LightGlueStick
Fast and Robust Joint Point-Line Matching

LightGlueStick adaptively adjusts its depth based on image difficulty, exiting after the 4th layer for the top image pair (easy) and the 7th layer for the bottom pair (hard). These pairs are processed in just 27ms and 42ms, respectively.

Installation

To install the software follow these instructions (tested on Ubuntu)

git clone https://github.com/aubingazhib/LightGlueStick.git
cd LightGlueStick
python -m venv venv
source venv/bin/activate
pip install .

LightGlueStick Inference

You can match features with:

python -m lightgluestick.run -img1 assets/img1.jpg -img2 assets/img2.jpg

To enable depth adaptivity, simply provide a depth confidence value:

python -m lightgluestick.run -img1 assets/img1.jpg -img2 assets/img2.jpg --depth_confidence 0.95

When depth adaptivity is enabled, the model will perform an early exit based on the specified confidence threshold and print the corresponding early exit layer.

Training 🏋️

The training code is available in a separate repository, GlueFactory. Within GlueFactory, you can not only train LightGlueStick, but also other deep matchers such as LightGlue and GlueStick, use multiple feature extractors, line extractors, robust estimators, as well as run evaluations on multiple benchmarks.

Licence 📜

Our code is licenced under Apache-2.0 license. However, bear in mind that it uses a SuperPoint backbone that has a non-commercial licence. Therefore, the overall system is non-commercial 😞.

BibTeX Citation

Please consider citing the following papers if you found this code useful:

@InProceedings{Ubingazhibov_2025_ICCV,
    author    = {Ubingazhibov, Aidyn and Pautrat, Rémi and Suárez, Iago and Liu, Shaohui and Pollefeys, Marc and Larsson, Viktor},
    title     = {LightGlueStick: a Fast and Robust Glue for Joint Point-Line Matching},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
    month     = {October},
    year      = {2025},
    pages     = {7244-7254}
}