ResPlan: A Large-Scale Vector-Graph Dataset of 17 000 Residential Floor Plans

Mohamed Abouagour, Eleftherios Garyfallidis · 2025

Read ResPlan: A Large-Scale Vector-Graph Dataset of 17 000 Residential Floor Plans by Mohamed Abouagour, Eleftherios Garyfallidis (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-14006.txt.

Introduction

Mohamed Abouagour and Garyfallidis release ResPlan, a dataset of 17,000 residential floor plans in vector form. Each plan marks walls, doors and windows. It labels rooms under a 17-class scheme, from kitchens to balconies. Coordinates come in both pixels and meters. Every plan also carries a room-connectivity graph with four kinds of edges.

Older datasets are raster images only, or cover whole building floors. ResPlan instead gives complete single homes, from apartments to villas. Plans average just over eight usable rooms. The release ships with loading code, a standard split, and baselines for three tasks. Those tasks are room labeling, floor-plan generation, and plan-to-graph extraction.

Catalog record

CREATORMohamed Abouagour, Eleftherios Garyfallidis
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-14006

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

ResPlan: A Large-Scale Vector-Graph Dataset of 17 000 Residential Floor Plans Mohamed Abouagour, Eleftherios Garyfallidis Abstract We introduce ResPlan , a large-scale dataset of 17,000 detailed, structurally rich, and realistic residential floor plans, created to advance spatial AI research. Each plan includes precise annotations of architectural elements (walls, doors, windows, balconies) and functional spaces (e.g., kitchens, bedrooms, bathrooms). ResPlan addresses key limitations of existing datasets such as RPLAN [ 12 ] and MSD [ 11 ] by offering enhanced visual fidelity and greater structural diversity, reflecting realistic and non-idealized residential layouts. Designed as a versatile, general-purpose resource, ResPlan supports a wide range of applications—including robotics, reinforcement learning, generative AI, virtual/augmented reality, simulations, and game development. Plans are available in both geometric and graph-based formats, enabling direct integration into simulation engines and fast 3D conversion. A key contribution is an open-source pipeline for geometry cleaning, alignment, and annotation refinement. Additionally, ResPlan includes structured representations of room connectivity, supporting graph-based spatial reasoning tasks. Finally, we present comparative analyses with existing benchmarks and outline several open benchmark tasks enabled by ResPlan. Ultimately, ResPlan offers a significant advance in scale, realism, and usability, providing a robust foundation for developing and benchmarking next-generation spatial intelligence systems. Figure 1: Example floor plan from the ResPlan dataset. 1 Introduction Floor plans are fundamental abstractions of indoor environments, encoding the spatial organization of rooms, structural elements, and access points. They are critical to a wide range of domains, including architectural design, robotics, indoor navigation, and generative modeling. In recent years, the availability of annotated floor plan datasets has catalyzed progress in tasks such as floor plan generation, image-to-plan translation, and embodied agent navigation. However, existing datasets often exhibit limitations in terms of structural complexity, geometric fidelity, and representation format. For example, the widely used RPLAN dataset [ 12 ] provides over 80,000 annotated floor plans in raster format. Despite its scale, RPLAN predominantly consists of simple, single-unit layouts with mostly rectangular room shapes. Moreover, its representation leads to irregular door symbol geometries and overlapping elements (all depicted in a single image layer), hindering precise geometric analysis and complicating downstream applications. More recently, the Modified Swiss Dwellings ( MSD ) dataset [ 11 ] addressed some of these issues by introducing a collection of multi-unit building layouts comprising 5,300 floor plans and 18,900 distinct apartments. While the MSD corpus provides valuable coverage of multi-unit floor plates, each record represents an entire floor plate encompassing multiple residential units. As a result, researchers aiming to extract self-contained apartment layouts must undertake non-trivial preprocessing. This includes filtering geometries by apartment identifiers and reconstructing missing architectural features—such as party walls and entrance doors—that are implicit or absent once adjacent units are excluded. This requirement introduces significant data-engineering overhead and reduces the dataset’s immediate utility for applications that assume complete, standalone unit-level floor plans. Beyond RPLAN and MSD, other datasets exhibit their own shortcomings. The LIFULL Homes dataset [ 7 ] contains millions of floor plan images scraped from a real-estate platform, but only a small fraction of these plans have been fully vectorized or semantically annotated, greatly limiting their utility for direct use in machine learning pipelines. For instance, the Raster-to-Graph system by Hu et al. [ 4 ] produced a vectorized subset of about 10,800 plans from LIFULL, but this remains a tiny portion of the overall corpus. Conversely, CubiCasa5K [ 6 ] offers 5,000 floor plan images with rich annotations (over 80 object categories including furniture), yet its scale is relatively small for training modern models. Collectively, these limitations highlight the gap between currently available datasets and the needs of applications requiring high-fidelity, unit-level representations of interior spaces. 2 Our Contribution To address this gap, we introduce ResPlan , a new large-scale dataset of high-fidelity residential floor plans designed to support both generative and simulation-based applications. ResPlan comprises 17,000 vector-based floor-plan instances, each capturing rich architectural details and diverse spatial configurations—ranging from standard apartments to complex, multi-wing homes. Unlike previous datasets, ResPlan is created with simulation readiness and structured representation in mind: all layouts are provided in a clean vector-graphic format and include graph-based representations of room connectivity, enabling seamless integration with learning models, virtual environments, and robotic simulators. In summary, the key contributions of ResPlan are: • Scale and Structural Complexity: ResPlan significantly advances the scale and complexity of available floor plan data. It contains 17k layouts encompassing a broad variety of residential designs, including many with irregular (non-rectangular) room shapes and intricate topologies. The dataset spans a wide distribution of home sizes, including several examples with five or more bedrooms (configurations that are rare or absent in prior datasets), thereby exceeding the diversity of layouts found in earlier benchmarks. • Realism and Annotation Quality: All floor plans are derived from real-world building designs (collected from public real estate listings), ensuring realistic architectural features and proportions. A rigorous cleaning process and manual verification have been applied to guarantee precision—especially for critical elements like wall alignments and door placements. This results in high-fidelity layouts that closely reflect practical design choices. • Rich Semantic Annotations: Each floor plan is annotated with a standardized taxonomy of room types and structural components. We unify inconsistent nomenclature across sources (e.g., MasterRoom or GuestRoom are both normalized to Bedroom ) and clearly distinguish structural elements (e.g., interior vs. exterior walls). These semantic labels are embedded in the vector data, making it straightforward to query functional spaces (bedrooms, kitchens, bathrooms, living areas, balconies, etc.) and structural elements (doors, windows, walls) across all layouts. • Graph-Structured Representation: In addition to geometric floor plan data, ResPlan provides an explicit graph representation for each layout. Rooms are represented as nodes (with attributes such as room type and area), and adjacencies between rooms are captured as typed edges (e.g., door or open-arch connections). This graph structure enables structured reasoning and facilitates the use of graph neural networks, constraint solvers, or symbolic planners on floor plan data (illustrated in Fig. 3 ). • Simulation-Ready Vector Format and Tools: ResPlan’s vector-graphic format allows direct conversion to 3D environments without laborious preprocessing. All walls, doors, and windows are represented as clean polygons with consistent thickness, ensuring they can be extruded into watertight 3D meshes. We provide an open-source processing pipeline that converts raw floor plan drawings into this standardized format, enabling researchers to reproduce our data construction. We also propose a suite of benchmark tasks (with evaluation metrics) built on ResPlan to spur future research (see Section 5 ). Table 1: Comparison of ResPlan with existing floor plan datasets. Feature ResPlan (Ours) RPLAN [ 12 ] MSD [ 11 ] LIFULL [ 7 ] CubiCasa5K [ 6 ] # Samples \sim 17,000 \sim 80,000 5,372 (18.9k units) \sim 5,000,000 5,000 Source Real estate listings Asian real estate Swiss dwellings Japanese real estate Multiple countries Format Vector + Graph Raster (256 \times 256) Vector + Graph Raster (images) Raster + Vector (SVG) Curation Cleaned, Filtered Manual Cleaned, Filtered Scraped (raw) Manual (+furniture) Vector Native Yes No (converted) Yes No Yes Unit-Level Plans Yes Yes No (extraction needed) Yes Yes Multi-Unit per Floor No No Yes No No Graph Connectivity Yes No Yes No No ResPlan offers a high-quality, simulation-ready benchmark for structured floor plan learning

Read the complete text