The problem — Instance segmentation evaluation has historically split across two metrics: detection performance (did the model find the right objects?) and segmentation quality (how well did it draw their boundaries?). A model can score well on detection while producing coarse masks, or achieve sharp masks on easy objects while missing the hard ones. Panoptic Quality was designed to hold both dimensions accountable in a single number.
What it is / how it works — Proposed by Kirillov et al. at CVPR 2019, PQ decomposes into two factors: Segmentation Quality (SQ), the mean IoU over all true-positive matched pairs, and Recognition Quality (RQ), the F1 score computed from true positives, false positives, and false negatives at an IoU ≥ 0.5 matching threshold. Together: PQ = SQ × RQ. This factored form is interpretable — a drop in PQ can be immediately attributed to worse boundary drawing (falling SQ) or worse detection coverage (falling RQ). PQ was originally defined for panoptic segmentation (semantic + instance), but the instance-only variant is widely adopted in cell biology benchmarks and features in workflow-level evaluation recommendations.
Metrics Reloaded endorses PQ for instance segmentation tasks where both detection and delineation quality matter — which covers the majority of high-content imaging workflows.
Where it breaks — The IoU ≥ 0.5 matching threshold inherited from COCO may be too lenient for high-content microscopy. A predicted nucleus that covers half the reference is called a match, even though the missed half will corrupt per-cell morphology features. In dense tissue sections with overlapping nuclei, the binary match/no-match verdict can systematically penalize partial overlaps that would be acceptable biologically. A 2023 analysis in Scientific Reports argues that PQ should be used cautiously for cell nuclei segmentation in digital pathology specifically because the fixed threshold ignores class imbalance between large stromal cells and small epithelial nuclei.
Like all instance metrics, PQ requires per-instance ground-truth annotations — expensive to produce and always reflecting human boundary decisions, as Ground Truth Is a Design Problem explains. The Fovea operating stance is to report PQ as one of several metrics (alongside SEG and split/merge counts) and always evaluate on the instrument and staining protocol of deployment, not a benchmark set.