Average Precision (AP)

Imagine reviewing a detector’s predictions from most confident to least confident and asking: “How reliably does it find the real objects before it starts making too many mistakes?” Average Precision (AP) captures the answer in one score, rewarding detectors that find objects accurately and rank their best predictions first.

For one object class, such as cars, every predicted bounding box has a confidence score. Predictions are sorted from highest to lowest score, then matched against the labeled objects in the image. A prediction counts as a true positive when it has the correct class and overlaps an unused ground-truth box enough, measured by Intersection over Union (IoU). Otherwise, it is a false positive. Missed labeled objects contribute to false negatives.

• Precision asks: among predictions accepted so far, what fraction is correct?
• Recall asks: what fraction of all real objects has been found?
• Plotting precision against recall at progressively lower confidence thresholds produces a precision–recall curve.

AP is the area under this curve. A high AP means the detector maintains strong precision while reaching high recall; it finds most objects without flooding the result with incorrect boxes.

AP always depends on an IoU threshold. AP@0.5 treats a detection as correct when IoU is at least 0.50, while AP@0.75 demands tighter box placement. The COCO benchmark’s headline AP averages results across IoU thresholds from 0.50 to 0.95, making it sensitive to both recognition and localization quality. For example, a production-line detector can identify every defective bottle but still earn a lower strict AP if its boxes are poorly aligned. Mean Average Precision (mAP) then averages AP across classes, providing one headline score for a multi-class detector. This makes AP central when comparing models such as YOLO, Faster R-CNN, and DETR: accuracy alone cannot reveal whether their confidence-ranked detections are genuinely dependable.