s3dis

What it is

S3DIS (Armeni et al., Stanford 2D-3D-Semantics / Joint 2D-3D-Semantic Data, 2017) is a real indoor scan corpus of 6 large-scale areas (over 6,000 m², 271 rooms) drawn from 3 buildings of mainly educational and office use — offices, hallways, lobbies, conference rooms, lounges, copy rooms, open spaces, storage, WCs, auditoria. It ships mutually-registered 2D / 2.5D / 3D modalities with instance-level semantic annotations over 12 object categories (wall, floor, ceiling, beam, column, window, door, sofa, desk, chair, bookcase, board), plus registered point clouds and per-room geometry.

Its value here is being the cleanest source of real non-residential room geometry we can actually mirror. Our loadable floorplans are all residential; the one real non-residential source we track — the scan bundle indoor-scene-scans — is flagged gated, statistics-only, do NOT mirror. S3DIS is small enough (we need only room polygons + semantics, not the ~70k RGB images) to fully ingest as a floorplan corpus of a real institutional building.

Papers: "Joint 2D-3D-Semantic Data for Indoor Scene Understanding" (arXiv:1702.01105) and the CVPR 2016 "3D Semantic Parsing of Large-Scale Indoor Spaces". Portal: buildingparser.stanford.edu.

Verified specs (from the mirrored data, 2026-07-13)

Aligned v1.2 mirrored and parsed — 271 rooms across 6 areas, 10,293 files / 18.05 GB, 0 upload failures.

• On-disk layout: Area_N/<room>/<room>.txt (full-room XYZRGB point cloud) + Area_N/<room>/Annotations/<class>_<n>.txt (per-object point subsets).

• Point format: one point per line, X Y Z R G B space-separated, metres (aligned = canonical axis-aligned frame → floor outline extracts cleanly).

• Room-type distribution (the reason this set exists here — real institutional spread vs structured3d's 1 office):

  office	hallway	storage	WC	conferenceRoom	lounge	lobby	pantry	auditorium	openspace	copyRoom
  156	61	19	11	11	3	3	3	2	1	1

• Semantic classes present: wall, floor, ceiling, beam, column, window, door, table, chair, sofa, bookcase, board, stairs, clutter (+ a stray Icon macOS artefact in a couple of dirs — harmless, ignore in the converter).

• Floor/wall class points → room polygon; table/chair/bookcase/board → real furniture-arrangement prior for gis/furnish.

Acquisition & storage

• Status: acquired. Mirrored to s3://monad-knowledge/datasets/s3dis/ (Aligned v1.2 only — non-aligned zip + both .mat packs deliberately skipped; the imagery-heavy f304 2D-3D-S parent skipped entirely). Streamed zip-entry → S3 (no local extraction, laptop had only ~16 GB free). _manifest.json carries per-file facets (area, room, room_type, kind, object_class, bytes) plus rooms_by_type / annotations_by_class rollups, so room selection is a one-GET in-memory filter.
• Loading (S3-only, bounded): read datasets/s3dis/_manifest.json, filter files[] to the rooms you want (kind=="room_cloud"), get_bytes those; pull Annotations/* only for the objects you need. Room clouds are multi-MB text — stream/parse, never sync the whole 18 GB.
• Converter: notebooks/floorplans/converters/06_s3dis_s3.ipynb (mirrors 03b_structured3d_s3.ipynb). Reads the manifest → selects rooms (facet filter, richest-first, scoped by AREAS/ROOM_TYPES/N_ROOMS_PER_AREA) → fetches only the floor + furniture annotation files → room polygon = concave_hull(0.6) of grid-downsampled (5 cm) floor-class points with a convex-hull floor → perimeter walls → gis.loader.load_site (one Area = one single-level site; dry-run unless APPLY=True). Furniture is IP-113-native: each room's real chair/table/sofa/ bookcase/column objects become a FURNITURE arrangement via ensure_arrangement + write_items (footprint = object min-rotated-rect, height = Z-extent; crosswalk chair→chair, table→reading_table, sofa→sofa, bookcase→ shelf_single, column→column), aligned to the loader's origin_offset_cm. Validated end-to-end on real Area_1/Area_2/Area_4 rooms (2026-07-14) — polygon areas + furniture counts sane (auditorium → 293 seats), all cells run clean in dry-run. Known limitation: rooms are separate scans, so inter-room adjacency/doorways are NOT reconstructed — each room is independently walled (cross-room door transitions via the door-class points are a follow-on). Heights (z_min/z_max + gis rebuild-3d) are a follow-on, like the Structured3D height pass.
• Research-use terms — bucket private, do not republish.

Why it matters here

• Real institutional geometry we can load. Answers the "how do we know public-space geometry is realistic?" question indoor-scene-scans raises — with a real educational/office building the placement-oracle can run on directly, not just derive statistics from.
• Furniture-arrangement calibration. Real bookcase/desk/chair layouts per room type tune the gis/furnish theme grammar for offices and library-like spaces — complementing the first-party fiit-library-roomplan-2026-05-29 gold reference with a second, larger real institutional sample.
• Complements ArchCAD. archcad-400k gives drafted public-building plans at scale; S3DIS gives scanned real institutional rooms with furniture — drafted breadth + scanned ground truth.