image keypoint

An image keypoint is a small, distinctive spot in a picture that a computer can find again reliably — even after the scene shifts, the camera moves, or the lighting changes. Think of it as a visual landmark. Just as you navigate a city by remembering corners and signs rather than every brick, a robot picks out a sparse set of standout points instead of trying to track every pixel, which would be hopeless because most pixels look just like their neighbors.

What makes a good keypoint is being unmistakable from every direction. A point in the middle of a blank white wall is useless: slide your window of view in any direction and it looks identical, so you cannot tell if anything moved. A point along a straight edge is only half-useful: you can tell when it shifts across the edge, but sliding along the edge looks the same. A corner, where two edges meet, is the prize — nudge it any way at all and the local pattern changes, so the robot can locate it precisely. This is why classic corner detectors are the workhorses of keypoint finding, and why the two prized traits are repeatability (the same physical spot is found again in a new image) and distinctiveness (it does not get confused with other spots).

Keypoints are the anchors of much of robot vision. By finding the same physical points across two photos, a robot can stitch a panorama, track its own motion frame to frame, build a map while localizing itself in it, or estimate the depth of objects from a stereo pair. Each keypoint is usually paired with a feature descriptor — a compact fingerprint of its surroundings — so the same point can be recognized and matched between different views.