Main contributions

  • A new scene parsing dataset (ADE20k). This new dataset contains 150 object and stuff classes. Each image has been labeled into objects, parts of objects, and in some cases even parts of parts.
  • A novel Cascade Segmentation network is proposed for parsing the scene.

Dataset overview

  • 20,210 training image, 2,000 validation images and 3,000 testing image.

Cascade segmentation model

Experiment results