LightGlue

Why we included this project

Anyone building 3D reconstruction, visual localization, or SLAM systems will recognize the problem LightGlue solves: given two images with keypoints and descriptors already extracted by a detector such as SuperPoint or DISK, it returns the correspondences between them. An adaptive mechanism skips work on easy image pairs and spends extra compute only where matching is genuinely hard, which keeps accuracy close to heavier matchers while cutting inference time. That speed matters on robots and in mobile pipelines, where latency budgets are tight. Pretrained weights ship for several common feature extractors, and the Hugging Face Transformers integration means a few lines of Python are enough to run it without custom build steps. If your pipeline already produces sparse local features, this saves you from writing a matcher from scratch.