Non-Maximum Suppression Algorithm

Non-maximum suppression is an algorithm for thinning a cluster of redundant nearby detections down to the single best one, applied wherever an underlying detection method produces more responses than the number of genuine features actually present. In edge detection, it examines the gradient magnitude at each pixel along the direction of the gradient and keeps that pixel only if its value is the local maximum among its immediate neighbors in that direction, discarding the others so a real edge is reported as a single pixel-wide line rather than a blurred band; in modern object detection, the same underlying idea is applied to a set of overlapping candidate bounding boxes, keeping the highest scoring box in each cluster and discarding every other box that overlaps it beyond a set threshold. John Canny's 1986 edge detector popularized the technique for edge thinning, and later object detection systems adopted the same principle to turn a dense set of overlapping candidate boxes into one clean detection per object.