What is panoptic quality?
Panoptic quality is the metric designed for panoptic segmentation, the task that labels every pixel with both a class and, for countable objects, an instance identity. PQ factors neatly into two parts: segmentation quality, the average intersection over union of correctly matched segments, and recognition quality, an F1-style score over whether segments were detected at all. Multiplying them gives a single, interpretable number.
This structure lets you see whether a model's errors come from poor masks or from missing and hallucinated objects.
Key takeaways
- PQ is the standard metric for panoptic segmentation.
- It decomposes into segmentation quality and recognition quality.
- The split shows whether errors are in masks or in detection.
How it works
Predicted and ground-truth segments are matched when their intersection over union exceeds 0.5, which makes the matching unambiguous. Segmentation quality averages the IoU of matched segments, recognition quality is the F1 over matched, missed, and false segments, and PQ is their product. It is computed for both amorphous regions like sky or road and countable objects like people or cars.
Why it matters
Panoptic segmentation is central to scene understanding for robotics and autonomous systems, where you need both the layout of the scene and the individual objects in it. PQ gives a single, comparable score for that combined task, and its decomposition makes it a useful diagnostic, not just a leaderboard number.
Frequently asked questions
How is PQ different from mIoU?
mIoU scores semantic segmentation by class overlap. PQ also accounts for distinguishing individual object instances and for detection errors, which mIoU does not capture.
What are stuff and things in panoptic segmentation?
Stuff refers to uncountable regions like grass or sky, and things refers to countable objects like cars or people. PQ evaluates both.
Related terms