What it does

Matterport3D bundles raw and annotated data from 90 properties scanned with a Matterport Pro Camera, plus scripts and baseline models for scene-understanding tasks. The repo holds the data organization docs, benchmark definitions, and MIT-licensed loading/viewing code; the actual dataset requires signing a terms-of-use form and emailing it from an institutional address.

The interesting bit

This is from 2017, yet it is still the go-to reference for indoor RGB-D work because it pairs dense 3D meshes with semantic instance annotations and camera poses — the boring plumbing that makes multi-view learning actually reproducible.

Key highlights

• Annotations include color/depth images, camera poses, textured meshes, floor plans, region labels, and object-instance semantics
• Five benchmark tasks defined: keypoint matching, view overlap prediction, surface normal estimation, region classification, and semantic voxel labeling
• Loading and viewing utilities provided in the code directory
• Published at 3DV 2017; 1,217 stars suggests sustained relevance

Caveats

• Access is gated behind a signed terms-of-use form and institutional email verification
• The repo itself is mostly documentation and glue code; the value is the dataset, not novel algorithms

Verdict

Grab this if you are training or evaluating indoor scene understanding models and need a standard benchmark. Skip it if you are looking for shiny new architectures — the code here is strictly utilitarian.