Standard 8 stops to get here

Mean Average Precision (mAP)

The mean of average-precision scores across classes or queries: the headline metric for object detection and for ranked retrieval.

Your route here

8 stops · basics first
  1. Machine Learning ✓ understood

    Building systems that learn patterns from data instead of following hand-written rules, getting better at a task as they see more examples.

  2. Supervised Learning ✓ understood

    Learning from examples paired with the correct answer, so a model can predict answers for new inputs it hasn't seen.

  3. Classification ✓ understood

    A supervised learning task where the model assigns each input to one of a fixed set of categories, such as spam or not spam.

  4. Confusion Matrix ✓ understood

    A table showing true positives, true negatives, false positives, and false negatives for classification evaluation.

  5. Precision ✓ understood

    Of everything a model flagged as positive, the fraction that really was positive: TP / (TP + FP). It answers "when the model says yes, how often is it right?"

  6. Recall ✓ understood

    The share of actual positives a model correctly identifies, TP / (TP + FN); also called sensitivity or the true positive rate.

  7. Precision-Recall Curve ✓ understood

    A curve showing the tradeoff between precision and recall at different thresholds.

  8. Average Precision ✓ understood

    The weighted mean of precisions at each threshold, where the weight is the increase in recall from the previous threshold.

  9. Mean Average Precision (mAP) · you are here ✓ understood

Picture it

  1. 01 Rank the predictions By confidence, per class or per query
  2. 02 Mark each hit or miss Detection: a hit needs enough box overlap
  3. 03 Average precision One AP per class or query
  4. 04 Take the mean mAP: one number for the whole system
Notice the two averages: precision is averaged along each ranked list first, then those scores are averaged across classes or queries.

Mean average precision, usually written mAP, turns the quality of a ranked list of results into a single number, then averages that number over many lists. It is the standard headline score in two places: object detection and search. (Not to be confused with MAP, maximum a posteriori, a different idea that shares the acronym.)

The building block is average precision. Rank a model’s results from most to least confident and walk down the list; each time you reach a correct result, note the precision so far. Averaging those values gives AP, which is high only if the correct results sit near the top. AP summarizes one precision-recall curve. mAP is the mean of AP over every class, or every query.

In information retrieval, each query gets its own AP and mAP averages them. Manning, Raghavan and Schütze note that MAP became the most standard measure in the TREC evaluation community, valued for discriminating well between systems and staying stable.

In object detection

For object detection, a prediction only counts as correct if its box overlaps a real object enough, measured by IoU. The PASCAL VOC challenge required overlap above 50% and scored each class by AP. COCO made the test stricter: its primary metric averages AP over ten IoU thresholds, from 0.50 to 0.95, and over all 80 categories, which rewards boxes that fit tightly rather than roughly. COCO simply calls this AP and notes that it makes no distinction from mAP.

Reading a number

Because detection benchmarks compute mAP differently, a score is meaningless without its protocol. “[email protected]” (the VOC-style number) runs higher than COCO-style AP for the same model, so compare only numbers produced the same way.

Where it sits

Before this

Average Precision
Mean Average Precision (mAP)

Leads to

Nothing yet: a destination in its own right.

Explore nearby

Evaluation Intersection over Union A score from 0 to 1 for how well a predicted region matches the true one: the area they share divided by the area they cover together. Vision & Multimodal Object Detection Finding every object of interest in an image and giving each a class label, a confidence score and a bounding box. Evaluation Precision-Recall Curve A curve showing the tradeoff between precision and recall at different thresholds. Evaluation Precision Of everything a model flagged as positive, the fraction that really was positive: TP / (TP + FP). It answers "when the model says yes, how often is it right?" Evaluation Recall The share of actual positives a model correctly identifies, TP / (TP + FN); also called sensitivity or the true positive rate. Vision & Multimodal Non-Maximum Suppression Filtering overlapping detection boxes by keeping only the most confident predictions.

Sources

  1. Christopher D. Manning, Prabhakar Raghavan and Hinrich Schütze, Introduction to Information Retrieval . Cambridge University Press, 2008; section 8.4, Evaluation of ranked retrieval results
  2. Mark Everingham et al., "The PASCAL Visual Object Classes (VOC) Challenge" . International Journal of Computer Vision 88, 2010
  3. COCO Detection Evaluation . Official metric definitions for the COCO detection challenge