What it does

Structured3D is a dataset of 3,500 house designs from professional interior designers, rendered into photo-realistic 2D images with dense ground-truth annotations: semantic labels, depth, surface normals, albedo, room layouts, 3D bounding boxes, and wireframes. The repo provides Python visualization tools to inspect wireframes, floorplans, textured meshes, and perspective or panorama layouts.

The interesting bit

The data comes from real architectural design software (courtesy of Kujiale.com), not procedurally generated boxes. That means the room shapes, furniture arrangements, and lighting configurations reflect actual human design decisions—while still offering pixel-perfect labels impossible to collect in the physical world.

Key highlights

• 3.5K scenes with multiple lighting and furniture configurations per scene
• Annotations include semantic, depth, normal, albedo, layout, wireframe, plane, and 3D bounding box
• Standard train/val/test splits predefined (scenes 00000–02999 / 03000–03249 / 03250–03499)
• Visualization scripts for 3D wireframe, plane, mesh, layout (perspective + panorama), and floorplan
• ECCV 2020 paper with supplementary material and Codalab benchmark

Caveats

• Dataset requires signing a terms-of-use agreement form for access; not instant download
• Known invalid cases exist (listed in metadata/errata.txt); check before training
• Bounding box annotations were corrected in March 2020; older downloads may have wrong basis

Verdict

Grab this if you’re training room-layout estimators, 3D reconstruction models, or anything that needs structured indoor geometry with clean supervision. Skip if you need dynamic scenes, outdoor environments, or real sensor noise—these are static, synthetic interiors.