Taskonomy A large, high-quality dataset of various indoor scenes, including more than 4.5 million images of indoor scenes from more than 500 buildings.

The dataset covers 26 common tasks in the visual field, including but not limited to: 2D edge detection, plane normals, jigsaw puzzles, automatic coloring, room layout construction, object classification, and scene prediction. Each image has all the annotations for these 26 tasks.

The dataset also has the following features:

• Contains complete pixel-level geometry information obtained through aligned grids.
• Contains semantic information obtained through knowledge distillation of ImageNet, MS COCO, and MIT Places datasets.
• All consistent camera positions. Complete camera parameters.
• High-resolution images.
• It is 3 times larger than the ImageNet dataset.