Panoptic Segmentation

Panoptic segmentation unifies semantic and instance segmentation into a single output.

Every pixel receives a semantic class. Countable object categories (things, such as people and cars) also get an instance ID, while uncountable background regions (stuff, such as sky and road) keep only the class label.

The result is a complete, non-overlapping parse of the scene that captures both object identities and background context.

Standard benchmarks include COCO Panoptic and Cityscapes, with representative models such as Panoptic FPN and Mask2Former.