Before non-maximum suppression thins it, Faster R-CNN R50-FPN v2 hands back 66 candidates for this photograph: 59 for the dog and the ball, and 7 for classes the scene does not contain. They are ranked by score here and scored against two ground-truth boxes drawn by hand. Slide the score threshold: the operating point walks along the pooled curve, but the curve itself never moves — AP is a property of the ranking, not of the dial. Then read the two small plots: each class is scored on its own ranking, and mAP is the mean of their APs. Before you drag the IoU bar — the pooled curve measures 0.67 while both per-class curves fill the square. Which of those two numbers is the detector's fault?