philippds/modified-swiss-dwellings-enriched

Modified Swiss Dwellings (MSD), enriched

Floor plans of medium-to-large multi-apartment building complexes (ECCV 2024 benchmark MSD), each linking three modalities of one floor plan: image, geometry, and access graph.

1. Why this is here & what was done

Hosted on Hugging Face for reach and one-line loading by the ML community. The Swiss Dwellings (SD) license (CC BY 4.0) permits redistribution with attribution — so this also enriches the public MSD release, which had dropped all provenance columns.

We recovered them by finding that the MSD sample id is exactly the SD floor_id, then joining back to SD via an exact room-centroid match. This adds: SD identifiers (site/building/plan/unit/apartment/area), the original fine-grained room subtypes, and the full SD simulations (view/sun/noise/connectivity) and location (climate + walkability) metadata. Credit for the underlying data belongs to the original authors (see bottom).

2. Splits & configs

• Splits: train (4,572) and test (800) — 5,372 floor plans total.
• Configs (joined on id):
  • default — core MSD: images, coordinates, both graphs, ids, per-room attributes (27 cols)
  • simulations — id + 361 per-room sim_* metrics
  • locations — id + 508 per-building location_* / location_rating_* columns
  • geometries — the full flat MSD table: one row per geometry element (every room, wall, door, window) with WKT geom + all SD columns. Use this when you need the raw vector geometry of non-room elements (walls/openings), which the default graph omits.

3. Columns (default config)

Example values from the first sample (id=23, a floor with N=52 rooms and E=54 graph edges). Per-room arrays share one node order; graph edges are source[i]→target[i].

Column	Type	Example (id=23)
id	int32	23 (= SD floor_id)
site_id / building_id / plan_id	int32	79 / 8 / 8
struct_in	Image L	512×512 structural-elements mask
full_out	Image P	512×512 room-type mask (pixel = class 0–12, palettized)
struct_x / struct_y	float32[512]	world coords of the image axes, e.g. [-35.19, -35.03, …]
full_x / full_y	float32[512]	as above for full_out, e.g. [-15.80, -15.74, …]
graph_in_zoning_type	int16[N]	[0, 0, 3, 2, 0, 1, …] (8 zoning classes)
graph_in_edge_source / _target	int16[E]	[0,1,2,…] / [8,8,5,…]
graph_in_edge_connectivity	string[E]	["door","door","door", …] (values: passage / door / entrance)
graph_out_room_type	int16[N]	[0, 0, 8, 7, 0, 1, …] (13 room classes)
graph_out_subtype	string[N]	["ROOM","ROOM","BALCONY","BATHROOM", …] (orig. ~25-way subtype)
graph_out_centroid	float32[N][2]	[[25.44, 6.27], [31.70, 7.34], …]
graph_out_geometry	float32[N][…][2]	room polygon rings, [[[27.92,4.36],[24.60,3.25], …]]
graph_out_edge_source / _target / _connectivity	int16[E] / str[E]	access graph over rooms
graph_out_area_id	int64[N]	[718090, 382415, …] (SD primary key → other configs)
graph_out_unit_id	int32[N]	[21037, 21037, …] (null for public rooms)
graph_out_apartment_id	string[N]	["1f872b…", …] (null for public rooms)
graph_out_unit_usage	string[N]	["RESIDENTIAL", …, "PUBLIC"]
graph_out_elevation / graph_out_height	float32[N]	0.0 / 2.60 (metres)

Class legends — room_type: Bedroom, Livingroom, Kitchen, Dining, Corridor, Stairs, Storeroom, Bathroom, Balcony, Structure, Door, Entrance Door, Window; zoning_type: Zone1–4, Structure, Door, Entrance Door, Window.

simulations adds per-room sim_* (e.g. sim_layout_area, sim_sun_*, sim_noise_*); locations adds per-building location_* (climate normals, walkable POIs) and location_rating_*. Join them on id.

4. How to use

from datasets import load_dataset
REPO = "philippds/modified-swiss-dwellings-enriched"

ds  = load_dataset(REPO, split="train")                 # core (images + graphs)
sim = load_dataset(REPO, "simulations", split="train")  # per-room metrics
loc = load_dataset(REPO, "locations",   split="train")  # per-building metrics

# join extra metadata onto the core sample by id
merged = ds.to_pandas().merge(sim.to_pandas(), on="id")

Good for: floor-plan generation (image and/or graph conditioned), graph→layout and layout→graph learning, room-type / access-graph prediction, multi-apartment structure reasoning, and studies linking layout to environment (sun, view, noise, walkability).

Attribution (CC BY 4.0 — please cite both)

• MSD — van Engelenburg et al., MSD: A Benchmark Dataset for Floor Plan Generation of Building Complexes, ECCV 2024. arXiv:2407.10121
• Swiss Dwellings v3.0.0 (source of all recovered columns) — Standfest, M. et al. (2023), Zenodo. https://doi.org/10.5281/zenodo.7788422

With sincere thanks to the MSD authors and the Swiss Dwellings team; this repository only re-packages and re-links their open data and claims no new authorship.