Detection asks two questions at once and scores them separately.

Demonstration

Every detector optimises a sum of two losses: a classification loss over a label set, and a regression loss on four box coordinates, usually some form of intersection-over-union. Weighting them against each other is a hyperparameter, and it is a statement about which error you would rather make.

Three difficulties follow, and each produced a family of methods.

Class imbalance is extreme. A dense detector evaluates tens of thousands of candidate locations, of which a handful contain objects. Trained with ordinary cross-entropy, the easy background dominates the gradient. Focal loss down-weights well-classified examples so the rare positives are heard, and that single change is what made one-stage detectors competitive.

Assignment is a choice. Which prediction is responsible for which ground truth object? Anchor-based methods match by IoU against predefined boxes; DETR removes anchors entirely and solves a bipartite matching between a fixed set of queries and the objects present. The assignment rule is not a detail — it determines what the loss even means.

Duplicates need removing. Several predictions fire on one object, so classical detectors run non-maximum suppression as a post-process, with its own threshold. DETR’s set-based loss makes it unnecessary, which is the cleanest thing about that architecture.

Then the reporting. mAP averages precision over recall levels, over classes, and over IoU thresholds. COCO’s headline number averages IoU from 0.5 to 0.95. A model with excellent classification and sloppy boxes and one with the reverse can land on the same mAP.

Corollary

Ask for mAP@0.5 and mAP@0.75 separately. The gap between them is the box quality, and it is invisible in the averaged number that gets quoted.