IoU → NMS: The Small Geometry Behind Object Detection

Object detectors often predict several boxes around the same object. Non-Maximum Suppression turns that noisy set into a smaller set of useful detections.

Intersection over Union

IoU measures how much two boxes overlap relative to their combined area.

A value near zero means little overlap. A value near one means the boxes are nearly identical.

Why detectors produce duplicates

Dense prediction creates many candidate boxes. Several neighboring predictions can all assign high confidence to the same physical object.

That is expected behavior before post-processing.

NMS as a greedy rule

A classic NMS loop is intentionally simple:

1. Keep the highest-confidence box.
2. Compare it with the remaining boxes.
3. Suppress boxes whose IoU crosses a threshold.
4. Repeat with the next surviving box.

while boxes:
    best = pop_highest_score(boxes)
    keep.append(best)
    boxes = [b for b in boxes if iou(best, b) < threshold]

The threshold is a product decision

A low threshold suppresses aggressively and can merge nearby objects. A high threshold preserves more candidates and can leave duplicates.

So NMS is not merely a mathematical cleanup step. It encodes what your application considers "the same detection."