Description

Objaverse (Deitke et al., CVPR 2023) is a massive open-vocabulary 3D object bank — 800k+ annotated assets (Objaverse-XL: 10M+). It is what lets LLM-agent scene generators (Holodeck , I-Design) furnish arbitrary non-residential themes — libraries, cafeterias, lecture halls, museums — because the object list comes from language + retrieval, not a fixed residential training set.

Modality / size

  • 800k / 10M+ .glb meshes with captions + tags. Access via pip install objaverse; objaverse.load_objects([uid]).
  • Caveats: per-object licenses are heterogeneous (Sketchfab-sourced; many CC-BY, some non-commercial) — filter per-UID before any redistributive use. Scale is not normalised (objects arrive at arbitrary units) — rescale to the class's known real height (Holodeck implements exactly this CLIP-retrieval + rescale loop).

Role in this project

  • Only relevant if we add the LLM-agent placement layer (Option B in the brainstorm): the agent picks a furniture class per theme, retrieves an Objaverse mesh, rescales, and we keep the oriented box. Cosmetic for CSI/peds (we need the box, not the mesh) — keep off the critical path.
  • Breadth source for the non-residential asset gap that abo and 3d-front don't cover (institutional seating, shelving variety).

Acquisition & storage

  • License: aggregate ODC-By v1.0 (Objaverse-XL) / per-object licenses — verify before any redistribution. Do NOT mirror to our S3; load per-UID on demand.
  • No bulk download needed — the retrieval loop pulls only the handful of UIDs a scene references.

Why it matters here

  • The pragmatic decision (see brainstorm): we do not need Objaverse for the grammar placer. It becomes useful only for the optional Holodeck-style agent layer or for pretty 3D on the web atlas.