This repository hosts the batch inference code (inference_pose.py and inference_point.py) of LightGlue.
The architecture is based on adaptive pruning techniques, in both network width and depth - check out the paper for more details.
inference_point.py: Save the corresponding feature point pairs between two frames inferred by lightglue into a. txt file, with four numbers for each line being x1 , y1 ; X2, y2.
inference_pose.py: It only calculates the camera pose based on the feature points on the rgb image. After my own verification, the calculated pose seems to be not very accurate. If there is rgb-d information, "inference_point.py" is used to calculate the feature points between frames. Then rgb-d is applied to the 3D point cloud to calculate the pose, such as the simplest least squares method.
There are pretrained weights of LightGlue with SuperPoint and DISK local features.
The data needs to be in the following format:
<data_root> # datadir
├── <scene001>
├── images # RGB images
├── <00000>.jpg
├── <00001>.jpg
├── <00002>.jpg
...
├── <scene002>
├── images # RGB images
├── <00000>.jpg
├── <00001>.jpg
├── <00002>.jpg
...
├── <scene003>
├── images # RGB images
├── <00000>.jpg
├── <00001>.jpg
├── <00002>.jpg
...
- Image Matching WebUI: a web GUI to easily compare different matchers, including LightGlue.
If you use any ideas from the paper or code from this repo, please consider citing:
@inproceedings{lindenberger2023lightglue,
author = {Philipp Lindenberger and
Paul-Edouard Sarlin and
Marc Pollefeys},
title = {{LightGlue: Local Feature Matching at Light Speed}},
booktitle = {ICCV},
year = {2023}
}