Abstract

Our aim is to create a unified semantic dataset which—by enlarging the number of classesand the diversity of the shared classes, aims to provide a more generic benchmark for training and evaluation of semantic segmentation methods.

The created dataset contains images obtained from different datasets re-labeled with a bigger variety of labels, obtaining a total of 293 semantic labels distributed among 145,555 training images and 7,614 validation images (each of one with its own annotation).

The datasets agglutinated to compose the proposed dataset are:

• COCO-Stuff, COCO-Stuff Dataset:Tran/Val 2017 Images and Annotations

• Cityscapes, Cityscapes Dataset

• ADE20K, ADE20K Dataset

• TASKONOMY, TASKONOMY Dataset

• Mapillary, Mapillary Dataset