ResPlan: A Large-Scale Vector-Graph Dataset of 17,000 Residential Floor Plans
The paper introduces ResPlan, a large-scale vector-graph dataset comprising 17,000 metric-scale residential floor plans with detailed room annotations and connectivity graphs, designed to advance research in semantic labeling, generative modeling, and spatial reasoning beyond the limitations of existing raster-based or floor-plate-level datasets.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to teach a robot how to understand a house. You could show it a photograph of a floor plan, like a picture of a map drawn on paper. The robot can learn to recognize that a dark line is a wall and a colored square is a bedroom. But a photograph is just a flat image; it doesn't tell the robot how the rooms are connected, how big they really are in meters, or which door leads to which room. It's like looking at a puzzle from the side—you can see the colors, but you can't see how the pieces fit together.
To build a truly smart robot that can design or reason about homes, scientists need more than just pictures. They need "vector" data, which is like a digital blueprint where every wall and door is a precise mathematical shape with exact measurements. They also need a "graph," which is a fancy word for a map of connections, showing that the kitchen is next to the dining room, but the bathroom is only reachable through a specific door. For a long time, researchers had to choose: they could have a huge library of simple pictures, or a small library of detailed blueprints, but rarely both. This new paper introduces a massive new tool that finally gives them everything they need in one package.
The Big New Map: ResPlan
Meet ResPlan, a giant new dataset that acts like a super-powered library for teaching computers about house layouts. The researchers at Indiana University have gathered 17,000 residential floor plans and turned them into a format that computers can really understand. Think of it as taking a messy pile of real estate brochures and turning them into a perfect, digital Lego set where every piece knows exactly where it belongs.
Unlike older collections that were just low-resolution pictures (like a blurry photo of a map), ResPlan provides vector geometry. This means every wall, door, and window is stored as a precise shape with real-world measurements in meters. If a wall is 21 centimeters thick in the real world, the computer knows it's 21 centimeters thick, not just "a thick line."
But the real magic is the graph. In this dataset, every room is a "node" (like a dot on a map), and the connections between them are "edges" (lines connecting the dots). The researchers didn't just draw lines; they labeled them with four specific types of connections:
- via_door: You can walk from this room to that one through a door.
- via_window: You can see or pass through a window.
- adjacency: The rooms share a wall but don't have a direct opening.
- direct: The rooms touch directly without a barrier.
This level of detail allows computers to do things they couldn't do before, like figuring out if a bathroom is isolated or if a kitchen is too far from the living room.
Why This Matters: Breaking the "Simple House" Barrier
Before ResPlan, the most popular dataset for this kind of work was called RPLAN. RPLAN had a lot of plans (about 80,000), but they were mostly simple, small apartments with an average of about 6.7 rooms. The most complex house in that dataset had only 8 functional rooms.
ResPlan changes the game. It includes everything from small apartments to massive, multi-wing villas. The average house in ResPlan has 8.1 functional rooms, and some have as many as 22. In fact, 34.6% of the houses in the ResPlan test set are more complex than the most complex house in the old RPLAN dataset. This is a big deal because it means researchers can finally train their AI on "real" big houses, not just tiny, simple ones.
The dataset also covers a wide variety of home types, including apartments, villas, and multi-wing residences, all with consistent labeling for 17 different categories (like bedrooms, bathrooms, kitchens, balconies, and storage).
What the Computer Can Do Now
The authors didn't just dump the data; they tested it to see what computers could learn from it. They set up three challenges:
- Naming the Rooms: Can the computer look at a graph of connections and guess what each room is? Using a smart algorithm called GraphSAGE, the computer got it right 95.1% of the time. This is much better than older methods, proving that understanding how rooms connect helps the computer "know" that a room next to a kitchen is likely a dining area, even if it looks like a bedroom.
- Building New Houses: Can the computer generate a new floor plan if you give it a list of rooms and a map of how they should connect? The results were mixed but interesting. Some methods could get the number of rooms right but ended up with weird shapes (like a giant living room swallowing everything). However, the data showed that as houses get bigger (more than 9 rooms), the computer's ability to keep the layout logical starts to drop, suggesting that building complex homes is still a hard puzzle for AI.
- Reading the Map: Can the computer look at a floor plan and draw the connection graph? They found that simply checking if rooms touch each other was very good at finding connections (98.2% accuracy), but figuring out how they connect (door vs. window) required a bit more brainpower.
The Catch and the Future
While ResPlan is a huge step forward, the authors are careful to point out its limits. All 17,000 plans come from real estate listings in South Asia. This means the houses reflect the style, size, and layout habits of that specific region. If you try to use this data to design a house in New York or Tokyo, the computer might get confused because the "rules" of house building are different there. The data suggests that models trained on ResPlan won't automatically work for Western or East Asian architecture without more training.
Also, these are single-floor plans. The dataset doesn't include stairs, multiple stories, or furniture. It's a 2D map of the floor, not a 3D model of the whole building.
Despite these limits, ResPlan is a massive resource. It gives researchers a "canonical split" (a standard way to divide the data for testing) and code to help anyone start experimenting. By providing 17,000 detailed, connected, and measured floor plans, it removes the old trade-off where scientists had to choose between having a lot of data or having high-quality data. Now, they can have both, opening the door for smarter, more capable AI that can truly understand the complex geometry of our homes.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.