Matterport3D

Matterport3D is a large-scale RGB-D dataset introduced in 2017 by researchers from Princeton, Stanford, Facebook, and Matterport. It contains 10,800 panoramic views captured from 194,400 RGB-D images across 90 building-scale indoor scenes (primarily residential and commercial buildings). The dataset provides comprehensive annotations including dense surface reconstructions, camera poses, 2D and 3D semantic segmentations, and a taxonomy of 40+ semantic categories. The scenes are globally aligned, covering entire buildings rather than isolated rooms. Matterport3D has become a foundational dataset for indoor scene understanding, 3D reconstruction, semantic segmentation, and embodied AI research. It serves as a key benchmark for navigation tasks in simulation platforms including Habitat and Gibson.