Non-Maximum Suppression: Deduplicating Bounding Boxes with IoU

One car, three boxes. NMS keeps the best. Panel one - The over-detection problem: - One object, many overlapping boxes - Each box carries a confidence score - Keep the top-confidence box - Discard significant overlaps Panel two - Measuring overlap with IoU: - IoU = intersection / union area - 0 = no overlap, 1 = identical - Above 0.5, it is redundant Panel three - The greedy loop in action: - Sort by confidence, high to low - Keep best 0.95, cut IoU over 0.5 - Repeat until kept or cut Panel four - Falsify the greedy dial: - 0.95/0.88/0.82: one survivor - Two real cars over 0.5: one wrongly cut - Greedy reads overlap, not class - Soft-NMS and class-wise NMS fix it Sticky note - The one line: - One box per object - Threshold is a dial, not a win