Correspondence Transformers With Asymmetric Feature Learning and Matching Flow Super-Resolution（ACTR）

This is the official code for ACTR implemented with PyTorch.

Environment Settings

git clone https://github.com/YXSUNMADMAX/ACTR
cd ACTR
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

• Result on SPair-71k: python test.py --datapath "/path_to_dataset" --pretrained "/path_to_pretrained_model/spair" --benchmark spair

• Results on PF-PASCAL: python test.py --datapath "/path_to_dataset" --pretrained "/path_to_pretrained_model/pfpascal" --benchmark pfpascal

Acknowledgement

We borrow code from public projects (Thanks a lot !!!). We mainly borrow code from CATs.

BibTeX

If you find this research useful, please consider citing:

@inproceedings{sun2023correspondence,
  title={Correspondence Transformers With Asymmetric Feature Learning and Matching Flow Super-Resolution},
  author={Sun, Yixuan and Zhao, Dongyang and Yin, Zhangyue and Huang, Yiwen and Gui, Tao and Zhang, Wenqiang and Ge, Weifeng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={17787--17796},
  year={2023}
}