mean average precision (mAP)

A detector's average precision, averaged again across every object class, into one summary number.

Mean average precision (mAP) summarizes a detector's overall quality as the mean, across every object class, of that class's average precision (the area under its own precision-recall curve, swept over confidence threshold). mAP is only comparable between two detectors evaluated on the same dataset with the same IoU-threshold convention.

Examples

A detector scoring AP $=0.85$ on cars, $0.60$ on bicycles, and $0.40$ on pedestrians has mAP $=0.617$ across those three classes -- but only comparable to another detector's mAP computed on the same dataset and IoU convention.