MSD: A Benchmark Dataset for Floor Plan Generation of Building Complexes

Casper van Engelenburg^*,1, Fatemeh Mostafavi¹, Emanuel Kuhn¹, Yuntae Jeon², Michael Franzen³, Matthias Standfest³, Jan van Gemert¹, Seyran Khademi¹.

¹ Delft University of Technology
² Sungkyunkwan University
³ Indepedent Researcher

European Conference on Computer Vision (ECCV), 2024
^*Main author — c.c.j.vanengelenburg@tudelft.nl

arXiv Code Dataset

A realistic floor plan dataset — In this article, we propose and benchmark a new large-scale floor plan dataset, which we call [M]odified [S]wiss [D]wellings. MSD exceeds other floor plan dataset, such as RPLAN and LIFULL, in terms of complexity and better represents what could be found in the real-world.

Current models are not powerful enough — Most interestingly is that today's models that are designed for floor plan generation – mostly specialized neural networks – seem not robust and/or smart enough to tackle the complexity of MSD. Exciting new research in floor plan generation as well as understanding lies ahead.

Multimodality and coding guidelines — The floor plans come in various linked modalities: image, geometry, and graph. The main data container is the graph (`networkx.Graph()` or `torch_geometric.data.Data()`) on-top-of which the room shapes and types (as node-level attributes), the connectivity types (as edge-level attributes) and the full image (as graph-level attribute) are modelled. Additionally, we prepared clear coding guidelines on how to effectively load and visualize the floor plans and their corresponding access graphs.

MSD Dataset

MSD is an ML-ready dataset for floor plan generation and analysis at building-level scale. The MSD dataset is derived from the Swiss Dwellings database (v3.0.0). The MSD dataset contains 5372 highly-detailed floor plans of single- as well as multi-unit building complexes across Switzerland, hence extending the building scale w.r.t. of other well-known floor plan datasets. Some highlights:

Origin — MSD is the first European large-scale floor plan dataset originating from Switzerland;
Realism — MSD is the first large-scale floor plan dataset that contains information on the connectivity between the apartments; in addition, compass orientation is included;
Complexity — MSD exceeds other datasets in the the average number of corners per room, rooms per apartment, and apartments per floor; furthermore, MSD has the largest share of irregularly-shaped rooms;
Multimodality — MSD contains floor plan raster images, vectorized formats, and corresponding access graphs;
Diversity — MSD has the largest diversity of access graphs.

MSD compared to RPLAN and LIFULL — Rooms are colored on function (e.g., blue for "bedroom"). The functional diagrams (represented as graphs) are drawn on top of the floor plans. MSD (right) significantly differs from RPLAN (left) and LIFULL (middle), as it contains more complex and realistic floor plans.

Representation of floor plan data — MSD contains three types of data. (Left) The floor plan is the complete layout of the floor of the building, including the position and shape of the rooms, doors, and windows. The areas are labeled by color (see legends). There are two category systems: 1) category is based on the type of the area (’room type’), in which each area has a room type (e.g., "bedroom"); and 2) category is based on the zone of the area (’zone type’), in which each areas has a zone type (e.g., "zone1"). (Middle) The associated room and zoning graphs are depicted on the right of the floor plans. The node colors are equivalent to the colors of the floor plan. The position of each node is taken as the centroid of the area that the node represents. (Right) A binary image of the necessary structural components of the floor plan.

Area and unit distributions MSD — The unit distribution per floor (right), area distribution per floor (middle), and area distribution per unit (right) are plotted as histograms. The x-axis specifies the number of units or areas, and the y-axis specifies the frequency. From the unit distribution plot, it is apparent that MSD comprises mostly floor plans that contain between 2 to 9 units. MSD comprises mostly floor plans that have between 15 and 50 areas, with a peak of around 25. The area distribution per unit is similar to RPLAN and LIFULL, ranging between 3 and 15 areas per unit and a median around 7 areas per unit.

Models

We developed two baseline models to benchmark MSD: a diffusion- and segmentation-based. The former is built on-top-of HouseDiffusion (HD) — a state-of-the-art model for floor plan generation. The latter combines a U-Net and graph convolutional network. Details of the models can be found in the figures below.

Model 1: Modified HouseDiffusion (MHD) — Our diffusion-based baseline is built on-top-of HouseDiffusion (HD) — a state-of-the-art model for floor plan generation. A wall encoder is used to map the pre-processed building structure into corresponding wall embeddings. MHD expands HD by introducing an extra attention module (WCA) between the wall embeddings and corner features of the rooms. A graph attention network is separately trained to predict the room types from the zoning types, which are used to "color" the full layout.

Model 2: Graph-informed U-Net (UN) — UN takes the building structure (image) as input to a U-Net. The U-Net is composed of an encoder and decoder using the conventional up- and down-sampling 2D convolutions, resp., and includes skip connections between the encoder and decoder feature maps at equivalent feature map scales. A graph convolutional network is used to map the zoning graph to a feature vector which is concatenated to the latent space of the U-Net.

Results

Overall, the floor plans often look infeasible. We could, however, train MHD on RPLAN successfully (see suppl. mat.); hence, we believe that the poor results do not come from improper training. Instead, we attribute the somewhat poor results to the more complex benchmark we set: more complex graphs; more irregularly shaped rooms; unit connectivity; no axis alignment; etc.

Example generations of MHD and UN — Columns 1 and 2 show the inputs: the zone graph and building structure respectively. Columns 3 - 6 show the floor plans generated by the MHD variants. Columns 7 - 9 show the floor plans generated by the UN variants. Column 10 shows the ground truth.

MSD: A Benchmark Dataset for Floor Plan Generation of Building Complexes

MSD Dataset

Models

Results

Example generations of MHD and UN — Columns 1 and 2 show the inputs: the zone graph and building structure respectively. Columns 3 - 6 show the floor plans generated by the MHD variants. Columns 7 - 9 show the floor plans generated by the UN variants. Column 10 shows the ground truth.

Paper (incl. suppl. mat.)

BibTeX