From precision to average precision
Average precision summarizes a ranked list of results for one query. As you walk down the ranking, you record precision each time a relevant item appears, then average those values. A list that places relevant items near the top earns a high average precision.
Mean across queries
Mean average precision, or mAP, averages the average precision over many queries or classes. It rewards systems that consistently rank relevant items high across the whole workload, not just on a lucky query.
Where it appears
- Information retrieval uses mAP to compare search rankings.
- Object detection uses mAP across object classes, often at several overlap thresholds.
In detection, a predicted box counts as correct when its overlap with a true box passes a threshold. Reporting mAP across thresholds, such as the common range used in detection benchmarks, gives a fuller picture than a single cutoff.
Key idea
Average precision rewards ranking relevant items near the top of one query, and mean average precision averages it across queries or classes for retrieval and detection.