Official Implementation for paper Track Everything Everywhere Fast and Robustly, ECCV 2024.

Yunzhou Song ^1*, Jiahui Lei ^1*, Ziyun Wang ¹, Lingjie Liu ¹, Kostas Daniilidis ^1,2
1University of Pennsylvania, ²Archimedes, Athena RC, ^*equal contribution

The code is tested with python=3.8 and torch=2.2.0+cu118.

git clone --recurse-submodules https://github.com/TimSong412/OmniTrackFast
cd OmniTrackFast
conda create -n omni python=3.8
conda activate omni
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

1. Please refer to the preprocessing instructions for preparing input data for training. We also provide some processed data that you can download, unzip and directly train on.

2. With processed input data, run the following command to start training:

   python train.py --config configs/default.txt --data_dir {sequence_directory}
   

   You can also only train on flow supervision without long-term matching by

   python train.py --config configs/nomatch.txt --datadir {sequence_directory}
   

   You can view visualizations on tensorboard by running tensorboard --logdir logs/. By default, the script trains 100k iterations which takes less than 1 hour on a 2080Ti GPU.

1. Please download the benchmark annotations used by OmniMotion here and our checkpoints form here. Please unzip the benchmark directory into the dataset directory as dataset/tapvid_XXXX/annotations
2. Run the following command to run checkpoint evaluation:

   python run_eval.py
   

1. Any video sequence less than 20 frames should be expanded to over 20 or longer, otherwise the NVP blocks are too many to fit the sequence.
2. We use torch.jit and torch.autograd.Function in the non-linear NVP blocks to manually accelerate the forward and backward process. You may disable the jit configure if any runtime error occurs.
3. Generally, a video sequence of 100 frames consumes 5GB on a 2080Ti GPU for optimization.

@article{song2024track,
  title={Track Everything Everywhere Fast and Robustly},
  author={Song, Yunzhou and Lei, Jiahui and Wang, Ziyun and Liu, Lingjie and Daniilidis, Kostas},
  journal={arXiv preprint arXiv:2403.17931},
  year={2024}
}