Mean Average Precision (mAP) is the primary evaluation metric for object detection models. It summarizes how well a model handles both classification (is this object a car or a person?) and localization (how tightly does the predicted box fit the actual object?) into a single number between 0 and 1.

The calculation works from the bottom up. For each class, predictions are ranked by confidence score. Walking down this ranked list, precision and recall are computed at each point, generating a precision-recall curve. The area under this curve is the Average Precision (AP) for that class. Mean AP averages this value across all classes. The COCO benchmark standard, mAP@[0.5:0.95], averages AP across IoU thresholds from 0.5 to 0.95 in steps of 0.05, rewarding models that produce tight, well-localized boxes. The simpler AP@0.5 ("AP50") only requires 50% overlap between prediction and ground truth, and is more forgiving of localization errors.

When comparing models, mAP numbers only make sense if computed on the same dataset with the same evaluation protocol. COCO mAP, PASCAL VOC mAP, and custom dataset mAP use different IoU thresholds, class counts, and evaluation details. A model reporting 55% mAP on COCO is strong; 55% on a custom 3-class dataset may be mediocre. Always check the evaluation protocol alongside the number.