archcad-400k

What it is

ArchCAD-400K (Luo et al., arXiv 2025) is a large-scale architectural CAD drawing dataset built for panoptic symbol spotting — 5,538 highly standardized drawings decomposed into 413,062 chunks, with primitive-level vector annotations across 27 categories (structural, non-structural, and symbols). It is ~26× larger than the previous largest CAD dataset (FloorPlanCAD).

The reason it matters here is its building-type mix: unlike every floorplan corpus we currently load — all residential (resplan, zind, 3d-front, and structured3d, which has a single office across 3,500 scenes) — residential structures are only 14% of ArchCAD; the majority are offices and large-scale public / commercial facilities (average full-drawing area ~11,000 m²). It is the missing public-building analogue of ResPlan: the corpus the placement-oracle can run on when the venue is a library, office block, mall or civic building rather than a flat.

Paper: "ArchCAD-400K: An Open Large-Scale Architectural CAD Dataset and New Baseline for Panoptic Symbol Spotting" — arXiv:2503.22346. Repo: github.com/ArchiAI-LAB/ArchCAD. Data: huggingface.co/datasets/jackluoluo/ArchCAD.

Specs (repo structure verified; content pending access)

HF repo layout (data/, confirmed 2026-07-13 via the public tree API — the blobs themselves are still gated behind author review):

So the mirror is the ~800 MB vector layer (svg+json+point+caption) out of a ~2.6 GB repo — png rasters have no consumer in our vector→PostGIS path.

• Scale: 5,538 drawings / 413,062 chunks; 27 semantic categories.
• Units / CRS: metric CAD coordinates — confirm by parsing a real svg/json pair (the project page quotes area in m² but not the coordinate convention). The converter must read the drawing's own unit scale, not assume.
• Building-type taxonomy: "offices + public/commercial" confirmed; explicit library/museum labels not enumerated publicly — inspect the per-drawing caption/json metadata once unlocked to filter the venue types we want.

Per the datasets contract, do not trust the README — parse a handful of real drawings for the true primitive schema, unit scale and building-type facet before writing the manifest.

Acquisition & storage

• Status: gated — not yet in our S3. Access requested 2026-07-13 (HF user Sibyx, token in .env HF_TOKEN); as of that date the request is awaiting repo-author review (file resolve returns HTTP 403; the tree API is public). Non-commercial research use only. Mirror the moment it's approved.
• When acquired, follow the structured3d precedent — mirror the vector layer only (svg.zip+json.zip+point.zip+caption.zip, ~800 MB), skip png.zip (1.83 GB rasters). Target s3://monad-knowledge/datasets/archcad-400k/ with a _manifest.json carrying per-drawing facets (building_type, area_m2, n_rooms, bbox, chunk_ids) so scene selection needs no prefix listing.
• Converter: new notebooks/floorplans/converters/NN_archcad_s3.ipynb, mirroring 01_resplan.ipynb / 03b_structured3d_s3.ipynb — reads the manifest, selects candidate public-building drawings, fetches only those, converts the primitive graph → room polygons + walls + doors, loads into the IP-040 IndoorGML PostGIS schema via gis.loader.load_site (dry-run unless APPLY=True).
• Redistribution restricted — our bucket is private; do not republish.

Why it matters here

• Closes the public-building floorplan gap. The placement-oracle (c-placement-oracle) and all coverage×footfall crowd work currently run on residential geometry only. ArchCAD supplies offices/public/commercial layouts at scale — the venue diversity the AP-placement and CSI-counting claims need to generalise beyond flats.
• Furnishable & simulatable. Once in PostGIS, drawings feed the gis/furnish grammar placer → jupedsim crowds → Sionna CSI, exactly like the residential corpora — but now over large public spaces where footfall and coverage actually matter.
• Real, not synthetic. Complements the synthetic-clean structured3d / 3d-front geometry and the real-scan calibration bundle (indoor-scene-scans) with real drafted public-building plans.