Code for ECCV-ILR 2024 Workshop paper Leveraging Object Priors for Point Tracking
Create conda environment for this code base:
conda create -n mask_pips python=3.8
conda activate mask_pips
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
pip install -r requirements.txt
Use export_mp4_dataset.py file to generate clips from the PointOdyssey training set following Pips++ and then run python train.py to start training.
To evaluate the performance on the datasets reported in the paper, use testing scripts in the pips2 directory with the saved model.
For TAP-VID-DAVIS: pip2/test_on_tap.py
For CroHD: pips2/test_on_cro.py
For PointOdyssey: pips2/test_on_pod.py
To replicate the performance from the paper use our trained weights for the reference model. For TAP-VID-DAVIS, we load full sequence into memory at once, for others we use S=36.