Why can a detector win on mAP at IoU 0.5 but lose on mAP averaged over 0.50 to 0.95?

## Two ways to aggregate Average precision for one class is computed *at a fixed IoU threshold*: the threshold decides which predictions match ground truth, and from those true/false positives a precision-recall curve is traced by sweeping the confidence cut-off. mAP is that averaged over classes. So mAP is really a function of one hidden parameter, and reporting it without saying the threshold is meaningless. The two conventions in wide use: - **mAP at IoU 0.5** - a single, forgiving threshold. Classic and easy to read. - **mAP averaged over 0.50 to 0.95 in steps of 0.05** - ten separate evaluations, meaned. This is the headline number of the COCO-style protocol. Because a match at a high threshold is automatically a match at every lower one, per-threshold AP is non-increasing in the threshold. The averaged number is therefore always at or below [email protected], and the *size* of the gap is the diagnostic. ## What the gap measures Suppose your model finds nearly every object and orders its detections well, but each box is drawn a little too big, or is consistently shifted. At threshold 0.5, that costs nothing: a box that overlaps two thirds of the object still passes. At 0.75 a fair share of those boxes fall out of the match and turn into false positives *and* leave their objects unclaimed as misses, so precision and recall both drop. At 0.9 almost nothing survives. Average the ten terms and you get a small number even though [email protected] looked excellent. That is the confident-but-sloppy detector: strong classification head, weak or under-trained box regression. It wins any comparison run at 0.5 and loses badly on the averaged metric, and if your team ships on the 0.5 number you will discover the problem only when a downstream consumer starts cropping images at those boxes. The mirror image exists too. A model whose boxes are extremely tight but whose confidence scores are poorly ordered - some true detections scored below some false ones - suffers on the precision-recall curve at every threshold, but its loss is roughly constant across thresholds, whereas its rival's grows. Comparing the two curves rather than the two numbers is what tells you this. ## Reading it in practice The move is to plot AP against IoU threshold for each model, per class: - **High at 0.5, steep fall by 0.75** - localisation problem. Look at box-loss choice, feature resolution, label tightness, and whether your objects are small relative to the stride of the features the box head reads. - **Low everywhere, flat-ish shape** - detection or ranking problem: the objects are being missed or scored badly, and tightening boxes will not help. - **A cliff between two adjacent thresholds for one class only** - suspect systematic annotation convention, for example whether labels include an object's shadow, handle or occluded part. If humans drew the boxes to a different convention than your model learned, the two disagree by a constant margin that a threshold sweep exposes exactly. That last point is worth stressing: a strict metric measures agreement with the *annotation convention*, not with physical truth. If annotators are themselves inconsistent to within IoU 0.85, then AP at 0.9 is measuring label noise and no model can win it. Knowing your labels' own agreement level tells you which thresholds are informative on your data. ## Which one to report This is a product decision, and the honest answer names the consumer: - If a box triggers an alert, feeds a counter, or is drawn for a human who then looks at the region, 0.5 is a defensible operating criterion. - If a box is used to crop for a downstream reader, to measure a physical dimension, or to plan a grasp, loose boxes are failures and the strict average is closer to the truth. - If you are comparing to published work, report the same convention it used, and report both when they disagree. The worst outcome is a team that tunes for months against a threshold nobody chose deliberately. Pick the threshold that matches how the boxes are consumed, state it beside every number, and keep the sweep as a diagnostic so that a localisation regression cannot hide behind an unchanged headline figure.