LGFCTR
June 11, 2024 ยท View on GitHub
Code for "LGFCTR: Local and Global Feature Convolutional Transformer for Image Matching"
Data preparation
MegaDepth dataset
Since the preprocessed undistorted MegaDepth dataset provided in D2-Net has been not available, we use the original MegaDepth dataset
MegaDepth indices
Download and unzip MegaDepth indices following LoFTR
Build dataset symlinks
Symlink the datasets and indices to the data directory following LoFTR
Pretrained model
We provide the outdoor weights of LGFCTR in the Google Drive.
Requirements
Following LoFTR
Please follow LoFTR to prepare the environment.
Pip by yourselves
In addition, you can also install the requirements by yourselves. More specifically, we use pytorch==1.9.1+cu111, pytorch-lightning==1.3.5, opencv-python==4.5.5.64, torchmetrics==0.6.0 and kornia==0.6.11. Other requirements can be installed easily by pip.
Pip with our enviornment.yaml
conda env create -f environment.yaml
conda activate lgfctr
Reproduce
Training
You can reproduce the training by
sh scripts/reproduce_train/outdoor_ds.sh
Evaluation
You can reproduce the evaluation on MegaDepth dataset by
sh scripts/reproduce_test/outdoor_ds.sh
Demos
Visualize a single pair of images
We provide a demo for visualizing a single pair of images. You can specify img_path0 and img_path1 for your images, save_dir for your save directory, topk for the number of matches shown, img_resize for resized longer dimension, and is_original for outputing the original images.
cd vis
python vis_single_pair.py --img_path0 your_img_path0 --img_path1 your_img_path1 --save_dir your_save_dir --topk 1000 --img_resize 640 --is_original True
Visualize multi-scale attention weights
We provide a demo for visualizing multi-scale attention weights of a single pair of images. Besides arguments mentioned above, you can specify dpi for the dpi of outputs, and change the Line 41 to specify which index of resolutions and CTR for visualizations.
python vis_attention.py
Acknowledgements
This repository was developed from LoFTR, and we are grateful for their implementations.
Citation
If you find this code useful for your research, please use the following BibTeX entry.
@article{zhong2023lgfctr,
title={LGFCTR: Local and Global Feature Convolutional Transformer for Image Matching},
author={Zhong, Wenhao and Jiang, Jie},
journal={arXiv preprint arXiv:2311.17571},
year={2023}
}