ResPlan: Vector-Graph Residential Floor Plans

• ResPlan is a large-scale vector-graph dataset of 17,000 detailed residential floor plans, featuring precise architectural annotations and connectivity graphs.
• It offers clean vector geometries with uniform metadata on architectural elements and functional spaces, supporting robust spatial analysis.
• An open-source pipeline standardizes geometry cleaning, alignment, and annotation refinement, enabling reproducible and simulation-ready research.

ResPlan is a large-scale vector-graph dataset of 17,000 residential floor plans introduced to advance spatial AI research through detailed, structurally rich, and realistic unit-level layouts rather than idealized or raster-only abstractions. Each plan is provided with precise annotations of architectural elements and functional spaces, together with structured room connectivity, so that geometry, semantics, and topology are available in a single resource. The dataset is designed as a general-purpose foundation for robotics, reinforcement learning, generative AI, virtual and augmented reality, simulation, and game development, while also supplying an open-source pipeline for geometry cleaning, alignment, and annotation refinement (Abouagour et al., 19 Aug 2025).

1. Scope and defining characteristics

ResPlan is explicitly a residential floor-plan corpus, and its primary unit of representation is the individual dwelling rather than a larger building-level aggregate. The plans span compact apartments through large, multi-wing homes with irregular room shapes, complex topologies, and realistic proportions, reflecting non-idealized real-world layouts sourced from public real estate listings. This emphasis on realism is central to the dataset’s role in spatial AI: it is intended not merely as a geometric archive, but as a benchmark substrate in which architectural structure, semantics, and downstream usability are jointly prioritized.

The dataset’s core defining property is its vector-graph formulation. Floor plans are not only stored as clean vector geometries suitable for precise spatial computation; they are also accompanied by room connectivity graphs that expose adjacency and passage structure directly. This dual representation distinguishes ResPlan from resources that require raster-to-vector conversion, graph reconstruction, or unit extraction before they can be used in simulation or symbolic reasoning.

A recurring implication of the design is that ResPlan treats floor plans as structured environments rather than images. This suggests a shift from purely visual floor-plan modeling toward integrated geometric-semantic-topological learning.

2. Representation, annotations, and schema

ResPlan provides all plans in clean vector format, with detailed annotations covering both architectural elements and functional spaces. Architectural annotations distinguish walls by type, including interior, exterior, and neighbor/party walls; they also include doors, with the main/front door identified separately, as well as windows, balconies, stairs, garden/land, and related structural features. Functional spaces are semantically labeled using a unified taxonomy, with instance-level polygons for categories such as bedrooms, bathrooms, kitchens, living rooms, storage, parking, and pools (Abouagour et al., 19 Aug 2025).

Each sample is stored as a JSON dictionary keyed by semantic category. Values are POLYGON or MULTIPOLYGON geometries, and each record also includes metadata such as unitType, total floor area, planID, and wall_depth. A representative schema is:

{ "bedroom": "MULTIPOLYGON(...)", "living": "MULTIPOLYGON(...)", "garden": "MULTIPOLYGON EMPTY", "front_door": "POLYGON(...)", "wall_depth": 0.15 }

The schema is intentionally uniform. Categories that are absent in a given plan are still present as MULTIPOLYGON EMPTY, which removes special-case handling in downstream pipelines. The geometric and semantic fields are described as uniform and always present, and index stability is preserved so that the mapping between annotated room polygons and graph node indices is one-to-one and persistent.

ResPlan also includes metadata beyond raw geometry. Unit type, total floor area, and plan identity support stratified analysis, retrieval, and benchmark partitioning. Because wall depth is stored once per record and structural categories are normalized, direct downstream conversion and batch processing become substantially simpler.

3. Cleaning, alignment, and annotation refinement pipeline

A major contribution of ResPlan is its open-source pipeline for converting heterogeneous source material into a normalized vector-graph corpus. The pipeline begins with source ingestion, parsing public real estate plan documents such as CAD or PDF files and extracting polygons and text labels. It then performs geometry cleaning, including snapping and merging walls to enforce correct alignments, fixing small gaps or overlaps, removing duplicate or sliver polygons, and enforcing uniform wall thickness for conversion consistency (Abouagour et al., 19 Aug 2025).

The next stage is annotation refinement. Divergent room and feature names are unified into a standardized ontology; the example given is the normalization of labels such as "Master Room" and "Guest Room" to "Bedroom". Structural distinctions are also refined, for example separating neighbor party walls from simple partitions. Door polygons are validated and precisely aligned to adjacent wall openings, using both automatic processing and human verification. Error correction further includes merging tiny nooks with adjacent spaces or relabeling ambiguous rooms.

A final manual verification stage checks door placement, connectivity, and annotation errors. The paper emphasizes that the full pipeline is public and implemented in Python, so that new corpora can be re-parsed or extended using the same process. In encyclopedic terms, this is significant because the dataset is not only a released corpus but also a reproducible curation methodology.

4. Connectivity graphs and simulation-readiness

Every ResPlan sample includes a derived room connectivity graph. In this graph, nodes correspond to room polygons and carry attributes such as room type, centroid, and area. Edges are typed as door, arch, or shared-wall, thereby distinguishing direct circulation, large open passage, and adjacency without passage (Abouagour et al., 19 Aug 2025).

The graph is provided both as JSON node-edge lists and as NetworkX objects for Python. Cross-referencing between geometry and graph is designed to be constant-time, with polygon and graph-node index correspondence reported as $O(1)$ by index. This feature makes the dataset directly suitable for graph neural networks, symbolic planning, and constraint-based spatial reasoning, since geometric primitives and topological relations remain synchronized.

Simulation-readiness is another central design goal. Because plans are provided in vector form and all geometries can be reliably extruded to 3D with no significant preprocessing, ResPlan supports fast conversion to simulation assets. Door and window polygons define mesh cutouts, while shared wall depth supports consistent extrusion. A common misconception is to equate simulation-readiness with photorealistic furnishing or scene completion; in ResPlan, the term refers to the reliability of geometry for direct 3D conversion and downstream use in embodied environments.

5. Benchmark tasks enabled by the dataset

ResPlan defines four representative open benchmark tasks that jointly span generation, labeling, translation, and reconstruction. These tasks formalize the dataset’s intended use not only as training data but also as an evaluation substrate.

Task	Objective	Evaluation
Constrained Floor Plan Generation	Generate a valid plan from constraints	Boundary IoU, room type match, graph edit distance, FID
Semantic Room Labeling	Assign functional room labels to unlabeled geometry	Accuracy, per-class precision/recall/F1, mean IoU
Plan–Graph Bidirectional Translation	Extract graph from plan or generate plan from graph	Edge precision/recall, graph edit distance, topology matching, IoU
3D Layout Reconstruction	Produce a closed 3D mesh from 2D vector plans	Volumetric IoU, mesh quality/watertightness, structural validity

For constrained floor plan generation, the objective is to produce a valid plan from constraints such as an outline polygon, target room set, or adjacency graph. The listed baselines are House-GAN++, Graph2Plan, and diffusion-based generators. For semantic room labeling, the baselines include rule-based heuristics, CNNs for image segmentation, and GNNs for node classification. For plan–graph bidirectional translation, the benchmark splits into Plan→Graph and Graph→Plan, with metrics such as graph edit distance, topology matching, and geometric plausibility. For 3D layout reconstruction, the objective is a closed 3D mesh, with baselines including rule-based wall/element extrusion and learned height prediction (Abouagour et al., 19 Aug 2025).

This benchmark design places ResPlan at the junction of geometric learning, scene understanding, and structured generation. A plausible implication is that the dataset is intended to support evaluation regimes where a single corpus can test geometry, semantics, and topology under a shared schema.

6. Relation to prior floor-plan resources and adjacent methods

ResPlan is positioned explicitly against existing datasets including RPLAN, MSD (Modified Swiss Dwellings), LIFULL, and CubiCasa5K. The comparison emphasizes several differences: direct vector-graph representation rather than raster-only storage, explicit and complete unit-level plans, rich unified semantic annotation, explicit room connectivity graphs, and simulation-readiness without significant preprocessing (Abouagour et al., 19 Aug 2025).

The paper’s comparative table reports the following broad pattern. ResPlan contains approximately 17,000 floor plans in vector + graph form; RPLAN is raster only (256×256); MSD is vector + graph but is not unit-level and requires extraction; LIFULL is primarily raster; and CubiCasa5K is raster + vector with a stronger furniture focus. ResPlan is also described as having a wider distribution of room counts, with nearly 50% of plans being 2-bedroom while also including hundreds of large, multi-bedroom, multi-wing layouts missing in RPLAN and MSD.

In the broader research landscape, ResPlan’s benchmark tasks intersect with earlier work on automated floor-plan generation. For example, “Automated Generation of Dimensioned Rectangular Floorplans” develops a linear optimization framework for rectangular arrangements under minimum width and aspect ratio constraints and demonstrates regeneration of legacy rectangular floorplans from images (Upasani et al., 2019). ResPlan addresses a different layer of the problem: it provides realistic annotated data and graph structure for learning and benchmarking, whereas the earlier optimization-centered formulation focuses on feasible synthesis under explicit geometric constraints. This suggests complementarity rather than methodological overlap.

7. Applications, impact, and research significance

The immediate use cases named for ResPlan include generative AI, spatial reasoning, robotics and reinforcement learning, VR/AR and game development, simulation, and automated design tools. In generative modeling, the dataset supports GANs, diffusion models, and graph-based plan synthesis. In spatial reasoning, its structured graphs enable symbolic reasoning, constraint satisfaction, and architectural validation. In robotics and RL, the vector representation enables direct 3D conversion for navigation, mapping, and embodied tasks. In virtual environments and games, the plans support rapid population of realistic spaces and subsequent stylization or extension (Abouagour et al., 19 Aug 2025).

The paper also attributes a broader community role to the dataset. ResPlan is described as bridging the gap between architectural realism and ML/simulation-readiness, standardizing a unified and open pipeline for reproducible research, and supporting methodology development across geometry, semantics, and topology in one high-diversity resource. Its open, permissive release and reproducible processing pipeline reinforce that role.

Taken together, ResPlan can be understood as a structured benchmark infrastructure for spatial intelligence: a corpus in which realistic residential geometry, normalized semantics, and explicit connectivity are co-released so that floor-plan research need not choose between realism, annotation richness, and operational usability.