PascalVOC-SP

Assign one of 21 semantic-segmentation labels, including background, to every superpixel node. A superpixel receives the ground-truth class of the pixel at the superpixel’s center-of-mass coordinate.

Each of 11,355 images from the Semantic Boundaries Dataset augmentation of Pascal VOC 2011 becomes one graph. SLIC with compactness 30 produces at most 500 image regions, and two regions are connected when they share a boundary in the default region-adjacency graph.

Each node has 12 RGB summary values—the mean, standard deviation, maximum, and minimum in each channel—plus its two-dimensional center-of-mass coordinate. Each edge stores the average Sobel response and pixel count along the shared boundary.

The original training set remains training data. The original validation set is divided equally into validation and test subsets using a stratified image-level meta-label obtained by majority vote over non-background node labels, yielding 8,498/1,428/1,429 graphs. The score is class-weighted macro F1 over nodes.

This is not pixel-level segmentation; labels and predictions are quantized to superpixels, and labeling only from the center pixel can misrepresent mixed regions. LRGB also releases compactness-10 and 8-nearest-neighbor graph variants, whose structures and edge features differ from the default compactness-30 region-boundary version.