What is Non-Max Suppression (NMS), and why is it crucial in object detection?

NMS is the tidy-up step that turns a messy pile of raw predictions into one box per object. The algorithm is short:

def nms(boxes, scores, iou_thresh=0.5):
    """Return indices of boxes to keep."""
    order = scores.argsort()[::-1]  # high score first
    keep = []
    while len(order) > 0:
        i = order[0]
        keep.append(i)
        # drop boxes that overlap the winner too much
        rest = order[1:]
        order = rest[[iou(boxes[i], boxes[j]) < iou_thresh for j in rest]]
    return keep

Why the threshold matters

• Too high (e.g. 0.9): near-duplicate boxes survive, so you over-count.
• Too low (e.g. 0.2): two genuinely close objects get merged into one, so you under-count.
• A common default is around 0.45-0.5 for general scenes; tune it when objects are densely packed.

Variants like Soft-NMS decay scores instead of hard-deleting boxes, which helps in crowded scenes where real objects overlap heavily.

Hands-on tutorials