PATS: Patch Area Transportation with Subdivision for Local Feature Matching

Project Page | Paper

PATS: Patch Area Transportation with Subdivision for Local Feature Matching
Junjie Ni* Yijin Li*, Zhaoyang Huang, Hongsheng Li, Hujun Bao, Zhaopeng Cui, Guofeng Zhang
CVPR 2023

TODO List

Training script

Download Link

We provide the download link to

Pretrained models trained on MegaDepth and ScanNet, which are labeled as outdoor and indoor, respectively.
MegaDepth pairs and scenes (placed in a folder named megadepth_parameters).
The demo data, which is a sequence of images captured from near to far.

Run PATS

Installation

conda env create -f environment.yaml
cd setup
python setup.py install
cd ..

Prepare the data and pretrained model

Download from the above link, and place the data and model weights as below:

pats
├── data
│   ├── MegaDepth_v1
│   ├── megadepth_parameters 
│   ├── ScanNet
│   ├── yfcc100M
│   └── demo
└── weights
    ├── indoor_coarse.pt
    ├── indoor_fine.pt
    ├── indoor_third.pt
    ├── outdoor_coarse.pt
    ├── outdoor_fine.pt
    └── outdoor_third.pt

Evaluate on MegaDepth/YFCC/ScanNet dataset

python evaluate.py configs/test_megadepth.yaml
python evaluate.py configs/test_yfcc.yaml
python evaluate.py configs/test_scannet.yaml

Run the demo

python demo.py configs/test_demo.yaml

Citation

If you find this code useful for your research, please use the following BibTeX entry.

@inproceedings{pats2023,
  title={PATS: Patch Area Transportation with Subdivision for Local Feature Matching},
  author={Junjie Ni, Yijin Li, Zhaoyang Huang, Hongsheng Li, Hujun Bao, Zhaopeng Cui, Guofeng Zhang},
  booktitle={The IEEE/CVF Computer Vision and Pattern Recognition Conference (CVPR)},
  year={2023}
}

Acknowledgements

We would like to thank the authors of SuperGlue and LoFTR for open-sourcing their projects.

zju3dv/pats