\u0007\u0002\b\u0007\t \u000b\f\u0007\r\u000e\u000f\u000b\r\u000b\f\u0007\f\u000e\u0004\u000e \u0010\u0011 \u0012\u0013 \u0012\u0011 \u0014\u0013 \u0014\u0011 \u0015\u0013 \u0015\u0011 \u0011\u0013 \u0011\u0011 \u0000\u0001\u0000\u0001\u0002\u0003\u0004\u0005 \u0012\u0012\u0016\u0015 \u0014\u0012\u0016\u0011 \u0014\u0017\u0016\u0013 \u0015\u0017\u0016\u0010 \u0012\u0012\u0016\u0014 \u0014\u0014\u0016\u0013 \u0014\u0018\u0016\u0010 \u0015\u0017\u0016\u0010 \u0014\u0013\u0016\u0011 \u0015\u0013\u0016\u0010 \u0015\u0014\u0016\u0011 \u0011\u0013\u0016\u0015 \u0000 \u000b\u0019\u0002\u001a\u001b\u000e \u0004\u000b\u0001\u0007\u001c\u0019\u001d\u001d\u001e \u001f\u000b\u0006 \u000b\u0007 \u000b\u000e!\"\u000b\u0001\u001a\u001b\u0019#$\u001e $\u000b%\u0005&\u001f' \u001c\u001a\u001f(\u001d#\u001e Figure 2. Performance comparisons between the proposed SemiDETR and other SSOD methods, including PseCo [17] and DenseTeacher [46]. Designing an SSOD framework for DETR-based detectors is non-trivial. Concretely, DETR-based detectors take a one-to-one assignment strategy where the bipartitematching algorithm forces each ground-truth (GT) bounding box to match a candidate proposal as positive, treating remains as negatives. It goes well when the ground-truth bounding boxes are accurate. However, directly integrating DETR-based framework with SSOD is problematic, as illustrated in Fig. 1 (a) where a DETR-SSOD vanilla framework utilizes DETR-based detectors to perform pseudo labeling on unlabeled images. In the Teacher-Student architecture, the teacher model usually generates noisy pseudo bounding boxes on the unlabeled images. When the pseudo bounding box is inaccurate, the one-to-one assignment strategy is doomed to match a single inaccurate proposal as positive, leaving all other potential correct proposals as negative, thus causing learning inefficiency. As a comparison, the one-to-many assignment strategy adopted in the traditional object detectors maintains a set of positive proposals, having a higher chance of containing the correct positive proposal. On the one hand, the one-to-one assignment strategy enjoys the merits of NMS-free end-to-end detection but suffers the training inefficiency under semi-supervised scenarios; on the other hand, the one-to-many assignment strategy obtains candidate proposal set with better quality making the detector optimized more efficiently but inevitably resulted in duplicate predictions. Designing a DETR-based SSOD framework that embraces these two merits could bring the performance to the next level. Additionally, the consistency-based regularization commonly used in current SSOD methods becomes infeasible in DETR-based SSOD. Specifically, current SSOD methods [4, 11, 14, 17] utilize consistency-based regularization to help object detectors learn potential feature invariance by imposing consistency constraints on the outputs of pairs-wise inputs (such as scale consistency [4, 11, 17], weak-strong consistency [14], etc.). Since the input features are deterministic in traditional object detectors, there is a one-to-one correspondence between the inputs and outputs, which makes the consistency constraint convenient to implement. However, this is not the case in DETRbased detectors. DETR-based detectors [3, 16, 21, 44, 48] use randomly initialized learnable object queries as inputs and constantly update the query features through the attention mechanism. As the query features update, the corresponding prediction results constantly change, which has been verified in [16]. In other words, there is no deterministic correspondence between the input object queries and its output prediction results, which prevents consistency regularization from being applied to DETR-based detectors. According to the above analysis, we propose a new DETR-based SSOD framework based on the TeacherStudent architecture, which we term Semi-DETR presented in Fig. 1 (b). Concretely, we propose a Stage-wise Hybrid Matching module that imposes two stages of training using the one-to-many assignment and the one-to-one assignment, respectively. The first stage aims to improve the training efficiency via the one-to-many assignment strategy and thus provide high-quality pseudo labels for the second stage of one-to-one assignment training. Besides, we introduce a Cross-view Query Consistency module that constructs cross-view object queries to eliminate the requirement of finding deterministic correspondence of object queries and aids the detector in learning semantically invariant characteristics of object queries between two augmented views. Furthermore, we devise a Cost-based Pseudo Label Mining module based on the Gaussian Mixture Model (GMM) that dynamically mines reliable pseudo boxes for consistency learning according to their matching cost distribution. Differently, Semi-DETR is tailored for DETR-based framework, which achieves new SOTA performance compared to the previous best SSOD methods. To sum up, this paper has the following contributions: \u2022 We present a new DETR-based SSOD method based on the Teacher-Student architecture, called SemiDETR. To our best knowledge, we are the first to examine the DETR-based detectors on SSOD, and we identify core issues in integrating DETR-based detectors with the SSOD framework. \u2022 We propose a stage-wise hybrid matching method that combines the one-to-many assignment and one-to-one assignment strategies to address the training inefficiency caused by the inherent one-to-one assignment within DETR-based detectors when applied to SSOD. \u2022 We introduce a consistency-based regularization scheme and a cost-based pseudo-label mining algorithm for DETR-based detectors to help learn semantic feature invariance of object queries from different augmented views. \f\u2022 Extensive experiments show that our Semi-DETR method outperforms all previous state-of-the-art methods by clear margins under various SSOD settings on both MS COCO and Pascal VOC benchmark datasets. 2. Related Work Semi-Supervised Object Detection. In SSOD, Pseudo Labeling [22,26,30,35,36,38,39,41, 47] and Consistency-based Regularization [4,11,14,17,23, 25, 28, 46] are two commonly used strategies. A detailed description can be found in the supplementary document. However, most of these works are based on the traditional detectors, e.g. Faster RCNN [28], which involves many hand-crafted components, e.g anchor box, NMS, etc. Our Semi-DETR is significantly different from previous works: (1) we explored the challenges of the DETR-based object detectors on SSOD, which, to our best knowledge, is the first systematic research endeavor in SSOD; (2) our SemiDETR method is tailored for the DETR-based detectors, which eliminates the training efficiency caused by bipartite matching with the noisy pseudo labels and presents a new consistency scheme for set-based detectors. End-to-End Object Detection with Transformer. The pioneering work DETR [3] introduced transformers into object detection to eliminate the need for complex handcrafted components in traditional object detectors. Many follow-up works have been dedicated to solving the slow convergence and high complexity issues of DETR [16, 21, 24,42,44,48]. Recently, DINO [44] combined with a variety of improvements related to DETR, such as query selection [42,48], contrastive query denoising [16], and achieved SOTA performance across various object detection benchmark datasets with excellent convergence speed. Complementary to these, we aim to extend the study of DETRbased detectors to SSOD and present Semi-DETR, which is a tailored design for SSOD. Our framework is agnostic to the choice of DETR-based detectors and could easily integrate with all DETR-based detectors. Omni-DETR [37] is a DETR-based object detector designed for omni-supervised object detection. It is not designed specifically for SSOD as admitted in their paper, but it is extended to the SSOD task by introducing a simple pseudo-label filtering scheme. Our Semi-DETR is significantly different from Omni-DETR in the following aspects: (1) Different motivations for model design; (2) Different training strategy; (3) Significant performance improvement. The detailed discussion is in Supplementary Document. 3. Semi-DETR 3.1. Preliminary We aim to address the problem of semi-supervised DETR-based object detection, where a labeled image set Ds = {xs i, ys i }Ns i=1 and an unlabeled image set Du = {xu i }Nu i=1 are available during training. Ns and Nu denote the amount of labeled and unlabeled images. For the labeled images xs, the annotations ys contain the coordinates and object categories of all bounding boxes. 3.2. Overview The overall framework of our proposed Semi-DETR is illustrated in Fig. 3. Following the popular teacher-student paradigm [31] for SSOD, our proposed Semi-DETR adopts a pair of teacher and student models with exactly the same network architecture. Here we adopt DINO [44] as an example while the overall framework of Semi-DETR is compatible with other DETR-based detectors. Specifically, in each training iteration, weak-augmented and strongaugmented unlabeled images are fed to the teacher and student, respectively. Then the pseudo labels generated by the teacher with confidence scores larger than \u03c4s served as supervisions for training the student. The parameters of the student are updated by back-propagation, while the parameters of the teacher model are the exponential moving average (EMA) of the student. Our main contribution contains three new components: stage-wise hybrid matching, cross-view query consistency, and cost-based pseudo-label mining, which address the core issues of DETR-based SSOD. In the following sections, we introduce more details of our proposed Semi-DETR. 3.3. Stage-wise Hybrid Matching DETR-based frameworks rely on one-to-one assignment for end-to-end object detection. For DETR-based SSOD framework, an optimal one-to-one assignment \u02c6 \u03c3o2o can be obtained by performing the Hungarian algorithm between the predictions of the student and pseudo-labels generated by the teacher:    \\ha t  {\\ sig ma } _ { o2o } = \\u n d er se t  { \\sig ma \\in \\xi _{N}}{\\arg \\min } \\sum _{i=1}^{N} \\mathcal {C}_{\\operatorname {match}}\\left (\\hat {y}^{t}_{i}, \\hat {y}^{s}_{\\sigma (i)}\\right ) \\label {eq:one2one}  (1) where \u03beN is the set of permutations of N elements and Cmatch \u0010 \u02c6 yt i, \u02c6 ys \u03c3(i) \u0011 is the matching cost between the pseudolabels \u02c6 yt and the prediction of the student with index \u03c3(i). However, in the early stage of SSOD training, the pseudo-labels generated by the teacher are usually inaccurate and unreliable, which imposes a high risk of generating sparse and low-quality proposals under the one-to-one assignment strategy. To exploit multiple positive queries to realize efficient semi-supervised learning, we propose to replace the one-to-one assignment with the one-to-many assignment:    \\ha t   { \\sig ma  }_{o2 m }   =\\l eft \\{ \\ u nd er s et  {\\bm   { \\ s igma _i} \\in C^M_{N}}{\\arg \\min } \\sum _{j=1}^{M} \\mathcal {C}_{\\operatorname {match}}\\left (\\hat {y}^{t}_{i}, \\hat {y}^{s}_{\\bm {\\sigma _i(j)}}\\right )\\right \\}_{i=1}^{|\\hat {y}^t|}. \\label {eq:one2many}  y (2) \fTeacher Weak Aug Strong Aug Unlabeled Data Labeled Data EMA Update One-to-Many Assignment One-to-One Assignment NMS \ud835\udc3f! \ud835\udc3f! \ud835\udc3f\" + + Evolve to NMS Free Detector Score Filter Pseudo GT Boxes for Consistency Object  queries Object  queries Cost-based Pseudo Label Mining Student Stage-2 Student Stage-1 Cross-view query  consistency \u2026 \u2026 \u2026 \u2026 Figure 3. Overview of Semi-DETR. Our framework is based on Teacher-Student architecture. Specifically, a multi-stage training strategy is derived to avoid incorrect bipartite matching with low-quality pseudo labels. Hybrid Matching with one-to-many assignment is applied in the first stage to generate higher quality pseudo labels for the following one-to-one training stage. Besides, a cross-view query consistency loss is designed to further enhance the consistency learning for the whole training process, where the pseudo boxes are filtered by a costbased GMM mining module. where CM N is the combination of M and N, which denotes that a subset of M proposals is assigned to each pseudo box \u02c6 yt i. Following [7,8], we utilize a high-order combination of classification score s and the IoU value u as the matching cost metric: m  m =  s^{\\alpha }\\cdot u^{\\beta } \\label {eq:match_criterion}  (3) where \u03b1 and \u03b2 control the effect of classification score and IoU during the assignment, and following [7], we set \u03b1 = 1, \u03b2 = 6 by default. With the one-to-many assignment, M proposals with the largest m values are selected as positive samples while regarding the remaining proposals as negative ones. We train the model with one-to-many assignment for T1 iterations in the early stage of semi-supervised training. Following [7,18], the classification loss and regression loss are also modified at this stage:   \\m ath c al { L }^{ o 2m}_{c l s}=\\sum _{i=1}^{N_{\\text {pos }}}\\left |\\hat {m}_{i}-s_{i}\\right |^{\\gamma } B C E\\left (s_{i}, \\hat {m}_{i}\\right )+\\sum _{j=1}^{N_{\\text {neg }}} s_{j}^{\\gamma } B C E\\left (s_{j}, 0\\right ) \\label {eq:o2m_cls_loss}  m (4)   \\m ath c al { L }^{ o2m}_{r e g}=\\sum _{i=1}^{N_{\\text {pos }}} \\hat {m}_{i} \\mathcal {L}_{G I o U}\\left (b_{i}, \\hat {b}_{i}\\right )+\\sum _{i=1}^{N_{\\text {pos }}} \\hat {m}_{i} \\mathcal {L}_{L_{1}}\\left (b_{i}, \\hat {b}_{i}\\right ) \\label {eq:o2m_reg_loss}  m (5)   \\m a thca l { L }^{o 2m} = \\mathcal {L}^{o2m}_{cls} + \\mathcal {L}^{o2m}_{reg}  (6) where \u03b3 is set to 2 by default. With multiple assigned positive proposals for each pseudo label, the potentially highquality positive proposals also get the chance to be optimized, which greatly improves the convergence speed and, in turn, obtains pseudo labels with better quality. However, Weak  Augmented View Strong  Augmented View RoI Align Deformable Cross-Attention Teacher Student Supervised by  pseudo ground truth Cross-view  queries Standard object queries Masked Deformable Self-Attention \u2026 \u2026 \u2026 \u2026 \u2026 \u2026 \u2026 \u2026 \u2026 \u2026 \ud835\udc3f! \ud835\udc3f\" Cross-view  queries Standard object queries Decoder  Layer X N  Decoder  Layer X N  Masked Deformable Self-Attention Deformable Cross-Attention Backbone Encoder Backbone Encoder Figure 4. Overview of the cross-view query consistency module. Query embeddings from the RoI features of pseudo labels on different views are cross-wise sent to the teacher and student decoders. The corresponding decoded features are enforced to be similar by a consistency loss. the multiple positive proposals for each pseudo label result in duplicate predictions. To mitigate this problem, we propose to switch back to the one-to-one assignment training in the second stage. By doing this, we enjoy the high-quality pseudo labels after the first stage training and gradually reduce duplicate predictions to reach an NMS-free detector with one-to-one assignment training at the second stage. The loss functions of this stage are the same as [44]:   \\m a thca l { L }^{o 2o} = \\mathcal {L}^{o2o}_{cls} + \\mathcal {L}^{o2o}_{reg}  (7) 3.4. Cross-view Query Consistency Traditionally, in non-DETR-based SSOD frameworks, consistency regularization can be employed conveniently \fby minimizing the difference between the output of teacher f\u03b8 and student f \u2032 \u03b8, given the same input x with different stochastic augmentation:   \\ m athc al {L}_{c}  = \\sum _{\\mathbf {x} \\in \\mathcal {D}_{u}} \\operatorname {MSE}\\left (f_{\\theta }(\\mathbf {x}), f_{\\theta }^{\\prime }(\\mathbf {x})\\right )  (8) However, for DETR-based frameworks, as there is no clear (or deterministic) correspondence between the input object queries and their output prediction results, conducting consistency regularization becomes infeasible. To overcome this issue, we propose a Cross-view Query Consistency module that enables the DETR-based framework to learn semantically invariant characteristics of object queries between different augmented views. Fig. 4 illustrates our proposed cross-view query consistency module. Specifically, for each unlabeled image, given a set of pseudo bounding boxes b, we process the RoI features extracted via RoIAlign [12] with several MLPs:   \\ begin {aligned} c_t  =  \\operatorname {M LP}(\\operatorname {RoIAlign}(F_t, b)) \\\\ c_s = \\operatorname {MLP}(\\operatorname {RoIAlign}(F_s, b)) \\label {eq:roi_align} \\end {aligned}  (9) where Ft and Fs denote the backbone feature of the teacher and student, respectively. Subsequently, ct and cs are regarded as cross-view query embeddings and attached to the original object queries in another view to serve as the input of the decoder:    \\b eg i n {aligned} \\ hat {o}_{ t },o _{ t } &= \\operato rnam e {Decoder}_t([c_s,q_t], E_t| A) \\\\ \\hat {o}_{s},o_{s} &= \\operatorname {Decoder}_s([c_t,q_s], E_s| A) \\end {aligned} \\label {eq:cross_view decoding}  (10) where q\u00b7 and E\u00b7 denote the original object queries and the encoded image features, respectively. \u02c6 o\u00b7 and o\u00b7 denote the decoded features of cross-view queries and original object queries. Note the subscript t and s indicate teacher and student, respectively. Following [16], the attention mask A is also employed to avoid information leakage. With the semantic guide of input cross-view queries embeddings, the correspondence of the decoded features can be naturally guaranteed, and we impose consistency loss as follows:   \\ mathc al {L}_{c}  =\\operatorname {MSE}(\\hat {o}_{s}, \\operatorname {detach}(\\hat {o}_{t}))  (11) 3.5. Cost-based Pseudo Label Mining To mine more pseudo boxes with meaningful semantic contents for the cross-view query consistency learning, we propose a cost-based pseudo label mining module that dynamically mines reliable pseudo boxes in the unlabeled data. Specifically, we perform an additional bipartite matching between the initial filtered pseudo boxes and the predicted proposals and utilize the matching cost to describe the reliability of the pseudo boxes:   C _ {ij} = \\la m bda  _1C_{Cls}(p _ i,\\ h at {p_j})   + \\lambda _2C_{GIoU}(b_i,\\hat { b_j}) + \\lambda _3C_{L_1}(b_i,\\hat {b_j}) \\label {eq:match_cost}  (12) where pi, bi represents the classification and regression result of i-th predicted proposals while \u02c6 pj, \u02c6 bj indicates the class label and box coordinates of j-th pseudo label. Subsequently, in each training batch, we cluster the initial pseudo boxes into two states by fitting a Gaussian Mixture Model for the matching cost distribution. As illustrated in Fig. 5, the matching cost aligns well with the quality of pseudo boxes. We further set the cost value of the clustering center of the reliable ones as the threshold and collect all pseudo boxes with lower cost than the threshold for the cross-view query consistency calculation. 3.6. Loss Function The final loss L is represented as follows:   \\be g in { align ed}  \\m a thca l {L}  & = \\ m ath b b {I} (t\\ l eq  T_1) \\cdot  ( \\m a thcal {L}^{o2m}_{sup}+w_u\\cdot \\mathcal {L}^{o2m}_{unsup}) \\\\ &+ \\mathbb {I}(t > T_1)\\cdot (\\mathcal {L}^{o2o}_{sup}+w_u\\cdot \\mathcal {L}^{o2o}_{unsup}) \\\\ &+ w_c\\cdot \\mathcal {L}_c \\end {aligned}  (13) where L\u00b7 sup and L\u00b7 unsup are the supervised loss and the unsupervised loss, respectively, containing both the classification loss and regression loss. The Lc means the cross-view consistency loss. The wu and wc are the unsupervised loss weight and consistency loss weight, which set wu = 4 and wc = 1 by default. t is the current training iteration and T1 is the duration time of the first stage training within the SHM module. 4. Experiments 4.1. Datasets and Evaluation Metrics We validate our method on the MS-COCO benchmark [20] and Pascal VOC datasets [6]. MS-COCO contains 80 classes with 118k labeled images in the train2017 set and 123k unlabeled images in the unlabeled2017 set. In addition, the val2017 set with 5k images is provided for validation. Following [39], we consider two evaluation settings to validate our method on the MS-COCO benchmark: (1) COCO-Partial. 1%, 5%, and 10% images of the COCO train2017 set are randomly sampled as the labeled training data, and the remaining images of train2017 are regarded as the unlabeled data. 5 different data folds are created for each data split to validate our method. The average of standard COCO mAP on the val2017 is adopted as our final performance metric. (2) COCO-Full. Under this setting, the entire train2017 is utilized as the labeled data, and unlabeled2017 is used as the additional unlabeled data. The standard COCO mAP on the val2017 is taken as the evaluation metric. Pascal VOC contains 20 classes with VOC2007 and VOC2012 provided as the labeled data and unlabeled data respectively. The evaluation metrics are the COCO-style AP50:95 and AP50 on the VOC2007 test set. \fReliable Unreliable (a) Distribution of Cost (b) Initial Pseudo Boxes (c) Unreliable Pseudo Boxes (d) Reliable Pseudo Boxes Figure 5. An illustration of Cost-based Pseudo Label Mining. We first take the image-level confidence score\u2019s mean with variance to get the initial pseudo labels, shown in (b), for each image and perform the Hungarian match to get the matching cost of each pseudo ground truth bounding box within a batch. Then, we fit a GMM model with these cost values, shown in (a). We argue that the pseudo boxes with lower cost are more likely to be the reliable pseudo boxes, so we take the lower threshold from the GMM model to filter the pseudo label again to obtain the final pseudo boxes presented in (d). 4.2. Implementation Details To avoid loss of generality, we choose Deformable DETR [48] and DINO to integrate into our Semi-DETR method. Following them, we use ResNet-50 [13] pretrained on ImageNet [5] as our backbone network. Focal Loss [19] is used for classification during training. Smooth L1 Loss and GIoU [29] Loss are used for regression. We set the number of object queries to 300 for Deformable DETR and 900 for DINO, respectively. For the training hyperparameters, following [39]: (1) For the COCO-Partial benchmark, we train Semi-DETR for 120k iterations on 8 GPUs with 5 images per GPU. The first stage with one-to-many assignment is set to 60k iterations. The ratio of the labeled data and unlabeled data is set to 1:4. The weight of the unsupervised loss is set to \u03b1 = 4.0. (2) For the COCOFull benchmark, we double the training time with COCOunlabeled to 240k, where the first stage with one-to-many assignment is set to 180k iterations. The batch size is set to 64 on 8 GPUs with 8 images per GPU. The ratios of labeled data and unlabeled data are set to 1:1, and the loss weight of unlabeled data is set to \u03b1 = 2.0. (3) For the Pascal VOC benchmark, we train Semi-DETR for 60k iterations where The first stage with one-to-many assignment is set to 40k iterations. Other settings are kept the same with COCO-Partial. For all experiments, the confidence threshold is set to 0.4. We utilize Adam [15] with a learning rate of 0.001, and no learning rate decay scheme is used. The teacher model is updated through EMA with a momentum of 0.999. Besides, we follow the same data prepossessing, and augmentation pipeline in [39] without modifications. 4.3. Comparison with SOTA methods We compare our Semi-DETR method with current SOTA SSOD methods on both MS-COCO and Pascal VOC datasets. We present the superiority of Semi-DETR in the following aspects: (1) comparisons to two-stage and onestage detectors, (2) comparisons to DETR-based detectors, and (3) generalization ability. COCO-Partial benchmark. According to Tab. 1, SemiDETR shows significant superiority over current SOTA SSOD methods across all experiment settings in COCOPartial. Concretely, (1) compared to SOTA two-stage and one-stage detectors, Semi-DETR outperforms PseCo (experiment 3) by 2.77, 2.00, 2.05 mAP with Deformable DETR (by 8.07, 7.60, 7.44 mAP with DINO) under the 1%, 5%, 10% settings and beats Dense Teacher (experiment 5) by 2.82, 1.49, 0.97 mAP with Deformable DETR (by 8.12, 7.09, 6.37 mAP with DINO) under the 1%, 5%, 10% settings. Obviously, Semi-DETR is a better semi-supervised object detector, and it does not require hand-crafted components used in two-stage and one-stage detectors; (2) we construct two DETR-based baselines, namely DETR under supervised training only and a simple pseudo labeling Teacher-Student architecture integrating DETR with SSOD. By comparing experiments 7-10 (or experiments 11-14), Semi-DETR outperforms the supervised baseline by 14.20, 10.80, 8.90 mAP with Deformable DETR (12.50, 10.60, 8.50 mAP with DINO) and surpasses the SSOD baseline by 5.80, 3.40, 3.30 mAP with Deformable DETR (2.10, 2.10, 1.90 mAP with DINO). This demonstrates that simply integrating DETR-based detectors with Teacher-Student architecture is not optimal. (3) we use Deformable DETR and DINO to show the generalization ability of our Semi-DETR method. Apparently, Semi-DETR consistently boosts the performance of both detectors over the corresponding baselines by clear margins (experiments 7-14). With stronger detectors like DINO, Semi-DETR still enjoys a notable performance improvement. COCO-Full benchmark. According to Tab. 3, when adding additional unlabeled2017 data, Semi-DETR with Deformable DETR enjoys 3.6 mAP performance gain and reaches 47.2 mAP, surpassing PseCo and Dense Teacher by 1.1 and 1.1 mAP, respectively. This further manifests the effectiveness of Semi-DETR. Besides, under stronger baselines like DINO, Semi-DETR still shows obvious performance gain (+1.8 mAP), which outperforms PseCo and Dense Teacher by 4.3 and 4.3 mAP respectively, and gener\fTable 1. Comparisons with SOTA SSOD methods under the COCO-Partial setting. All results are the average of all 5 folds. Def-DETR denotes Deformable DETR. Sup Only denotes supervised only baseline. Category Method ID 1% 5% 10% Unbiased Teacher 1 20.75 \u00b1 0.12 28.27 \u00b1 0.11 31.50 \u00b1 0.10 Two-Stage Soft-Teacher 2 20.46 \u00b1 0.39 30.74 \u00b1 0.08 34.04 \u00b1 0.14 PseCo 3 22.43 \u00b1 0.36 32.50 \u00b1 0.08 36.06 \u00b1 0.24 DSL 4 22.03 \u00b1 0.28 30.87 \u00b1 0.24 36.22 \u00b1 0.18 One-Stage Dense Teacher 5 22.38 \u00b1 0.31 33.01 \u00b1 0.14 37.13 \u00b1 0.12 Unbiased Teacher v2 6 22.71 \u00b1 0.42 30.08 \u00b1 0.04 32.61 \u00b1 0.03 Omi-DETR(Def-DETR) 7 18.60 30.20 34.10 Def-DETR(Sup only) 8 11.00 \u00b1 0.24 23.70 \u00b1 0.13 29.20 \u00b1 0.11 Def-DETR SSOD(Baseline) 9 19.40 \u00b1 0.31 31.10 \u00b1 0.21 34.80 \u00b1 0.09 Semi-DETR(Def-DETR) 10 25.20 \u00b1 0.23 34.50 \u00b1 0.18 38.10 \u00b1 0.14 End-to-End DINO(Sup only) 11 18.00 \u00b1 0.21 29.50 \u00b1 0.16 35.00 \u00b1 0.12 DINO SSOD (Baseline) 12 28.40 \u00b1 0.21 38.00 \u00b1 0.13 41.60 \u00b1 0.11 Omi-DETR(DINO) 13 27.60 37.70 41.30 Semi-DETR(DINO) 14 30.50 \u00b1 0.30 40.10 \u00b1 0.15 43.50 \u00b1 0.10 Table 2. Comparisons with SOTA SSOD methods under the Pascal VOC setting. Def-DETR denotes Deformable DETR. Sup Only denotes supervised only baseline. Category Method AP50 AP50:95 Unbiased Teacher 77.37 48.69 Two-Stage Soft-Teacher PseCo DSL 80.70 56.80 One-Stage Dense Teacher 79.89 55.87 Unbiased Teacher v2 81.29 56.87 Def-DETR(Sup only) 74.50 46.20 Def-DETR SSOD(Baseline) 78.90 53.40 Semi-DETR(Def-DETR) 83.50 57.20 DINO(sup only) 81.20 59.60 End-to-End DINO SSOD (Baseline) 84.30 62.20 Semi-DETR(DINO) 86.10 65.20 ates a new SOTA performance of 50.4 mAP. Pascal VOC benchmark. Semi-DETR presents consistent performance improvements on the Pascal VOC benchmark as shown in Tab. 2. Generally, Semi-DETR outperforms the supervised baseline by 9.0 on AP50 and 11.0 on AP50:95 with Deformable DETR (by 4.9 on AP50 and 5.6 on AP50:95 with DINO). Furthermore, Semi-DETR beats all previous SOTA SSOD methods by significant margins on both AP50 and AP50:95. 4.4. Ablation Study We conduct extensive experiments to verify the effectiveness of Semi-DETR in the following aspects: (1) component effectiveness; (2) variants of Stage-wise Hybrid Matching (SHM); (3) effectiveness of Cross-view Query Consistency (CQC) and Cost-based Pseudo Label Mining (CPM); (4) hyper-parameters. All experiments are performed with DINO as the base detector on the 10% labeled images setting of the COCO-Partial benchmark. Table 3. Comparisons with SOTA SSOD methods under the COCO-Full setting. Def-DETR denotes Deformable DETR. Sup Only denotes supervised only baseline. Method 100% Unbiased Teacher 40.2 +1.1 \u2212 \u219241.3 Soft-Teacher 40.9 +3.6 \u2212 \u219244.5 PseCo 41.0 +5.1 \u2212 \u219246.1 DSL 40.2 +3.6 \u2212 \u219243.8 Dense Teacher 41.2 +3.6 \u2212 \u219246.1 Semi-DETR(Def-DETR) 43.6 +3.6 \u2212 \u219247.2 Semi-DETR(DINO) 48.6 +1.8 \u2212 \u219250.4 Component Effectiveness. According to Tab. 4, we perform four experiments to verify the effectiveness of each proposed component. We formulate a strong baseline that integrates DINO with SSOD via pseudo labeling in experiment 1. In general, our proposed components enjoy consistent performance improvements. Specifically, by introducing the SHM module, it outperforms the baseline by 1.1 mAP. Further integrating the CQC and CPM modules brings an extra 0.8 improvement. This shows that our proposed components are complementary to each other and proves the effectiveness of each component in our model. Variants of SHM. We examine the impact of different one-to-many assignment strategies within SHM in the first stage of training. Concretely, Max-IoU [28], ATSS [45] and SimOTA [9] are chosen as the alternatives. All models are trained for 60k iterations. As presented in Tab. 6, it is interesting to find that not all traditional one-to-many assignment methods are effective in DETR-based detectors. Max-IoU assignment strategy and ATSS show significant performance degradation when applied to the first stage, \fTable 4. Component effectiveness of Semi-DETR. SHM denotes the Stage-wise Hybrid Matching, CQC means Cross-view Query Consistency, and CPM represents Cost-based Pseudo Label Mining, respectively. ID SHM CQC CPM mAP AP50 AP75 1 41.6 58.3 45.1 2 \u2713 42.7 59.3 46.2 3 \u2713 \u2713 43.1 59.6 46.6 4 \u2713 \u2713 \u2713 43.5 59.7 46.8 even though they are commonly used in traditional object detectors. On the other hand, SimOTA shows comparable performance to our one-to-many assignment strategy. This is possibly caused by the fact that SimOTA and our method adopt a ranking-based one-to-many assignment strategy while Max-IoU and ATSS utilize hard or dynamic thresholding-based one-to-many assignment strategy, which leads to a significant difference number of assigned positive samples for each pseudo ground truth bounding box and thus suffers performance degradation. More analysis can be found in the supplementary document. Effectiveness of CQC+CPM. According to Tab. 5, we compare four different methods to generate pseudo labels for CQC and evaluate the precision and recall metrics of the generated pseudo labels. First, we present two methods (by setting a fixed classification score \u03c4s = 0.4 or by selecting Top-K pseudo labels with the highest confidence scores) that obtain pseudo labels with high precision (81.5% or 80.2%) and low recall (41.3% or 39.4%) but observe marginal performance gains. Then we present the Mean+Std method that aims to balance the precision (60.2%) and recall (54.0%) of pseudo labels via combining the image-level mean confidence score and variance \u03c4 = \u00b5 + \u03c3, which enjoys a better performance improvement (+0.4 mAP). Finally, our Cost-based GMM method achieves a better trade-off between the precision (77.6%) and recall (52.1%) metrics, which has a 0.8 performance gain. Table 5. Effects of different methods to filter pseudo labels for cross-view consistency training. Method mAP Precision Recall Fixed(0.4) 42.8 81.5% 41.3% Top-K(K=9) 42.9 80.2% 39.4% Mean + Std 43.1 60.2% 54.0% Cost-based GMM 43.5 77.6% 52.1% Hyperparameters. We study two types of hyperparameters in our model: (1) the pseudo label threshold \u03c4s; (2) the training iterations T1 of the first stage in SHM. For \u03c4s, according to Tab. 8, the best performance is achieved when \u03c4s = 0.4. Possibly, a lower threshold could introduce noisy pseudo labels, while a higher threshold could decrease the effective number of pseudo labels. For the Table 6. Effects of the different one-to-many assignment methods in the first stage. Strategy mAP AP50 AP75 Max-IoU 11.4 15.0 12.1 ATSS 18.7 30.5 18.9 SimOTA 42.5 59.9 45.2 Ours 42.8 59.8 46.0 Table 7. Effects of the training iteration T1 of the first stage using one-to-many assignment strategy in Stage-wise Hybrid Matching. T1 40k 60k 80k 100k 120k mAP 42.9 43.5 43.2 43.0 44.0 NMS-Free Y Y Y Y N Table 8. Effects of the pseudo label threshold \u03c4s. \u03c4s 0.2 0.3 0.4 0.5 0.6 mAP 42.6 43.0 43.5 43.2 42.8 training iterations of the first stage, according to Tab. 7, performing the one-to-many assignment strategy across both stages achieves 44.0 mAP at the cost of using NMS in the end. The appropriate training time of the first stage is at 60k iterations, which achieves the best performance of 43.5 mAP and does not require NMS post-process at the same time. 5."
+                }
+            ],
+            "Jiajun Wu": [
+                {
+                    "url": "http://arxiv.org/abs/2307.12105v2",
+                    "title": "$W$-boson Mass Anomaly from High-Dimensional Scalar Multiplets",
+                    "abstract": "In light of the recently discovered $W$-boson mass anomaly by the CDF\nCollaboration, we discuss two distinct mechanisms that could possibly explain\nthis anomaly through the introduction of high-dimensional $SU(2)_L$ scalar\nmultiplets. The first mechanism is the tree-level $W$-boson mass correction\ninduced by the vacuum expectation values of one or more $SU(2)_L$ scalar\nmultiplets with odd dimensions of $n\\geq 3$ and zero hypercharge of $Y=0$ in\norder to avoid the strong constraint from measurements of the $Z$-boson mass.\nThe second mechanism is to consider the one-loop level $W$-boson mass\ncorrection from a complex multiplet. In particular, we focus on the case with\nan additional scalar octuplet with $Y=7/2$. As a result, we find that both\nmechanisms can explain the $W$-boson mass anomaly without violating any other\ntheoretical or experimental constraints.",
+                    "authors": "JiaJun Wu, Chao-Qiang Geng, Da Huang",
+                    "published": "2023-07-22",
+                    "updated": "2023-10-26",
+                    "primary_cat": "hep-ph",
+                    "cats": [
+                        "hep-ph"
+                    ],
+                    "main_content": "INTRODUCTION Recently, the CDF-II Collaboration has reported the most precise measurement of the Wboson mass, showing that the observed W-boson mass of mCDF\u2212II W = 80433.5 \u00b1 9.4 MeV [1] deviates the latest Standard Model (SM) prediction of mSM W = 80357 \u00b1 6 MeV [2]. The significance of this anomaly is more than 7\u03c3, which indicates new physics (NP) beyond the SM (BSM). Therefore, it is a natural and pressing concern to introduce NP models to account for this anomalous W-boson mass. Among a multitude of BSM scenarios, extending the SM Higgs sector by incorporating extra SU(2)L multiplets [3\u201351] is a promising avenue, since the modification of the scalar sector is intricately connected to the underlying mechanism for the electroweak (EW) gauge symmetry breaking and the related hierarchy problem, which can be studied through the measurement of EW oblique parameters [52\u201357]. Furthermore, the added scalar multiplet has the potential to resolve many puzzles in the SM, such as the nature of dark matter (DM) [58\u201363], the origin of the matter-antimatter asymmetry [64\u201367], and the characteristics of the EW phase transition as well as its related stochastic gravitational wave signals [68\u2013 85]. Therefore, comprehending the structure of the scalar sector could lead to a more profound understanding of the big picture of the SM and the physics beyond it. Extensive studies have been carried out in the literature to explain the CDF-II W-boson mass anomaly with low-dimensional scalar multiplets, which include a scalar singlet [3\u20136], a second Higgs doublet [7\u201327], and a scalar triplet [28\u201344, 46\u201351]. More recently, we have explored scalar multiplet scenario up to a maximum of a septuplet [45]. In this letter, we aim to explain the W-boson mass anomaly with higher dimensional multiplets, at both tree and one-loop levels. In particular, for the one-loop solution, we shall focus on the case of a scalar octuplet with Y = 7/2 and zero vacuum expectation value (VEV), which is the highest dimension for a complex scalar allowed by the perturbative unitarity constraint [86]. This paper is organized as follows: In Sec. II, we investigate the possibility of the treelevel explanation of the CDF-II W-boson mass excess with a high-dimensional multiplet. In Sec. III, we present a detailed phenomenological study of a scalar octuplet with Y = 7/2, which is of physical interest to explain the CDF-II W-boson mass anomaly at the one-loop level. Finally, we conclude in Sec. IV. 2 \fII. TREE-LEVEL EXPLANATION OF W-BOSON MASS ANOMALY One may easily explain the W-boson mass anomaly by introducing a SU(2)L scalar multiplet with its VEV of the neutral component inducing an additional mass correction to the W-boson mass. However, such an idea is hampered by the fact that, since the W and Z-bosons have the common origin from the EW gauge symmetry breaking, the associated Z-boson mass should also be corrected, which was strongly constrained by the current experiments [2]. In this section, we show that if the added scalar multiplet is in an odd-dimensional presentation of SU(2)L with zero hypercharge, the above problem can be resolved automatically. Let us begin by considering a real SU(2)L multiplet \u03be of dimension n = 2k + 1 with k as a positive integer denoting the weak isospin SU(2)L representation. The hypercharge of \u03be is fixed to Y = 0. When \u03be and the SM Higgs doublet obtain their VEVs, v\u03be, and vH, from the spontaneous breaking of the EW gauge symmetry, the W and Z-boson mass terms can be written as follows D\u00b5H\u2020D\u00b5H + D\u00b5\u03be\u2020D\u00b5\u03be \u2283\u2212 \u00121 4g2v2 H + 1 2k(1 + k)v2 \u03be \u0013 W + \u00b5 W \u2212\u00b5 \u22121 8(g2 + g\u2032 2)v2 HZ\u00b5Z\u00b5 , (1) where D\u00b5 denotes the covariant derivatives of the scalar fields H and \u03be with g and g\u2032 the SM SU(2)L and U(1)Y gauge couplings, respectively. Therefore, the SM gauge boson masses are given by mW = 1 2g q v2 H + 2k(1 + k)v2 \u03be , mZ = vH 2 p g2 + g\u2032 2 , (2) indicating that the Z-boson mass is not corrected in the presence of the extra multiplet scalar VEV of v\u03be. It can be understood by the fact that, for Y = 0, the couplings of various components in the multiplet with the Z-boson are proportional to their electric charges, so that the neutral component and the associated VEV cannot interact with the Z boson. As a consequence, this scalar-multiplet-extended model can avoid the strong constraint from the Z-boson mass measurement [2]. Further, if we take the CDF-II value of the W-boson mass as the one in our model, then the VEV of the multiplet can be estimated as follows (\u2206v)2 v2 H \u22612k(1 + k)v2 \u03be v2 H = \u0012mCDF W mSM W \u00132 \u22121 \u223c[0.00090, 0.00201] , at 2\u03c3 C.L. , (3) 3 \fwhere the SM Higgs doublet VEV is taken to be vH = 246.22 GeV [2]. Moreover, note that the model can be further extended by introducing a series of SU(2)L scalar multiplets with vanishing hypercharges. In this case, we can still keep the salient feature that the Z-boson mass is not modified so that the related constraint is weak. Another critical constraint on the added scalar multiplet is provided by the constraint of the \u03c1 parameter, which can be related to the oblique parameter T as follows [2, 87] \u2206\u03c1 = \u03b1T. (4) In the light of the current constraint on T given by the updated electroweak global fit in Ref. [43] which includes the CDF-II W-boson mass and fixes S = U = 0, we have \u2206\u03c1 \u2208[0.00101, 0.00133] , at 2\u03c3 C.L. (5) Theoretically, our model predicts this important quantity as follows [2] \u03c1 = 1 + 2k(1 + k)v2 \u03be v2 H . (6) According to Eq. (5), we can obtain the following 2\u03c3 region for (\u2206v)2/v2 H (\u2206v)2 v2 H \u2208[0.00101, 0.00133] , at 2\u03c3 C.L. . (7) Obviously, the multiplet VEV range in Eq. (7) allowed by the \u03c1 parameter constraints is not in conflict with the CDF-II W-boson mass signal region in Eq. (3). Consequently, the CDFII mW anomaly can be solved in this multiplet scalar extension of the SM at the tree-level, while the data of \u03c1 strongly constrains the allowed parameter space. III. ONE-LOOP LEVEL EXPLANATION OF THE W-BOSON MASS ANOMALY In this section, we discuss the explanation of the W-boson mass anomaly with only the one-loop corrections from the addition of high-dimensional scalar multiplet fields. After a general review of the relation between EW oblique parameters and the one-loop W-boson mass correction, we shall focus as a concrete example on the case of a scalar octuplet with Y = 7/2 and a vanishing VEV. 4 \fA. Oblique Parameters and the W-Boson Mass The NP effects in the EW sector are usually encoded by the three oblique parameters, namely S, T, and U [52, 53], which are defined at the one-loop level as follows: S \u22614s2 Wc2 W \u03b1 \u0014 A\u2032 ZZ (0) \u2212c2 W \u2212s2 W cWsW A\u2032 Z\u03b3 (0) \u2212A\u2032 \u03b3\u03b3 (0) \u0015 , T \u2261 1 \u03b1m2 Z \u0014AWW (0) c2 W \u2212AZZ (0) \u0015 , U \u22614s2 W \u03b1 \u0014 A\u2032 WW (0) \u2212cW sW A\u2032 Z\u03b3 (0) \u2212A\u2032 \u03b3\u03b3 (0) \u0015 \u2212S , (8) where the functions A(\u2032) V V \u2032(q2) refer to the vacuum polarizations for EW gauge bosons V (\u2032) = \u03b3, W, Z. Moreover, as shown in Refs. [7, 53], the general one-loop corrections of non-SM scalars to the W-boson mass squared can be expressed in terms of these oblique parameters S, T, and U as follows \u2206m2 W = \u03b1c2 Wm2 Z c2 W \u2212s2 W \u0014 \u2212S 2 + c2 WT + c2 W \u2212s2 W 4s2 W U \u0015 , (9) where sW (cW) are (co)sine of the Weinberg angle. However, as demonstrated in Ref.[88], the parameters T and S are usually generated by dimension-6 operators, while U can only be induced by a dimension-8 operator, resulting in its significant suppression. Therefore, in a NP model augmented by a scalar multiplet, it is expected that the dominant one-loop contribution to the W-boson mass arises from T and S, and the impact of U can be neglected. Besides, Ref. [45] has explicitly confirmed this expectation by demonstrating that when the masses of extra scalars exceed 300 GeV and their multiplet dimensions are restricted to be smaller than 10, the correction of U to the W mass is at least one order of magnitude smaller than the leading ones from T and S. According to Refs.[45, 56], the contribution of a scalar multiplet \u03be to T can be expressed as: T\u03be = 1 4\u03c0s2 wm2 W k\u22121 X I=\u2212k N 2 I+1F \u0010 m2 \u03beQ I , m2 \u03beQ I+1 \u0011 , (10) where I = k, k \u22121, k \u22122, ......, \u2212k + 1, \u2212k, NI = p (k + I)(k \u2212I + 1)/2, Q = I + Y is referred to the electric charge, and the function F(A, B) is given by F (A, B) \u2261 \uf8f1 \uf8f4 \uf8f2 \uf8f4 \uf8f3 A+B 2 \u2212 AB A\u2212B ln A B , A \u0338= B , 0 , A = B . (11) 5 \fOn the other hand, the correction of \u03be to S can be expressed as follows: S\u03be = \u2212Y 3\u03c0 k X I=\u2212k I ln m2 \u03beQ I , (12) where \u03beQ I represents components of the scalar multiplet \u03be with its electric charge Q. B. Scalar octuplet Explanation of the W-boson Mass Anomaly According to Ref. [86], the highest dimension of a complex scalar multiplet allowed by the perturbative unitarity constraint is eight. Moreover, Ref. [45] has shown that a real scalar multiplet is unable to generate the mass splitting to produce nonzero values of T and S, which is required to explain the W-mass anomaly. Consequently, in the following, we will focus on the model by adding a complex octuplet scalar with Y = 7/2, which has not been discussed so far in the literature. Note that the potential in this model can be constructed as follows with the Higgs doublet H and the scalar octuplet \u03be V (H, \u03be) = \u2212\u00b52 HH\u2020H + \u03bbH \u0000H\u2020H \u00012 + \u00b52 \u03be\u03be\u2020\u03be + \u03bb1 \u0000\u03be\u2020\u03be \u00012 + \u03bb2 \u0000\u03be\u2020T a \u03a6\u03be \u00012 + \u03bb3 \u0000\u03be\u2020\u03be \u0001 \u0000H\u2020H \u0001 + \u03bb4 \u0000\u03be\u2020T a \u03be \u03be \u0001 \u0000H\u2020T a HH \u0001 + \u03bb5 \u0000\u03be\u2020T a \u03be T b \u03be \u03be \u00012 . (13) Here, we present the most generic interaction terms for a scalar octuplet \u03be. It is worth noting that the potential given in Eq.(13) exhibits an Z2 symmetry that arises naturally without the need for any additional assumptions. Note that the mass splitting can be only generated in Eq. (13) by the following term: O4 = \u03bb4 \u0000\u03be\u2020T a \u03be \u03be \u0001 \u0000H\u2020T a HH \u0001 . (14) Therefore, we will concentrate on this term in our subsequent phenomenological studies. Additionally, it is worth noting that |\u03bb4| is also constrained by perturbativity with |\u03bb4| < 4\u03c0 [89]. Hence, In our further discussions, we take into account this perturbative limit by requiring |\u03bb4| \u226410. For the model with a scalar octuplet \u03be of Y =7/2, the phenomenology can be classified into the following two distinct types based on the sign of \u03bb4: \u2022 Type A: when \u03bb4 > 0, the lightest particle in the octuplet is the most electrically charged one with its mass denoted as MC, so that ML = MC. 6 \f\u2022 Type B: when \u03bb4 < 0, the lightest particle in the octuplet is the electrically neutral one with its mass denoted as M0, indicating ML = M0. In light of the mass splittings among scalars from O4, this Y = 7/2 scalar octuplet has the potential to explain the W-mass anomaly with its following nonzero corrections to the parameters T and S: \u2206m2 W = \u03b1c2 Wm2 Z c2 W \u2212s2 W \u0014 \u2212S 2 + c2 WT \u0015 . (15) Figs. 1 and 2 display the parameter spaces in the ML\u221a \u2206m2 plane for the Type-A and Type-B scalar octuplet models, respectively. The horizontal axis denotes the mass of the lightest particle in the octuplet (ML), ranging from 1700 GeV to 5000 GeV, while the vertical axis denotes the mass splitting between adjacent components. The pink region in the figure indicates the parameter space allowed by EW global fits for T and S at the 2\u03c3 CL when U = 0, as reported in Ref. [88]. The cyan area corresponds to the parameter space that can explain the measured W-boson mass by CDF-II within the 2\u03c3 CL. The red solid line represents the scalar mass difference corresponding to the perturbative limit |\u03bb4| = 10.The results show that the Type-A model has a substantial amount of parameter space that can solve the CDF-II mW anomaly while satisfying the EW global fits and perturbative limits. On the other hand, for the Type-B model, it is seen from Fig. 2 that the CDF-II preferred region that explains the W-boson mass excess is entirely ruled out by the global fits of EW precision observables. IV."
+                }
+            ]
+        },
+        "edge_feat": {}
+    }
+}
\ No newline at end of file