Corners and Keypoints: Spots That Stay Identifiable

An edge tells you a boundary exists, but a long straight edge looks the same everywhere along its length, so it cannot tell you where you are. A corner does not have that problem. At a corner, brightness changes sharply in two different directions at once, so the local pattern is distinctive and hard to confuse with a nearby patch. This is the property that makes corners useful as keypoints: small image locations that can be found reliably and matched between two photographs of the same scene. The reason corners survive comparison is that they are stable under several transformations. Shift the image slightly and the corner is still a corner, just moved. Rotate the image and the two-direction change is still there. Scale it moderately and the corner remains locally distinctive, though very large scale changes eventually break it. A flat region fails all of these tests because every patch looks identical, and a straight edge fails because patches along it are indistinguishable from one another. Keypoint detectors therefore look for locations where the local pattern changes strongly in more than one direction, then attach a small descriptor — a compact summary of the surrounding patch — so that two keypoints can be compared numerically. Matching two images then becomes a problem of finding keypoints whose descriptors agree.