← AI Terminology
mAP - mean Average Precision
mAP (mean Average Precision) averages precision–recall performance across classes and detection thresholds — the headline metric for object detection and some retrieval tasks.
COCO mAP defines modern detection leaderboards.
COCO mAP defines modern detection leaderboards.
Why It Matters in AI
Accuracy is meaningless for detection with many boxes and classes. mAP summarises ranking quality of detections and is required literacy for computer-vision products and YOLO/DETR-style models.
Key Points
| Aspect | Description |
|---|---|
| Use | Detection, instance segmentation, sometimes retrieval |
| COCO | Average over IoU thresholds 0.5:0.95 |
| Idea | AP per class from PR curve; mean over classes |
| Tools | pycocotools evaluation |
| Caveat | Protocol details must match when comparing |
| Related | IoU, NMS, precision/recall |
Simple Analogy
Not only “did you find the cats?” but how precisely and completely you ranked every cat box across the whole photo album.
Common Usage Examples
- COCO eval mAP for detectors
- Report AP50 vs COCO mAP
- Track mAP when changing NMS
- Class-wise AP for long-tail analysis
Summary
In short: mAP summarises detection quality across classes and overlap thresholds — the standard scoreboard for object detectors.