Modified Swiss Dwellings Benchmark

• MSD is a comprehensive dataset of floor plans featuring complex, multi-apartment Swiss dwellings with diverse geometries and structural details.
• The dataset offers multiple modalities including images, native vector geometries, detailed room graphs, and binary structural representations.
• MSD benchmark challenges models to preserve room topology, geometric fidelity, and load-bearing constraints in realistic residential layouts.

Searching arXiv for the MSD papers and the underlying HouseDiffusion paper to ground the article and citations. Modified Swiss Dwellings (MSD) is a floor plan dataset and benchmark for generative and predictive modeling of complex residential building layouts, with an explicit emphasis on medium- to large-scale, multi-apartment dwellings rather than predominantly single-apartment plans. It was introduced to address a mismatch between widely used public datasets and real-world building complexity, and it provides images, native vector geometries, graphs, and binary building-structure inputs for each plan. MSD is also the basis of a benchmark showing that existing floor plan generation methods degrade substantially when confronted with irregular room polygons, inter-unit connectivity, preserved orientation, and predefined load-bearing structure (Engelenburg et al., 2024).

1. Origin and curation

MSD is curated from the Swiss Dwellings dataset and undergoes a multi-stage cleaning pipeline designed to retain only unique, residential, structurally meaningful plans. The reported stages are feature removal of non-floorplan geometries such as bathtubs and stairs, residential-only filtering, near-duplicate removal by retaining only the floor plan at the lowest elevation for each building plan ID, scale filtering to keep only layouts with at least 15 areas and at least two “zone 2” public spaces, and additional cleaning that yields the final curated set (Engelenburg et al., 2024).

The resulting dataset contains 5,372 unique, mid- to large-scale floor plans, more than 18,900 distinct apartments, and approximately 163,500 labeled room areas after near-duplicate removal. The dataset is explicitly positioned as the first large-scale floor plan dataset with a significant share of layouts of multi-apartment dwellings, and the associated benchmark states that code and data are open (Engelenburg et al., 2024).

The design objective is methodological rather than purely archival. Earlier public datasets are described as being dominated by simple, single-apartment dwellings, mostly Manhattan-shaped rooms, limited inter-apartment structure, and substantial redundancy. MSD is constructed to preserve the aspects that complicate downstream modeling: multiple units per floor, richer connectivity, non-axis-aligned geometry, and load-bearing constraints. A plausible implication is that MSD is intended not only as a training corpus but as a stress test for model classes that were calibrated on simpler domestic layouts.

2. Representation and annotation schema

Each MSD floor plan is represented in four modalities: an image with semantic room labels, native vector polygon geometries with coordinates in meters, graphs, and a binary image encoding building structure in the form of load-bearing features (Engelenburg et al., 2024). The native vector representation is important because the benchmark emphasizes that MSD is not image-vectorized after the fact; this minimizes vectorization artifacts relative to datasets derived primarily from raster sources.

At the room level, each area has three reported attributes: a polygonal shape, a room type, and a zoning type. The zoning labels are given as private, public, service, and outdoor. Graph annotations encode connectivity with nodes corresponding to areas and carrying position, geometry, type, and zone, and with edges labeled as passage, door, or front door. Two graph variants are provided: room-type graphs and zone-type graphs (Engelenburg et al., 2024).

The building-structure channel is not an auxiliary visualization but a modeled constraint. MSD includes extraction of load-bearing walls and columns, with thickened walls identified via quantile-based thresholding. This structural modality directly motivates conditional generation settings in which synthesized room layouts must be compatible with existing building fabric rather than unconstrained interior partitioning (Engelenburg et al., 2024).

3. Geometric and topological profile

MSD differs from prior floor plan datasets along several geometric and graph-theoretic axes. The benchmark reports an average of 8.68 corners per room, 8.75 rooms per unit, and 3.52 units per floor, compared with 4.54 corners per room in LIFULL and 5.04 in RPLAN. MSD also includes non-Manhattan rooms, room labels, zoning labels, and doors, while retaining original compass orientation instead of normalizing away real-world directional context (Engelenburg et al., 2024).

Topological diversity is summarized through graph entropy,

$$\mathrm{H}_g = - \sum_{g \in \mathcal{G}} p_g(g) \log p_g(g),$$

with reported values of 8.02 for MSD, 7.79 for LIFULL, and 4.56 for RPLAN (Engelenburg et al., 2024). In the benchmark interpretation, the higher value for MSD indicates a richer diversity of connection topologies.

A second distinguishing property is redundancy reduction. The dataset description reports minimal near-duplicates, less than 1%, in contrast to 50–70% in LIFULL and RPLAN. This matters for evaluation because strong performance on datasets with many near-duplicate plans can overstate generalization. MSD’s curation therefore attempts to make memorization less effective and structural reasoning more necessary (Engelenburg et al., 2024).

These characteristics jointly explain why MSD is materially harder than legacy corpora. The benchmark explicitly identifies several sources of difficulty: more areas per layout, multiple apartments per floor, rich inter-connectivity within and across units, irregular polygonal rooms, structural constraints from predefined load-bearing elements, graph-diverse inputs, and the absence of axis alignment (Engelenburg et al., 2024).

4. Benchmark tasks, metrics, and empirical results

The MSD benchmark evaluates both raster-style and vector-style approaches. The two reported families are a Graph-informed U-Net, with graph conditioning supplied by a Graph Convolutional Network and optional preprocessing of the building-structure input, and a modified HouseDiffusion pipeline, including variants with wall cross attention and either full polygons or minimum rotated rectangles (MRR) as room-shape targets (Engelenburg et al., 2024).

The evaluation uses Mean Intersection-over-Union and graph compatibility. The benchmark gives

$$\mathrm{MIoU}(\mathcal{Q};\mathcal{K}) = \frac{1}{N} \sum_{i=1}^{N} \sum_{c \in \mathcal{C}} \frac{R_c(Q_i) \cap R_c(K_i)}{R_c(Q_i) \cup R_c(K_i)}$$

and

$$\mathrm{Compatibility}(\mathcal{Q};\mathcal{K}) = \frac{1}{N} \sum_{i=1}^{N} \frac{1}{|\mathcal{E}^k_i|} \sum_{e \in \mathcal{E}^k_i} \mathbf{1}[e \in \mathcal{E}^q_i].$$

The first measures pixel-wise overlap; the second measures the fraction of target graph edges preserved in the prediction (Engelenburg et al., 2024).

The reported average results are as follows:

Method	MIoU	Compatibility
U-Net	32.5	n.a.
UN (with GCN)	40.6	n.a.
UN (with GCN + pre)	42.4	n.a.
MHD (POL)	10.9	80.3
MHD (POL) + WCA	17.9	71.1
MHD (MRR)	11.5	87.1
MHD (MRR) + WCA	21.8	76.2

The benchmark interpretation is nuanced. U-Net-based methods achieve substantially higher MIoU, indicating better pixel-level overlap, but they do not reliably produce explicit room graphs. Modified HouseDiffusion variants preserve connectivity topology more effectively, but they struggle with precise placement and shape fidelity. Wall cross-attention improves geometry alignment while reducing graph compatibility, and MRR simplification improves graph compatibility without solving area-fit limitations (Engelenburg et al., 2024).

A common misconception is that a higher segmentation metric is sufficient to solve floor plan generation. The benchmark directly counters this by separating raster overlap from graph preservation: UN variants are stronger on MIoU, whereas MHD variants are stronger on compatibility. Conversely, high graph compatibility does not imply accurate geometry. The MSD results therefore expose a multi-objective problem in which topology, semantics, and geometric fidelity remain only partially aligned.

5. HouseDiffusion adaptation for MSD

The technical report on adapting HouseDiffusion to MSD begins from the HouseDiffusion formulation in which each diffusion timestep contains an embedding for every room polygon corner, denoted $C^t_{i,j}$ with $i$ indexing the room and $j$ the corner within the room, and the model predicts 2D position noise per corner. The original transformer stack uses three attention mechanisms: Component-wise Self Attention among corners within the same room, Global Self Attention across all corners in a floorplan, and Relational Cross Attention between corners in connected rooms (Kuhn, 2023, Shabani et al., 2022).

MSD adds an external structural constraint: a binary mask for structural walls. To condition generation on that constraint, the adaptation inserts cross-attention layers into the transformer so that room-corner embeddings act as queries and structural-wall-corner embeddings act as keys and values, with a mask allowing each room corner to attend to all wall corners. Before that interaction, structural wall corners are themselves processed through self-attention transformer layers intended to propagate context among neighboring wall corners. The report states that this cross-attention is inserted in both continuous denoising steps for positions and discrete denoising steps for categorical attributes (Kuhn, 2023).

The wall geometry needed by HouseDiffusion is not given directly; it is extracted from the structural mask by a preprocessing pipeline. The reported steps are morphological thinning to skeletonize the wall mask, graph extraction from the skeletonized image using the sknw library, and post-processing to segment polylines into straight-line segments and filter very short segments. The output is a set of straight wall lines suitable for geometric attention (Kuhn, 2023).

A further adaptation concerns room polygons. Because MSD rooms have variable corner counts and higher geometric complexity than RPLAN, the report approximates each room polygon by its minimum rotated rectangle, defined as the smallest area rectangle containing the polygon regardless of axis alignment. This enforces a constant four-corner representation per room and therefore matches the model’s expectations more closely. At inference, overlapping rooms are ordered by area, plotted largest first, and when a rectangle is split into multiple pieces by structural walls, only the largest piece is assigned to the room. The report states that the MRR approximation attains an IoU of 0.76 against the ground-truth masks (Kuhn, 2023).

The MSD adaptation also addresses a label mismatch: MSD provides zoning types, whereas the original HouseDiffusion pipeline expects room types. The reported solution is a Graph Attention Network trained on zoning and connection types to predict room types per node, using several GATConv layers, with the best result at three layers, and an output concatenated with initial features before MLP prediction (Kuhn, 2023).

6. Limitations, interpretation, and open research directions

The benchmark and technical report converge on the conclusion that MSD exposes failure modes not apparent on simpler datasets. The technical report states qualitatively that predicted room polygons generally follow wall structure but lack precision, especially when there are many small rooms, and that room layouts resemble the ground truth visually while corner alignment to structural walls can still be poor. It further reports that structure-mask refinement improves IoU across all classes, but that the mean IoU between refined predicted polygons and ground-truth labels without background is approximately 0.23 on validation, far below the 0.76 achievable by the MRR approximation to the ground-truth masks themselves (Kuhn, 2023).

These findings delimit what structural conditioning currently accomplishes. It helps, but it does not guarantee tight geometric adherence to load-bearing walls. Likewise, rectangle simplification stabilizes representation and sometimes improves compatibility, but it reduces fidelity for highly irregular rooms and complex wall constraints. The reported absence of ablations for the extra transformer encoder layer and for the placement of cross-attention means that some architectural choices remain empirical rather than isolated causal findings (Kuhn, 2023).

The open directions named in the papers are specific. They include moving beyond minimum rotated rectangles toward direct modeling of variable-length polygons or more flexible parametric representations; exploring alternative encodings of wall masks or wall graphs, including direct raster input and more sophisticated geometric representations; performing ablations on structural cross-attention placement and on structural transformer encoders; testing whether wall conditioning should appear only in certain denoising stages; improving data efficiency through augmentation, transfer learning, or semi-supervised techniques; and developing training or architectural mechanisms that encourage tighter wall alignment (Kuhn, 2023). In the benchmark framing, these limitations imply that structurally constrained, graph-conditioned, multi-apartment floor plan generation remains an open problem and that MSD functions as a benchmark for floor plan machine understanding rather than a task already solved by existing methods (Engelenburg et al., 2024).