What it is

ABO (Collins et al., CVPR 2022) — 147,702 real Amazon product listings with multilingual metadata and dimensions, 398,212 catalog images, and 7,953 high-quality glTF 3D models with PBR materials. Here it is the grounding source for the furniture catalog's two needs: metric footprints and materials.

We mirrored the metadata subset only (acquired-partial): the listings/ product metadata carries item_dimensions + material + product_type, which is all the catalog consumes. The 165 GB 3D-model tar and the image sets are excluded (we need numbers, not meshes).

Verified specs (from exploration, not the README)

  • Subset on our S3: 16 × listings/metadata/listings_{0..f}.json.gz (gzipped ndjson, one product/line) + README.md, listings/README.md, LICENSE-CC-BY-4.0.txt, derived_catalog_priors.json, _manifest.json. 87.5 MB, 21 objects.
  • Dimensions: item_dimensions.{width,length,height}.normalized_value.{unit,value} — units are inches (mostly) or centimetres; abo_priors._to_m converts. ABO length = the depth axis.
  • Materials: material[0].value is free text ("Wood", "Metal", "Fabric", "PETG…") → keyword-mapped to BoundaryMaterial.
  • Furniture product_type coverage (listings with usable dims): CHAIR 1652, SOFA 973, TABLE 841 (conflates coffee+dining — height median 0.52 m), CABINET 222 (wood, H≈0.99 m), SHELF 226 (wall shelves, not tall stacks), DESK 169 (wood, 0.95×0.71×0.76 m), OTTOMAN/STOOL_SEATING/PLANTER present. No library stack / lecture seat-row / cafeteria long-table types → those stay on ergonomics standards.

Acquisition & storage

  • Openly downloadable from AWS Open Data (no credentials): aws s3 sync --no-sign-request s3://amazon-berkeley-objects/listings/metadata/ …
  • Our mirror: s3://monad-knowledge/datasets/abo/ (load via the ABO pipeline notebook / monad_knowledge.gis.furnish.abo_priors).
  • License: CC-BY-4.0 — commercial-OK, attribution to Amazon.com (the stale LICENSE-CC-BY-NC-4.0.txt in the bucket is superseded by the top-level README's CC-BY-4.0 declaration).
  • Do not mirror the 3D-model / image tars — meshes stay upstream.

Why it matters here

  • Grounds monad_knowledge/gis/furnish/catalog.py dimensions + materials for the theme-aware placer (IP-092 follow-on provenance='computed' cells) — the box+material representation the CSI ray-tracer (Sionna RT Differentiable Ray Tracing for Radio Propagation Modeling, IP-101 scene_overrides.furniture) and jupedsim walkable-area subtraction consume.
  • Pipeline + build guide: docs/FURNITURE-CATALOG.md; priors tool monad_knowledge.gis.furnish.abo_priors.