Non-Maximum Suppression: Cleaning Up Overlapping Detection Boxes

Detector Thresholds

Confidence threshold 0.30

IoU threshold (NMS) 0.45

Show suppressed boxes

Live Readout

Above confidence

Kept after NMS

Suppressed (overlap)

Kept detection Suppressed proposal Below confidence threshold Ground-truth object

How it works

A real object detector (YOLO, Faster R-CNN, SSD…) rarely emits one clean box per object — it emits dozens of overlapping candidate boxes around every object, each with a confidence score. Non-Maximum Suppression (NMS) is the greedy algorithm that turns that noisy cloud into one box per object.

IoU(A, B) = area(A ∩ B) / area(A ∪ B)

NMS(boxes, confT, iouT):
  boxes = { b in boxes : score(b) ≥ confT }
  sort boxes by score, descending
  kept = []
  while boxes not empty:
    best = boxes.pop_highest_score()
    kept.append(best)
    boxes = { b in boxes : IoU(best, b) < iouT }
  return kept

Confidence threshold — discards low-confidence proposals before NMS even runs (grey boxes below it never compete).
IoU threshold — how much two boxes must overlap before the weaker one is thrown away. Low values suppress aggressively (fewer, sparser boxes); high values let near-duplicates survive.
Greedy loop — the highest-scoring remaining box is always kept, then every box that overlaps it past the IoU threshold is removed, and the process repeats on what's left.

Real-world relevance: every modern detection pipeline runs NMS (or a learned variant like Soft-NMS) as its very last step, right after the CNN backbone produces raw candidate boxes — without it, a single car would be reported as ten overlapping boxes instead of one.

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.