Local Feature Matching

Local feature matching identifies corresponding image locations despite changes in viewpoint, scale, rotation, lighting, or sensing modality. Systems may detect keypoints and compare descriptors, or use detector-free dense and semi-dense methods, often with transformers, coarse-to-fine refinement, and geometric filtering. They are evaluated on tasks such as homography and pose estimation, visual localization, structure from motion, and SLAM, using benchmarks including HPatches, MegaDepth, ScanNet, and WxBS. A central tension is achieving robust matches across difficult conditions while controlling runtime and memory; standard benchmark accuracy also may not predict performance under domain shifts or deployment constraints.