ModRes: A modification of the ResPlan dataset ready for floor plan analysis and graph machine learning

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "ModRes: A modification of the ResPlan dataset ready for floor plan analysis and graph machine learning".

Jane: The gist The ModRes dataset is a cleaned and enriched version of ResPlan, containing 8,431 valid and unique residential floor plans with structured geometric, topological,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we've established what ModRes is—a cleaned set of eight thousand four hundred thirty-one unique floor plans—and now we need to talk about what the actual research paper titled "ModRes: A modification of the ResPlan dataset ready for floor plan analysis and graph machine learning" actually says about the methodology <ref:2610.17178#pg1,ModRes: A modification of the ResPlan dataset ready for floor plan analysis>.

Jane: I think the key takeaway here is that they explicitly state their goal was to solve the preprocessing problem that made ResPlan hard to use reliably. They are focused on making it ready-to-use, which means fixing those errors before anyone tries to apply complex spatial analysis or graph machine learning techniques to it.

Lu: I read about how they filter for essential rooms and remove doors of unrealistically small dimensions, which shows a commitment to validity over just having a lot of data points. It’s about quality input for the machine learning models that follow >

Tom: Right, so the filtering process is the mechanism they used to ensure the results obtained using these floor plans will have validity and applicability to real-life residences.

Meng: I'm curious about how they handled those duplicate floor plans, because that’s where I see a major hurdle when we try to scale up models on real traffic or real building data.

Lalam: They tackled that by comparing vectors representing key values like the number of bedrooms and total area for each room type to eliminate those exact duplicates, which is a smart way to handle redundancy in training data >

Jane: And they also used size filtering based on a normal distribution to remove plans with unrealistic recorded areas, which helps keep the data realistic.

Tom: So, beyond just cleaning the data itself, what are the actual representations they give us for analysis? I need to know what kind of structures we can build on top of this.

Lu: They provide three complementary dual graphs: a primary planar graph representing walls and intersections, a dual graph where nodes are architectural components like rooms and doors, and another one that connects rooms through doors >

Tom: That’s a lot of representations for one floor plan. It seems they are trying to give us options depending on whether we want to focus on wall connectivity or room relationships >

Jane: Exactly, because those graphs are enriched with features like room type, area, perimeter, and compactness. This means you can use the continuous and categorical data in different ways for your analysis >

Meng: From my side, I need to know what the cost is to run this kind of graph representation; does it blow up memory when we try to process all eight thousand four hundred thirty-one plans <ref:2610.17178#pg1>?

Lalam: They provide these graphs in CSV and JSON formats, which means they are ready for standard data science tools and can be loaded into whatever system we use >

Tom: So the paper is essentially delivering a fully prepared dataset where the complexity of preprocessing has been done upfront. This gets us past the initial hurdle of getting usable input >

The paper's summary: Tom: Now let's talk about what ModRes actually summarizes in its introduction and abstract. It boils down to this: they took a dataset that was problematic because it had errors and duplicates, and they created a version that is ready to be used for floor plan analysis and graph machine learning.

Jane: The main point is the problem statement: publicly available datasets often require substantial preprocessing before they can be used reliably, so ModRes presents this cleaned version of the ResPlan dataset.

Lu: They are addressing the lack of readily available datasets in the Architecture, Engineering and Construction sector because floor plans aren't often released due to confidentiality risks >

Tom: So they are providing a solution to that access problem by creating a large-scale, validated resource that can be used for training machine learning models on building layouts.

Meng: If we think about this from a deployment perspective, having eight thousand four hundred thirty-one unique and valid plans means we have enough data points to start prototyping some classification models without immediately hitting the wall of bad data > <ref:2610.17178#pg1>

Lalam: And they emphasize that they provide multiple dual graph types for each floor plan, allowing users to select the representation appropriate for their specific task, which is a huge flexibility feature >

Jane: That flexibility comes from providing those different graph structures—the planar graph, the room connectivity graph, and others—which cater to different analytical needs >

Tom: So they are not just handing us a file; they are hand-delivering a multi-faceted data source where we can choose our analytical lens based on whether we need spatial analysis or something more graph-based >

Lu: And they make sure the topological and semantic representations are structured in a way that machine learning models can process them directly, which is crucial for the subsequent steps >

Jane: That structured input allows us to move from just looking at a drawing to running actual computations on the layout, like calculating accessibility or visibility scores >

The paper's improvements: Tom: Let's talk about the specific modifications they suggest as improvements over the original ResPlan dataset. They focused heavily on the cleaning workflow, which involved filtering for essential rooms and removing doors of unrealistically small dimensions.

Jane: They also made sure to remove plans with missing geometry or disconnected sections, which is vital because if a building plan doesn't connect properly, any spatial analysis will just break down >

Lu: A key improvement mentioned is the use of size filtering based on a normal distribution to remove unrealistic recorded areas, which helps filter out outliers that could skew our statistical analysis >

Tom: And they also eliminated duplicate floor plans by comparing vectors representing key values like room count and total area for each room type, which directly addresses the bias problem we talked about earlier.

Meng: So the improvements are focused on robustness—making sure the dataset is reliable enough to train models that actually generalize when we apply them to new, unseen buildings >

Lalam: They also provided graph-level features like Area, Net Area, and Perimeter calculated from the geometry, and node-level features like numerically encoded room type and compactness metrics for each node >

Jane: Those specific metrics are what make the graph representations useful; they turn raw geometric data into meaningful inputs for the machine learning models >

Tom: So they’re giving us pre-calculated metrics that save us from having to calculate all those area and perimeter values ourselves before we even start training anything.

Conclusion: Tom: So, wrapping up on this ModRes paper, the main implication is that they have delivered a practical resource for data-driven research on residential floor plans by correcting and extending ResPlan. It makes a single collection of floor plans available in several mutually consistent forms.

Jane: What this means for us in the industry is that we can move between spatial analysis, graph-based modeling, and building simulations using these different representations, which supports movement between those different computational methods >

Lu: For future work, they suggest focusing on incorporating multi-storey layouts because the current dataset is limited to single-storey buildings, which limits its generalisability to larger structures >

Meng: And another thing I'd add is that the dataset currently contains a large number of two-bedroom houses, which might limit the applicability of models we train on models trained on larger homes >

Lalam: Also, they noted that the dataset lacks data on furniture and interior fixtures, which are important for indoor navigation and robotics applications >

Tom: So, the main value here is making this dataset a benchmark for evaluating computational methods in residential floor plan analysis by providing these consistent forms.

Jane: Exactly. It’s a practical resource because it supports movement between spatial analysis, graph-based modeling, and building simulations >

Lu: I think the structure they provide—multiple dual graphs for each floor plan—is really powerful because it allows users to select the representation appropriate for their specific task >

Meng: And that flexibility is good, but we still need more data on different housing types to make these models truly applicable across the board >

Lalam: So, ModRes gives us a strong foundation for research right now, and future development should focus on incorporating multi-storey layouts and richer semantic annotations >

Tom: That’s the summary of ModRes. We have seen how they cleaned and enriched ResPlan to create a robust dataset for floor plan analysis and graph machine learning.

Jane: It’s a resource that lets us use those geometric, topological, and semantic representations in different ways for our specific computational needs >

Lu: And we should keep an eye on their future work regarding multi-storey layouts.

Meng: From an engineering side, the next step would be to see if these models hold up when we try to apply them to more complex or diverse housing stocks.

Lalam: ModRes is available on GitHub at https://github.com/emmashelley/ModRes, and it’s a resource for researchers looking to get started with this kind of data >