Feature Matching

Finds pairs of feature points across two images that correspond to the same physical point.

Feature matching finds pairs of features between two images (or two point clouds) that correspond to the same physical point. The classic pipeline: extract feature points and descriptors on each image (such as SIFT or ORB); find nearest neighbors by descriptor distance, using Euclidean distance for floating-point descriptors and Hamming distance for binary descriptors like ORB’s; then apply the ratio test proposed by David Lowe to discard ambiguous matches — if the distance ratio between the nearest and second-nearest neighbor is too close to 1, the match is dropped; finally, fit a geometric model with RANSAC to reject outliers that don’t fit. Recently, SuperGlue and LightGlue use attention networks to learn matching directly, and are more reliable under large viewpoint and lighting changes. Matching results underlie camera pose estimation, triangulation, visual SLAM, and structure from motion.

ExampleETH Zurich’s LightGlue (ICCV 2023) improves on SuperGlue by adaptively stopping computation early based on how easy or hard an image pair is, matching noticeably faster when two images overlap a lot and look similar.

Also called

Keypoint Matching, Correspondence Matching

Related

Feature Points · Random Sample Consensus · SuperPoint / SuperGlue / LightGlue · Structure from Motion · Visual SLAM · Epipolar Geometry

Sources

Wikipedia: Scale-invariant feature transform（含 Lowe 比值检验） (Chinese)
LightGlue: Local Feature Matching at Light Speed