Beyond Global Cues: Unveiling the Power of Fine Details in Image Matching

Paper

Introduction

The official code of our ICME 2024 paper.

Dependencies

• Python 3 = 3.8.12
• PyTorch = 1.9.1

Run the demo of our method

python  ./demo.py

The default will resize images to 640x640.

Weights

Weights can be found here

Data preparation

For details of data preparation, please refer to this.

Evaluation on MegaDepth

For evaluation, the images are resized to 640x640.

# with shell script
bash ./scripts/reproduce_test/outdoor_ds.sh

you can reproduce results in paper, namely,

AUC@5    AUC@10  AUC@20  Prec
48.77    65.20   77.80   97.29 

Evaluation on YFCC100M

For the script for evaluation on YFCC100M, you can refer to superglue.

When the RANSAC threshold is set to 1.0, you can reproduce results in paper.

After preparing the code for release, we tuned the hyperparameters and made some improvements. For example, when the RANSAC threshold is set to 0.3, the performance of our proposed method can be boosted again:

Evaluation Results (mean over 15 pairs):
AUC@5    AUC@10  AUC@20  Prec
42.55    61.69   76.75   93.87 

About training

If you want to train yourself, please

python  ./train.py

Due to the limited computational resources, the images are resized to 640x640 during training. You are suggested to use the larger sizes, for example, 832x832 for training since many works have proven that the large sizes can improve the performance.

Citation

If you use any ideas from the paper or code from this repo, please consider cite our paper through IEEE Xplore.