Requirements

• CUDA 12.0
• Python 3.6 (or later)
• torch 2.1.0
• torchvision 0.16.0
• opencv_python 4.10.0.84
• kornia 0.7.0

Data preparation

Install

# create virtual environment

conda create -n DenseAffine python=3.10

conda activate DenseAffine

# install DenseAffine requirements

cd DenseAffine

pip install -r requirements.txt

Get start

Download the weights at https://pan.baidu.com/s/1EdsAqKJ1HKVS5uLPMAxDSQ?pwd=idw2 password: idw2 or https://drive.google.com/file/d/1W82QJ5lgrsVQql30k_NZWZE5YH9lhgaE/view?usp=drive_link. Put it in weights folder.

python demo/affine_feature_estimator.py

The matching results of the images are saved in the results folder.

Download the KITTI sequens 04 for relative pose estimation. The URL of the dataset is https://pan.baidu.com/s/1EdsAqKJ1HKVS5uLPMAxDSQ?pwd=idw2 password: idw2 or https://drive.google.com/file/d/1W82QJ5lgrsVQql30k_NZWZE5YH9lhgaE/view?usp=drive_link.

python demo/affine_feature_relative_pose.py

Acknowledgements

We have used code and been inspired by https://github.com/Parskatt/dkm, https://github.com/laoyandujiang/S3Esti, and https://github.com/ducha-aiki/affnet, https://github.com/DagnyT/hardnet, https://github.com/Reagan1311/LOCATE, https://github.com/danini/graph-cut-ransac. Sincere thanks to these authors for their codes.

cite

If you find this work useful, please cite:

@inproceedings{sun2025learning,
  title={Learning affine correspondences by integrating geometric constraints},
  author={Sun, Pengju and Guan, Banglei and Yu, Zhenbao and Shang, Yang and Yu, Qifeng and Barath, Daniel},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
  pages={27038--27048},
  year={2025}
}