Mask Mean Average Precision

The intersection over union (IoU) threshold determines how much overlap a predicted mask must have with its ground-truth mask to count as a matched detection. A threshold therefore changes which predictions contribute to the precision-recall calculation. Reporting mask mAP at defined thresholds helps show whether a model maintains performance when mask overlap requirements become more demanding.

Class-wise averaging prevents performance on common or easily detected objects from completely determining the overall result. The metric first calculates average precision for each class, then combines those values into a mean. This makes the score useful when an engineering system must recognize and segment several object categories rather than optimize for only one.

Average precision is calculated from the precision-recall curve produced by mask detections and their matches to ground-truth masks. This captures performance across the system’s detection results instead of relying on a single operating point. When benchmarks also average across multiple IoU thresholds, the final score summarizes mask quality under several matching requirements.

Bounding-box metrics evaluate whether predicted rectangles correspond adequately to objects, whereas mask mAP evaluates the predicted pixels within each object instance. Two systems can therefore show similar box-level performance while differing in how accurately they delineate object shapes. This distinction matters when boundary or region accuracy affects the engineering task.

A typical calculation groups predictions and ground-truth masks by object class, compares corresponding regions using IoU, and applies the selected IoU threshold to determine matches. The resulting detections generate a precision-recall curve from which average precision is calculated for each class. Those class scores, and sometimes threshold-specific scores, are then averaged.

Engineers use the metric when they need an objective comparison of instance-segmentation models, especially in robotics, autonomous systems, industrial inspection, or medical imaging. It provides a common basis for evaluating both object identification and pixel-level delineation. Comparing scores across models can reveal differences that an evaluation based only on bounding boxes may overlook.

Teams should consider the object classes included, the IoU thresholds used, and whether the reported value averages one threshold or several. These choices affect what the score represents. Examining the underlying class-wise average precision values can show whether the overall result reflects consistent segmentation performance or uneven results across the evaluated categories.