CATs++: Boosting Cost Aggregation with Convolutions and Transformers (TPAMI'22)

For more information, check out the paper on [arXiv]. Also check out project page here [Project Page]

Network

Our model is illustrated below:

Environment Settings

git clone https://github.com/KU-CVLAB/CATs-PlusPlus.git
cd CATs-PlusPlus

conda env create -f environment.yml

Evaluation

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

Result on SPair-71k:

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

Results on PF-PASCAL:

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

Results on PF-WILLOW:

  python test.py --pretrained "/path_to_pretrained_model/pfpascal" --benchmark pfwillow --thres {bbox|bbox-kp}

Acknowledgement

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

BibTeX

If you find this research useful, please consider citing:

@article{cho2022cats++,
  title={CATs++: Boosting Cost Aggregation with Convolutions and Transformers},
  author={Cho, Seokju and Hong, Sunghwan and Kim, Seungryong},
  journal={arXiv preprint arXiv:2202.06817},
  year={2022}
}