Mean average precision, or mAP, is the standard headline metric for object detection and instance segmentation. It averages the average precision across all classes, and often across a range of intersection over union thresholds, into a single score that summarizes overall detection quality.
Mean average precision averages the average precision computed for each class, producing one number that reflects how well a detector performs across the whole label set. Modern benchmarks usually go a step further and average over multiple intersection over union thresholds as well, rewarding models that localize objects tightly, not just approximately.
The result is a threshold-independent, class-balanced summary that lets you rank detection and segmentation models against each other.
Key takeaways
mAP averages average precision across all classes.
Benchmarks often also average across a range of IoU thresholds.
It is the standard comparison metric for detection and segmentation.
What mAP captures
Common mAP conventions and what they reward.
Common mAP conventions and what they reward.
Convention
What it measures
mAP at IoU 0.5
Correct detections with loose localization
mAP at IoU 0.75
Correct detections with tight localization
mAP averaged over 0.5 to 0.95
Overall quality across localization strictness
How it works
For each class, the precision-recall curve is built and reduced to an average precision at a given IoU threshold. Those per-class values are averaged to get mAP at that threshold, and the popular COCO-style metric then averages mAP across IoU thresholds from 0.5 to 0.95. Because the recipe varies between benchmarks, mAP numbers are only comparable when computed the same way.
Why it matters
mAP is the number leaderboards and papers quote, so it shapes how detection progress is measured and communicated. Its class averaging also stops a model from looking good just by nailing the common classes while ignoring rare ones, which matters for real datasets with long tails.
Frequently asked questions
What does mAP at 0.5 to 0.95 mean?
It is the mean average precision averaged over IoU thresholds from 0.5 to 0.95, a stricter and more complete measure of localization quality.
Are mAP scores comparable across datasets?
Not directly. mAP depends on the classes, the data, and the exact computation, so it is meaningful mainly within the same benchmark.