ScanNet

ScanNet is a large-scale dataset and benchmark for 3D indoor scene understanding. It provides richly annotated, densely reconstructed 3D scenes derived from RGB-D scans of real indoor environments, including homes and offices. Each scene is represented as a textured 3D mesh with per-voxel or per-point semantic labels and, in many cases, instance annotations for common objects such as furniture and appliances.

The data release includes synchronized RGB and depth frames, camera poses, and the resulting 3D reconstructions, along with label files that map geometry to semantic and instance identities. The dataset facilitates a range of 3D understanding tasks, including semantic segmentation, instance segmentation, and 3D object detection. A variety of evaluation metrics are used, typically involving mean intersection-over-union (mIoU) across semantic classes and object-level metrics for instance detection.

ScanNet originated in the late 2010s, with an updated version known as ScanNet v2 that expanded the number of scenes and improved annotation quality. The dataset is widely used by researchers to train and benchmark deep learning models for 3D scene understanding, serving as a standard reference in academia and industry for evaluating methods on real-world indoor environments.

Access to the data is provided for research purposes, generally requiring agreement to licensing terms and a registration process. ScanNet has influenced subsequent datasets and methods in 3D vision, contributing foundational benchmarks for processing and interpreting complex indoor scenes.