Descriptor matching and Lowe's ratio test

WHAT THIS TESTS The interviewer checks whether you understand that raw nearest-neighbor matching is noisy and how a relative test cleanly removes ambiguous correspondences.

A GOOD ANSWER COVERS Matching starts by comparing descriptors. For each keypoint descriptor in the first image you search the second image for the descriptor with the smallest distance, Euclidean for float descriptors like SIFT, Hamming for binary descriptors like ORB. Taking the single nearest neighbor produces many false matches, especially in repetitive or textureless regions. Lowe's ratio test improves this by finding the two nearest neighbors and comparing the distance to the closest with the distance to the second closest. If the closest is not substantially nearer than the second, the correspondence is ambiguous and rejected; only matches where the ratio of closest to second-closest distance is below a threshold around 0.7 to 0.8 are kept.

COMMON WRONG ANSWERS Using a fixed absolute distance threshold, which fails because good distances vary by region and descriptor. Keeping every nearest neighbor. Misstating the ratio direction, the closest must be much smaller than the second-closest to pass.

LIKELY FOLLOW-UPS Why not an absolute threshold. How does cross-checking or RANSAC complement the ratio test. What threshold value is typical and what does tightening it do.

ONE CONCRETE EXAMPLE Imagine matching descriptors across two photos of a building with many similar windows. A keypoint on one window corner has a true correspondence on the matching window, but several other windows look nearly identical. Without the ratio test, the nearest neighbor might be the wrong window by a tiny margin, producing a confident but false match. With the ratio test you find the two best candidates; because multiple windows are near-identical, the closest and second-closest distances are almost equal, so their ratio is near one, above the 0.75 threshold, and the match is discarded as ambiguous. For a truly distinctive feature, say a unique sign, the correct match is far closer than any other, the ratio is small, and it passes. The test thus exploits distinctiveness: a reliable match must stand out from all alternatives, not merely be the closest by a hair. RANSAC is then often applied to enforce geometric consistency among the surviving matches.