Hugging Face's logo

Datasets:
MCG-NJU
/
Tracking-Any-Granularity

Languages:
English
ArXiv:
Tags:
video-object-segmentation
single-object-tracking
point-tracking
computer-vision
benchmark
License:
Dataset card Files
xet
Community
1
Tracking-Any-Granularity / README.md
jiamingZ's picture
jiamingZ
Update README.md
6f88df2 verified
preview code
|
raw
history blame contribute delete
7.92 kB
metadata
license: cc-by-sa-4.0
task_categories:
  - video-tracking
tags:
  - video-object-segmentation
  - single-object-tracking
  - point-tracking
  - computer-vision
  - benchmark
language:
  - en
pretty_name: TAG
arxiv: 2510.18822
configs:
  - config_name: default
    data_files: '*.json'
    sep: "\t"

SAM 2++: Tracking Anything at Any Granularity

πŸ”₯ Evaluation Server | 🏠 Homepage | πŸ“„ Paper | πŸ”— GitHub

Download

We recommend using huggingface-cli to download:

pip install -U "huggingface_hub[cli]"
huggingface-cli download MCG-NJU/Tracking-Any-Granularity --repo-type dataset --local-dir ./Tracking-Any-Granularity --local-dir-use-symlinks False --max-workers 16

πŸ”₯ Latest News

  • [2024-10-27] To provide a benchmark for the task of language reference, such as 'Tracking by natural language specification' and 'Referring video object segmentation', we added the language description of the object in meta json.
  • [2024-10-24] SAM 2++ model and part of Tracking-Any-Granularity dataset are released. Check out the project page for more details.

Dataset Summary

Tracking-Any-Granularity (TAG) is a comprehensive dataset for training our unified model, termed Tracking-Any-Granularity (TAG), with annotations across three granularities: segmentation masks, bounding boxes, and key points.

Dataset Description

Our dataset includes a wide range of video sources, demonstrating strong diversity and serving as a solid benchmark for evaluating tracking performance. Each video sequence is annotated with 18 attributes representing different tracking challenges, which can appear simultaneously in the same video. Common challenges include motion blur, deformation, and partial occlusion, reflecting the dataset’s high difficulty. Most videos contain multiple attributes, indicating the dataset’s coverage of complex and diverse tracking scenarios.

TAG dataset

Benchmark Results

We evaluated many representative trackers on the valid and test splits of our dataset:

video object segmentation

Model π’₯ & β„± π’₯ β„± π’₯ & β„± π’₯ β„±
STCN 70.4 65.9 75 76.2 72.2 80.2
AOT-SwinB 78.1 73.1 83.2 80.9 76.4 85.4
DeAOT-SwinB 79.6 74.8 84.4 81.6 77.3 85.9
XMem 74.4 70.1 78.6 75.7 71.8 79.6
DEVA 77.9 73.1 82.6 82.1 78.0 86.1
Cutie-base+ 79.0 75.0 83.0 83.8 80.0 87.7
Cutie-base+ w/MEGA 80.3 76.5 84.2 84.9 81.3 88.5
OneVOS 80.1 75.2 85.1 81 76.5 85.4
OneVOS w/MOSE 79.3 74.3 84.3 82.4 78 86.7
JointFormer 76.6 72.8 80.5 79.1 75.5 82.7
SAM2++ 87.4 84.2 90.7 87.9 84.9 90.9

single object tracking

Model AUC P_Norm P AUC P_Norm P
OSTrack 74.8 84.4 72.7 69.7 78.8 69.9
SimTrack 71.1 80.5 68.1 64.1 72.4 60.5
MixViT w/ConvMAE 72.1 80.9 70.5 69.7 78.2 70.2
DropTrack 76.8 86.9 74.4 71.1 80.5 72.1
GRM 73.1 82.3 71.4 69.1 77.4 69.1
SeqTrack 77.0 85.8 76.1 69.8 79.4 71.5
ARTrack 76.8 85.8 75.7 71.1 78.7 70.9
ARTrack-V2 76.3 85.5 74.3 71.8 79.5 71.9
ROMTrack 75.6 85.4 73.7 71.3 80.8 72.8
HIPTrack 78.2 88.5 76.6 71.4 81 72.5
LoRAT 75.1 84.8 74.4 70.5 79.7 68.7
SAM2++ 80.7 89.7 77.8 78 85.7 81.5

point tracking

Model Acc Acc
pips 19.0 19.8
pips++ 20.9 23.1
CoTracker 23.3 22.3
CoTracker3 29.6 29.1
TAPTR 23.7 23.8
TAPIR 21.3 24.6
LocoTrack 25.2 30.2
Track-On 24.8 25.8
SAM2++ 35.3 37.7

Dataset Structure 

<ImageSets>
β”‚
β”œβ”€β”€ valid.txt
β”œβ”€β”€ test.txt

<valid/test.tar.gz>
β”‚
β”œβ”€β”€ Annotations
β”‚ β”‚ 
β”‚ β”œβ”€β”€ <video_name_1>
β”‚ β”‚ β”œβ”€β”€ 00000.png
β”‚ β”‚ β”œβ”€β”€ 00001.png
β”‚ β”‚ └── ...
β”‚ β”‚ 
β”‚ β”œβ”€β”€ <video_name_2>
β”‚ β”‚ β”œβ”€β”€ 00000.png
β”‚ β”‚ β”œβ”€β”€ 00001.png
β”‚ β”‚ └── ...
β”‚ β”‚ 
β”‚ β”œβ”€β”€ <video_name_...>
β”‚
β”œβ”€β”€ Points
β”‚ β”‚ 
β”‚ β”œβ”€β”€ <video_name_1>.npz
β”‚ β”œβ”€β”€ <video_name_2>.npz
β”‚ β”œβ”€β”€ <video_name_...>.npz
β”‚
β”œβ”€β”€ Boxes
β”‚ β”‚ 
β”‚ β”œβ”€β”€ <video_name_1>.txt
β”‚ β”œβ”€β”€ <video_name_2>.txt
β”‚ β”œβ”€β”€ <video_name_...>.txt
β”‚
β”œβ”€β”€ Visible
β”‚ β”‚ 
β”‚ β”œβ”€β”€ <video_name_1>.txt
β”‚ β”œβ”€β”€ <video_name_2>.txt
β”‚ β”œβ”€β”€ <video_name_...>.txt
β”‚ 
└── JPEGImages
  β”‚ 
  β”œβ”€β”€ <video_name_1>
  β”‚ β”œβ”€β”€ 00000.jpg
  β”‚ β”œβ”€β”€ 00001.jpg
  β”‚ └── ...
  β”‚ 
  β”œβ”€β”€ <video_name_2>
  β”‚ β”œβ”€β”€ 00000.jpg
  β”‚ β”œβ”€β”€ 00001.jpg
  β”‚ └── ...
  β”‚ 
  └── <video_name_...>

BibTeX

If you find Tracking-Any-Granularity helpful to your research, please consider citing our papers.

@article{zhang2025sam2trackinggranularity,
  title={SAM 2++: Tracking Anything at Any Granularity},
  author={Jiaming Zhang and Cheng Liang and Yichun Yang and Chenkai Zeng and Yutao Cui and Xinwen Zhang and Xin Zhou and Kai Ma and Gangshan Wu and Limin Wang},
  journal={arXiv preprint arXiv:2510.18822},
  url={https://arxiv.org/abs/2510.18822},
  year={2025}
}

License

Tracking-Any-Granularity is licensed under a CC BY-NC-SA 4.0 License. The data of Tracking-Any-Granularity is released for non-commercial research purpose only.