non-max suppression (NMS)

Collapsing a detector's many overlapping candidate boxes for the same object down to one, keeping the highest-confidence box.

Non-max suppression (NMS) removes redundant detections by, for each object class, repeatedly keeping the highest-confidence remaining box and discarding every other box whose IoU with it exceeds a threshold -- collapsing many overlapping candidate boxes produced for the same object down to one, while leaving separate, low-overlap detections of genuinely distinct objects alone.

Examples

Three overlapping boxes around the same dog, with confidences $0.95$ , $0.80$ , and $0.60$ , collapse to just the $0.95$ box once the other two are suppressed for exceeding the IoU threshold against it.