ScanNet is an instance-level RGB-D dataset of indoor scenes (2D & 3D data) containing 2.5 million views from more than 1,500 scans, annotated with 3D camera keypoint detection (3D camera poses), surface reconstruction, and instance-level semantic segmentation. It is a collection of labeled voxels, not points or objects.

This dataset is ScanNet v2, the latest version of the ScanNet dataset, which has collected 1,513 annotated scans with a surface coverage of approximately 90%. In the semantic segmentation task, this dataset is labeled as 20 categories of annotated 3D voxelized objects.

Using this dataset helps achieve state-of-the-art performance in 3D scene understanding tasks such as 3D object classification, semantic voxel labeling, and CAD model retrieval.