Intersection over Union

The overlap area of two regions divided by their combined area, used to judge how accurate a predicted box or mask is.

Intersection over union measures how much two regions overlap: intersection area divided by union area, with 1 meaning perfect overlap and 0 meaning no overlap at all — mathematically, this is the Jaccard index, proposed by Paul Jaccard in 1901. Vision tasks use it to judge whether a detected box or mask is accurate: a prediction counts as correct only if its IoU with the ground truth exceeds some threshold. PASCAL VOC uses a threshold of 0.5; COCO is stricter, computing average precision (AP) at thresholds from 0.50 to 0.95 in steps of 0.05 and then averaging them — the commonly cited AP50 and AP75 are the results at the 0.5 and 0.75 thresholds specifically. Segmentation tasks use the same formula, just applied to mask pixels instead of a box.

ExampleA ground-truth box and a predicted box each have an area of 100 and overlap by 60. Their union is then 140, giving an IoU of about 0.43 — below 0.5, so by the VOC standard this detection wouldn't count as correct.

Also called

IoU, Jaccard Index

Related

Object Detection · Bounding Box · Mean Average Precision · Non-Maximum Suppression · Instance Segmentation · Precision / Recall

Sources

Jaccard index - Wikipedia
COCO Detection Evaluation