Mean Average Precision

The most common accuracy metric for object detection and segmentation: average precision (AP) per class, averaged across classes.

Mean average precision (mAP) is the most common combined metric for object detection and instance segmentation. For a given class, all predictions are sorted by confidence, and a prediction only counts as correct if its intersection-over-union (IoU — the overlap area between predicted and ground-truth box, divided by their union area) with a ground-truth box exceeds a threshold; plotting this gives a precision-recall curve, and the area under that curve is the AP for that class, and averaging over all classes gives mAP. The number varies a lot depending on the threshold: the PASCAL VOC era commonly used IoU = 0.5 (mAP50), while COCO instead computes it at ten thresholds from 0.5 to 0.95 in steps of 0.05 and averages those (mAP50-95), which is a much stricter test of localization accuracy. Note that the “AP” reported in the COCO paper and its leaderboard has already been averaged over classes — the official documentation states explicitly that it makes no distinction between AP and mAP. When reading a robot perception paper, first check which threshold convention and which dataset the reported number uses.

ExampleThe Mask R-CNN paper reports a mask AP of 35.7 for the ResNet-101-FPN version on COCO — here, “AP” is the mAP averaged over 10 IoU thresholds and every class.

Also called

mAP, AP, mAP50, mAP50-95

Related

Intersection over Union · Precision / Recall · Object Detection · COCO / LVIS · Non-Maximum Suppression · Mask R-CNN

Sources

COCO Detection Evaluation（官方评测说明） (Chinese)
Ultralytics: Performance Metrics Deep Dive