IoU (Intersection over Union)
A metric measuring the overlap between a predicted and ground truth region, calculated as intersection divided by union.
IoU measures overlap of prediction and ground truth (intersection/union) – the universal metric for object detection and segmentation.
Explanation
IoU is universally used in object detection and segmentation. An IoU ≥ 0.5 is typically considered a "correct detection" (AP@50).
Marketing Relevance
IoU is the standard evaluation metric for all object detection and segmentation models – from YOLO to SAM.
Example
An object detection model achieves mAP@50 = 0.85, meaning 85% of predictions have IoU ≥ 0.5 with the ground truth.
Common Pitfalls
IoU threshold strongly influences results. High IoU values are harder to achieve for small objects.
Origin & History
IoU is based on the Jaccard Index (Paul Jaccard, 1901). In computer vision it became the standard metric for PASCAL VOC and later ImageNet/COCO detection benchmarks from the 2000s.
Comparisons & Differences
IoU (Intersection over Union) vs. Dice Coefficient
Dice = 2×intersection/(A+B); IoU = intersection/union. Dice weighs overlap more heavily and is more common in medical segmentation.
IoU (Intersection over Union) vs. mAP (Mean Average Precision)
IoU is an overlap metric for individual predictions. mAP aggregates precision across all predictions at various IoU thresholds.