Local All-Pair Correspondence for Point Tracking

Seokju Cho¹ · Jiahui Huang² · Jisu Nam¹ · Honggyu An¹ · Seungryong Kim¹ · Joon-Young Lee²

¹Korea University ²Adobe Research

ECCV 2024

LocoTrack is an incredibly efficient model, enabling near-dense point tracking in real-time. It is 6x faster than the previous state-of-the-art models.

📰 News

2024-07-22: LocoTrack is released.
2024-08-03: PyTorch inference and training code released.
2024-08-05: Interactive demo released.

Please stay tuned for an easy-to-use API for LocoTrack, coming soon!

🎮 Interactive Demo

Try our interactive demo on Huggingface. To run the demo locally, please follow these steps:

Install Dependencies: Ensure you have all the necessary packages by running:
```
pip install -r demo/requirements.txt
```
Run the Demo: Launch the interactive Gradio demo with:
```
python demo/demo.py
```

Training and Evaluation

For detailed instructions on training and evaluation, please refer to the README file for your chosen implementation:

JAX Implementation
PyTorch Implementation

Evaluation Dataset Preparation

First, download the evaluation datasets:

# TAP-Vid-DAVIS dataset
wget https://storage.googleapis.com/dm-tapnet/tapvid_davis.zip
unzip tapvid_davis.zip

# TAP-Vid-RGB-Stacking dataset
wget https://storage.googleapis.com/dm-tapnet/tapvid_rgb_stacking.zip
unzip tapvid_rgb_stacking.zip

# RoboTAP dataset
wget https://storage.googleapis.com/dm-tapnet/robotap/robotap.zip
unzip robotap.zip

For downloading TAP-Vid-Kinetics, please refer to official TAP-Vid repository.

Training Dataset Preparation

Download the panning-MOVi-E dataset used for training (approximately 273GB) from Huggingface using the following script. Git LFS should be installed to download the dataset. To install Git LFS, please refer to this link. Additionally, downloading instructions for the Huggingface dataset are available at this link

git clone git@hf.co:datasets/hamacojr/LocoTrack-panning-MOVi-E

📚 Citing this Work

Please use the following bibtex to cite our work:

@article{cho2024local,
  title={Local All-Pair Correspondence for Point Tracking},
  author={Cho, Seokju and Huang, Jiahui and Nam, Jisu and An, Honggyu and Kim, Seungryong and Lee, Joon-Young},
  journal={arXiv preprint arXiv:2407.15420},
  year={2024}
}

🙏 Acknowledgement

This project is largely based on the TAP repository. Thanks to the authors for their invaluable work and contributions.

cvlab-kaist/locotrack