LightGlue is a deep neural network for matching sparse local features across image pairs.
The tool solves the problem of efficiently matching keypoints and descriptors between two images. It uses an adaptive pruning mechanism that adjusts both network depth and width based on the difficulty of each image pair, allowing it to compute quickly on easy matches while reducing computational complexity for challenging ones. The network takes keypoints and descriptors extracted from each image as input and returns indices of corresponding points between them.
Developers should choose LightGlue if they need fast, accurate feature matching for computer vision tasks like pose estimation or image registration. The tool works with multiple local feature extractors, offering pretrained weights compatible with SuperPoint, DISK, ALIKED, and SIFT. For those seeking integration simplicity, the tool is available through Hugging Face Transformers, enabling inference with minimal Python code. The repository provides training and evaluation capabilities through a companion library for those wanting to train custom models.
The project maintains active engagement with the research community through accessible resources including a Colab notebook and interactive demo. Development focuses on practical usability, as evidenced by integration into a major machine learning framework and provision of multiple entry points for both inference and training workflows.
