Day 49: mAP: how detection is actually evaluated
The metric every detection paper reports
Detection needs a metric that accounts for both *classification* (right label?) and *localization* (box in the right place?). mAP — mean Average Precision — does exactly this. A detection counts as correct only if its class is right *and* its IoU with the ground-truth box exceeds a threshold. From those correct/incorrect calls you compute a precision-recall curve per class, take its area (Average Precision), and average across classes.
- AP@0.5 — the classic threshold: a box is a hit if IoU ≥ 0.5.
- mAP@[0.5:0.95] — the stricter COCO metric, averaging AP across IoU thresholds 0.5 to 0.95; rewards tight localization.
- mAP = mean of AP across all classes — one number summarizing the detector.
Why the stricter metric matters for measurement
For FitXpert, a loosely-placed box that still counts at IoU 0.5 may not be tight enough for accurate measurement downstream. The stricter mAP@[0.5:0.95] tells you whether boxes are *tightly* placed — which matters when pixel positions feed into centimeter estimates later this stage.
Key terms
- mAP (mean Average Precision)
- The standard detection metric: mean over classes of the area under each class's precision-recall curve, using IoU-thresholded correctness.
- AP@0.5
- Average Precision counting a detection correct if class matches and IoU ≥ 0.5.
- mAP@[0.5:0.95]
- COCO-style mAP averaged over IoU thresholds 0.5–0.95, rewarding tight localization.
For a detection to count as a true positive in mAP, what two conditions must hold?
