Track Anything Annotate

Track Anything Annotate is a flexible tool for tracking, segmentation, and annotation of videos. It allows creating datasets from videos in YOLO and COCO formats. It is based on Segment Anything 2 and allows specifying any objects for tracking and segmentation.

Track Anything Annotate is an open-source project developed with community support. Your feedback and suggestions for improvement are incredibly valuable to us. We're especially interested in learning how you use track-anything-annotate in your projects. If you're willing to share your experiences or use cases (even small, experimental ones), please email us at [email protected]. This will not only help us improve the tool but may also serve as a basis for mentioning your project in future publications.

Read this in other languages: English | Русский

---

Quick Start

🛠️ Installation via uv

# For CUDA  
uv sync --extra cu129

# For CPU  
uv sync --extra cpu

Download Models

uv run checkpoints/download_models.py

Run the Demo (Access at http://127.0.0.1:8080 )

gradio demo.py

Dataset Creation

Currently, it only works with one class.

Type of saving

• yolo
• coco

uv run annotation.py --video-path path_to_video --names-class name_class --type-save yolo

Instructions for creating a dataset via json

uv run annotate_json.py --video-path path_to_video --json-path path_to_json --type-save yolo

---

Installation via pip

Install for CUDA Windows

# Clone the repository:
git clone https://github.com/lnikioffic/track-anything-annotate.git
cd track-anything-annotate

# Install dependencies:
pip install -r requirements.txt --index-url https://download.pytorch.org/whl/cu129

# Download Models
python checkpoints/download_models.py

# Run the gradio demo.
python demo.py

# Dataset Creation
python annotation.py

Install Linux

# Clone the repository:
git clone https://github.com/lnikioffic/track-anything-annotate.git
cd track-anything-annotate

# Install dependencies:
pip install -r requirements.txt

# Download Models
python checkpoints/download_models.py

# Run the gradio demo.
python demo.py

# Dataset Creation
python annotation.py --video-path path_to_video --names-class name_class --type-save yolo

# or
python annotate_json.py --video-path path_to_video --json-path path_to_json --type-save yolo

🗺️ Roadmap and Improvements

• Tracking single class export in YOLO.
• New export formats: Adding support for COCO JSON.
• Multi-class annotation: Ability to track multiple different classes.
• New export formats: Adding support for Pascal VOC XML.
• Image annotation: Ability to collect and annotate your own dataset based on images.

📚 Citation

If you use this project in your work, please cite the paper:

@article{ivanov2025track,
    title={Track Anything Annotate: Video annotation and dataset generation of computer vision models},
    author={Ivanov, Nikita and Klimov, Mark and Glukhikh, Dmitry and Chernysheva, Tatiana and Glukhikh, Igor},
    journal={arXiv preprint arXiv:2505.17884},
    year={2025}
}