Intersection over Union (IoU)

When a vision system draws a box around a car, a face, or a damaged product, it needs a way to judge how well that box matches reality. Intersection over Union (IoU) provides that score by measuring the overlap between a predicted region and the correct, human-labeled region.

How the score is calculated

IoU compares two shapes—usually bounding boxes, but also pixel masks in segmentation. The intersection is the area shared by both regions. The union is the total area covered by either region, counting the shared area only once. The calculation is:

IoU = area of overlap / area covered by either region

The result ranges from 0 to 1. An IoU of 0 means the regions do not touch; 1 means they match exactly. Think of laying one transparent rectangle over another: IoU asks what fraction of the combined visible area is covered by both.

How detectors use IoU

• Training: candidate boxes with sufficiently high IoU against a labeled object become positive examples; low-overlap boxes become background examples.
• Evaluation: a detection is counted as correct only when its IoU passes a chosen threshold, such as 0.5. Metrics such as Average Precision test performance across thresholds.
• Non-maximum suppression (NMS): if a detector produces several highly overlapping boxes for the same person, IoU helps identify duplicates so the strongest prediction can remain.

Why it matters

Classification alone can say “there is a bicycle”; IoU checks whether the model found the bicycle’s location accurately. A box can have the right label but poor IoU if it is shifted, too large, or misses much of the object. This is especially important for autonomous driving, where a loose box around a pedestrian is risky, and for medical segmentation, where overlap with a tumor mask directly affects clinical usefulness. Libraries such as torchvision provide a box_iou function to compute these comparisons efficiently.