The COCO Stuff Segmentation Task aims to advance the semantic segmentation technology level of the Stuff Class. Compared with the target detection task that mainly solves the Thing Class (people, cars, elephants), this task focuses more on the Stuff Class (grass, wall, sky).

COCO Stuff refers to objects of specific size and shape, which are usually composed of parts, while Stuff Class is a background defined by a uniform or repeated pattern of fine-scale attributes, without a specific or unique spatial range and shape. Stuff covers about 66% pixels in COCO, which enables us to interpret and understand important aspects of the image: scene type, possible categories and locations, and geometric properties of the scene.