Glue Factory

Glue Factory is a computer vision library dedicated to the training and evaluation of deep neural networks for local feature extraction and matching. Developed by the Computer Vision and Geometry Group (CVG), it serves as a research playground for state-of-the-art models like LightGlue and GlueStick.

Key Features

• Reproducibility: Scripts and configurations to reproduce training for SOTA models like LightGlue and GlueStick.
• Modular Design: Flexible TwoViewPipeline architecture allowing mix-and-match of extractors, matchers, outlier filters, and solvers.
• Standard Benchmarks: Integrated evaluation pipelines for HPatches, MegaDepth-1500, ScanNet-1500, and ETH3D.
• Auto-downloading: Models and datasets are automatically downloaded upon first use.
• Training Loop: Robust training script supporting distributed training, mixed precision, and experiment tracking.

Supported Models

Glue Factory supports a variety of local feature workflows:

• Extractors: SuperPoint, SIFT, ALIKED, DISK, Key.Net, etc.
• Matchers: LightGlue, GlueStick, SuperGlue, Nearest Neighbor, RoMA.
• Line Features: DeepLSD, LSD, Wireframe extraction.

License

The code and trained models in Glue Factory are released with an Apache-2.0 license. Note that third-party models in gluefactory_nonfree may follow their own restrictive licenses.