COCO-SP

Predict one of 81 semantic labels for each superpixel node. The assigned target is the pixel annotation at the mean coordinate of the superpixel region.

The benchmark converts 123,286 MS COCO images into graphs. Its default construction uses SLIC superpixels with compactness 30 and at most 500 regions per image, joined by edges when the regions share a boundary.

Nodes contain 12 per-region RGB statistics—mean, standard deviation, maximum, and minimum per channel—and two center-of-mass coordinates. Region-boundary edges contain the mean Sobel response and the number of boundary pixels.

The original COCO validation set becomes a 5,000-image test set; 5,000 images sampled from the original training set form validation, leaving 113,286 for training. The official score is class-weighted macro F1 over superpixel nodes.

The task differs from original COCO instance segmentation and inherits approximation error from superpixelization and center-pixel labels. Optional LRGB variants change SLIC compactness or replace region-boundary edges with coordinate- or feature-based 8-nearest-neighbor graphs; results across these graph constructions are not directly comparable.