glue-factory

Why we included this project

Most local feature matching models arrive as pre-trained black boxes, which is a problem when your pipeline runs on imagery that looks nothing like the training data. Glue Factory, from the CVG group, is the codebase used to train and evaluate point and line matchers like LightGlue and GlueStick. Instead of accepting whatever weights ship, you can reproduce their training or retrain them on your own local features and datasets. Evaluation is built into the workflow rather than added at the end, with standard benchmarks such as HPatches and MegaDepth-1500 running on cached results so comparisons stay reproducible. The core code is Apache-2.0 and models and datasets download automatically, so going from clone to a custom-trained matcher takes little setup.