CATs: Cost Aggregation Transformers for Visual Correspondence NeurIPS'21

For more information, check out the paper on [arXiv].

Check out our journal extension! It will be appeared at TPAMI, but currenclty available at: [arXiv]. Also, the code implementation is available at : https://github.com/KU-CVLAB/CATs-PlusPlus

Network

Our model CATs is illustrated below:

Environment Settings

git clone https://github.com/SunghwanHong/CATs
cd CATs

conda create -n CATs python=3.6
conda activate CATs

pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install -U scikit-image
pip install git+https://github.com/albumentations-team/albumentations
pip install tensorboardX termcolor timm tqdm requests pandas

Evaluation

• Download pre-trained weights on Link
• All datasets are automatically downloaded into directory specified by argument datapath

Result on SPair-71k: (PCK 49.9%)

  python test.py --pretrained "/path_to_pretrained_model/spair" --benchmark spair

Result on SPair-71k, feature backbone frozen: (PCK 42.4%)

  python test.py --pretrained "/path_to_pretrained_model/spair_frozen" --benchmark spair

Results on PF-PASCAL: (PCK 75.4%, 92.6%, 96.4%)

  python test.py --pretrained "/path_to_pretrained_model/pfpascal" --benchmark pfpascal

Results on PF-PACAL, feature backbone frozen: (PCK 67.5%, 89.1%, 94.9%)

  python test.py --pretrained "/path_to_pretrained_model/pfpascal_frozen" --benchmark pfpascal

Acknowledgement

We borrow code from public projects (huge thanks to all the projects). We mainly borrow code from DHPF and GLU-Net.

BibTeX

If you find this research useful, please consider citing:

@inproceedings{cho2021cats,
  title={CATs: Cost Aggregation Transformers for Visual Correspondence},
  author={Cho, Seokju and Hong, Sunghwan and Jeon, Sangryul and Lee, Yunsung and Sohn, Kwanghoon and Kim, Seungryong},
  booktitle={Thirty-Fifth Conference on Neural Information Processing Systems},
  year={2021}
}