Description
3D-FRONT (Fu et al., ICCV 2021) is the dominant furnished-scene layout dataset: ~18,968 professionally-designed furnished rooms across 6,813 houses, populated with 13,151 furniture objects from the companion 3D-FUTURE asset set (~9,992 CAD models with textures + attributes). It is the training corpus behind almost every generative layout model — ATISS (arXiv-only, fetch manually), DiffuScene ↗, InstructScene, PhyScene.
Modality / size
- Scene layouts as per-object class + oriented 3D bounding box + orientation (exactly our box representation) + the 3D-FUTURE meshes.
- Residential only — bedroom / living / dining / study. No library stacks,
lecture rooms, cafeterias, corridors — the root cause of the field's
residential bias (see the brainstorm). Treat as a prior for the generic
room/officetheme, not the institutional ones.
Role in this project
- Optional learned prior for a future generative-layout comparison arm (ATISS/DiffuScene) against the authored grammar placer. Not on the critical path — the grammar placer needs no training data.
- 3D-FUTURE is a secondary dimension/material prior source alongside abo
for the furniture catalog (
docs/FURNITURE-CATALOG.md).
Acquisition & storage
- License: free for academic use; mesh redistribution restricted — do NOT mirror 3D-FUTURE meshes to our S3. Registration-gated on Tianchi/Alibaba.
- Request access → download → keep upstream. If any subset is staged, only the
layout JSONs (not meshes) under
datasets/3d-front-layouts/per the datasets contract.
Why it matters here
- The residential ceiling of this dataset is why the monad placer is an authored theme grammar (IP-092 follow-on) rather than a pretrained model — the non-residential themes we need are exactly what 3D-FRONT lacks.