cvg/glue-factory
Training library for local feature detection and matching
What it solves
Glue Factory provides a unified framework for training and evaluating deep neural networks that extract and match local visual features (points and lines) across different images. It simplifies the process of reproducing state-of-the-art matching models and testing them against standard computer vision benchmarks.
How it works
The library implements a two-stage training pipeline: first pre-training on large synthetic homography datasets and then fine-tuning on realistic datasets like MegaDepth. It supports a variety of local feature extractors (such as SuperPoint, SIFT, ALIKED, and DISK) and integrates them with deep matchers like LightGlue, GlueStick, and SuperGlue. The framework includes built-in auto-downloaders for datasets and tools for evaluating performance using robust estimators like PoseLib and OpenCV.
Who it’s for
It is designed for researchers and developers working in computer vision, specifically those focused on image matching, 3D reconstruction, and visual localization.
Highlights
- Comprehensive Model Support: Supports training and evaluation for LightGlue, GlueStick, and SuperGlue, with evaluation support for LoFTR.
- Benchmark Integration: Built-in evaluation scripts for HPatches, MegaDepth-1500, Scannet-1500, and ETH3D.
- Flexible Configuration: Uses OmegaConf for easy command-line overrides of training and evaluation parameters.
- Visual Inspection: Includes a viewer to visually inspect and compare matches between different methods on benchmark datasets.