MSeg is a multi-domain semantic segmentation composite dataset. This dataset unifies semantic segmentation datasets from different fields: COCO, ADE20K, Mapillary, IDD, BDD, Cityscapes, and SUN RGB-D. By coordinating classification, merging, and splitting classes, a unified classification with 194 categories is obtained.

To make the pixel-level annotations conform to a unified taxonomy, we performed a large-scale annotation effort on the Mechanical Turk platform and generated compatible annotations in the dataset by re-labeling object masks.