Scene Parsing through ADE20K Dataset

Scene parsing is one of the key problems in computer vision, but despite community data-collection efforts few datasets cover a wide range of scenes and object categories with dense, detailed annotations. The paper introduces and analyzes ADE20K, a dataset spanning diverse annotations of scenes, objects, parts of objects, and in some cases even parts of parts, and builds a scene-parsing benchmark on it with 150 object and stuff classes.

Several segmentation baseline models are evaluated on the benchmark, and the authors propose a novel network design, the Cascade Segmentation Module, that parses a scene into stuff, objects, and object parts in a cascade and improves over the baselines. They further show that the trained scene-parsing networks enable downstream applications such as image content removal and scene synthesis, underscoring the dataset's practical value.