Description

A bundle of real RGB-D / mesh scans with instance+semantic labels, tracked together because they serve one purpose here: ground-truth non-residential room geometry + furniture arrangements for calibrating and validating the grammar placer (which real datasets, unlike 3D-FRONT, actually contain offices, classrooms and common areas).

  • ScanNet (Dai et al., CVPR 2017) — 1,513 scans incl. offices, classrooms, common areas, copy rooms. source_url: scan-net.org (terms agreement).
  • Matterport3D (Chang et al., 3DV 2017) — 90 building-scale scans, region + object semantics, floor plans; mixed/non-residential buildings.
  • Replica (Straub et al., 2019) — 18 high-fidelity scenes incl. 5 office rooms; per-primitive semantics.
  • HM3DSem (Yadav et al., 2022) — 142k semantic instances over 216 spaces incl. stores + other public indoor spaces; broadest non-residential real coverage.

Role in this project

  • Calibration: measure real furniture density, size distributions and arrangement spacing per room type → tune the theme-grammar parameters (themes.py) and catalog dimension ranges so computed placements match reality.
  • Validation: compare a placer-furnished office/classroom against a scanned one (statistics, not pixel match) — the "how do we know it's realistic?" evidence a reviewer will ask for.
  • Complements the first-party gold reference: fiit-library-roomplan-2026-05-29 (IP-109 LiDAR, 103 measured furniture instances on fiit-library-floor-0).

Acquisition & storage

  • License: each requires a signed academic terms agreement (gated); none are freely redistributable. Pull only the room-type subset needed for a calibration pass — do NOT mirror wholesale to our S3.
  • We need extracted statistics (per-class dims, counts per m²), not the scans themselves — derive a small scan_priors.csv and keep the raw scans upstream.

Why it matters here

  • These are the only openly-studied sources of real non-residential furniture layouts; they turn the authored grammar from "plausible" into "calibrated".