Precision / Recall

Precision measures how many reported results were correct; recall measures how many of the true targets were found.

Precision and recall are a pair of basic metrics for detection, classification, and retrieval tasks. Results are split into true positives (TP, reported and correct), false positives (FP, reported but wrong — a false alarm), and false negatives (FN, should have been reported but wasn't — a miss). Precision is TP/(TP+FP), and recall is TP/(TP+FN). The two usually trade off against each other: lowering the confidence threshold finds more targets but also raises the false-alarm rate. F1 is their harmonic mean, 2PR/(P+R), capturing both in a single number. Object detection first uses IoU to decide whether a predicted box counts as a hit, and then computes these metrics; sweeping the threshold traces out a precision-recall curve, whose area underneath is average precision (AP), which averaged across classes gives mAP. Robots also use these metrics to evaluate grasp detectors, contact detectors, and success detectors.

ExampleA cup detector reports 100 boxes on a test set, 80 of which correctly match real cups, out of 120 actual cups in the set: precision is 80%, recall is about 67%, and F1 is about 0.73.

Also called

F1 Score, F1

Related

Intersection over Union · Mean Average Precision · Object Detection · Non-Maximum Suppression · Success Detector

Sources

Wikipedia: Precision and recall
scikit-learn: Precision, recall and F-measures