This is a PyTorch implementation of ASpanFormer for ECCV'22 paper, “ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformer”, and can be used to reproduce the results in the paper.

This work focuses on detector-free image matching. We propose a hierarchical attention framework for cross-view feature update, which adaptively adjusts attention span based on region-wise matchability.

This repo contains training, evaluation and basic demo scripts used in our paper.

A large part of the code base is borrowed from the LoFTR Repository under its own separate license, terms and conditions. The authors of this software are not responsible for the contents of third-party websites.

conda env create -f environment.yaml
conda activate ASpanFormer

Download model weights from here

Extract weights by

tar -xvf weights_aspanformer.tar

A demo to match one image pair is provided. To get a quick start,

cd demo
python demo.py

Please follow the training doc for data organization

cd scripts/reproduce_test
bash indoor.sh

Similar results as below should be obtained,

'auc@10': 0.46640095171012563,
'auc@20': 0.6407042320049785,
'auc@5': 0.26241231577189295,
'prec@5e-04': 0.8827665604024288,
'prec_flow@2e-03': 0.810938751342228

cd scripts/reproduce_test
bash outdoor.sh

Similar results as below should be obtained,

'auc@10': 0.7184113573584142,
'auc@20': 0.8333835724453831,
'auc@5': 0.5567622479156181,
'prec@5e-04': 0.9901741341790503,
'prec_flow@2e-03': 0.7188964321862907

cd scripts/reproduce_train
bash indoor.sh

cd scripts/reproduce_train
bash outdoor.sh

If you find this project useful, please cite:

@article{chen2022aspanformer,
  title={ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformer},
  author={Chen, Hongkai and Luo, Zixin and Zhou, Lei and Tian, Yurun and Zhen, Mingmin and Fang, Tian and McKinnon, David and Tsin, Yanghai and Quan, Long},
  journal={European Conference on Computer Vision (ECCV)},
  year={2022}
}