Full 2D-3D-Semantic is a large-scale indoor space image dataset, covering an indoor area of up to 6,000 square meters, containing more than 70,000 RGB images, as well as corresponding depth, surface normals, semantic annotations, global XYZ images (all in the form of regular and 360° equirectangular images), camera information, registration information, semantically annotated 3D meshes and point clouds. The dataset provides a mutual registration model between 2D, 2.5D and 3D domains, and has instance-level semantic and geometric annotations.

The dataset enables the development of joint and cross-modal learning models and potential unsupervised methods by exploiting regularities existing in large-scale indoor spaces.