The problem — Object detection and instance segmentation pipelines need a binary decision: did the model find this object, or not? A pixel-level overlap score like Dice gives a continuous value but does not answer the yes/no question. IoU at a threshold turns continuous overlap into a match verdict — but the threshold choice is a policy decision that profoundly affects what the metric rewards, and default values (IoU ≥ 0.5) are often too lenient for biology.
What it is / how it works — IoU computes the area (or volume) of the intersection of two regions divided by their union: IoU = |P ∩ G| / |P ∪ G|. It is symmetric, bounded in [0, 1], and reduces to Dice when applied to binary masks via the relation Dice = 2 × IoU / (1 + IoU). In object detection, a predicted bounding box is declared a true positive if its IoU with a ground-truth box exceeds a threshold (commonly 0.5 or 0.75); in instance segmentation the same logic applies to mask polygons. mAP is built on top of IoU: it sweeps confidence thresholds and averages precision at different IoU cutoffs (e.g., COCO evaluates at 0.5:0.05:0.95).
IoU is also the direct building block of the Cell Tracking Challenge SEG measure, which computes the mean IoU over matched reference cells, and of Panoptic Quality, whose segmentation quality term is the average IoU of matched instance pairs.
Where it breaks — IoU at IoU ≥ 0.5 is permissive: a prediction can miss half the reference object and still count as a match. For small structures — organelles, rare cell subpopulations, nuclei in crowded tissue — a 0.5 threshold allows boundary errors that would corrupt per-cell feature extraction. In spatial omics, even a small boundary misalignment misassigns transcripts to the wrong cell, a failure that IoU does not capture. In 3D light-sheet volumes, global IoU is dominated by cytoplasmic volume and is insensitive to membrane boundary errors that separate or merge adjacent cells.
Metrics Reloaded explicitly recommends matching IoU thresholds to the task's sensitivity requirements — a strict IoU ≥ 0.75 or better for high-content profiling pipelines where boundary quality matters. Evaluating Segmentation Models in Real Microscopy Workflows reinforces this: report IoU distributions and merge/split rates rather than a single threshold pass/fail, and validate on the instrument-and-stain combination that will be deployed.