Raster‑to‑Vector Floorplan Reconstruction (R2V) for Standardized Housing Layouts

1. Motivation

Floor‑plan images (raster graphics) are ubiquitous but lack structured geometric and semantic information. Directly matching raster images yields only coarse visual similarity, whereas the true essence of a floor‑plan lies in its vector representation composed of points, lines, and surfaces that encode walls, doors, windows, and room layouts.

To enable precise house‑search and floor‑plan interpretation, these raster images must be standardized and reconstructed into high‑level semantic vectors that capture dimensions, positions, and spatial relationships of architectural elements.

Beike has pioneered such technology, applying 3D point‑cloud reconstruction and monocular reconstruction in 3D scenes, and developing a 2D raster‑to‑vector pipeline (FrameX) that standardizes floor‑plans, powers AI‑driven interpretation, and supports novel house‑search experiences.

Since 2020, the standardized floor‑plan has powered a “scan‑to‑search” feature for end‑users, and in 2021 it dramatically improved new‑home floor‑plan digitization efficiency (over 5× speedup) and enabled AI‑generated interpretation reports.

2. Definition

2.1 Raster and Vector Images

A raster image is a conventional photograph; a vector floor‑plan is a JSON file describing points (corners), lines (walls), doors/windows, and rooms (surfaces) with spatial coordinates.

2.2 “Standardization” and “Vectorization”

Vectorization (R2V) extracts point‑line‑surface structures from a raster image. Standardization further converts these vectors into a format that conforms to Beike’s internal data schema. The standardized vector enables downstream feature extraction and accurate floor‑plan retrieval.

3. Related Work

Traditional CAD tools (AutoCAD, SketchUp) produce vector drawings, but commercial listings often expose only raster renders, losing structural information. Academic and industrial research on floor‑plan vectorization falls into two categories:

3.1 End‑to‑End Element Recognition – Directly detects walls, doors, windows from images (e.g., Deep Floor Plan Net). Fast but limited by segmentation accuracy and open‑set variability.

3.2 DNN + Optimization – Methods such as Furukawa’s Raster‑to‑Vector (R2V) and FloorSP first predict element heatmaps with a DNN, then enforce geometric and semantic constraints via integer programming. This yields higher accuracy at the cost of increased solving time.

Key challenges identified include the diversity of floor‑plan styles, the need for real‑time inference, and the computational burden of large integer programs.

4. Modeling

The R2V pipeline follows a bottom‑up approach (points → lines → surfaces) and consists of three modules: a DNN for semantic segmentation, an Integer Programming (IP) module for constraint‑based selection, and a vector standardization step.

4.1 DNN Module

The backbone is a Dilated Residual Network (DRN) that predicts pixel‑wise heatmaps for 21 point classes (13 corner types plus doors, windows, and accessories) under the Manhattan assumption (only horizontal and vertical lines). Focal Loss is employed to improve detection of these small targets, yielding a 1.13 % absolute gain in accuracy and 2.7 % in recall.

4.2 IP Module

Using the DNN heatmaps, candidate primitives (points, walls, openings) are generated. An integer program selects a subset that satisfies a rich set of constraints: one‑hot encoding, connectivity, mutual exclusion, loop consistency, and opening placement. The objective maximizes the number of selected elements while respecting all constraints.

To meet product latency requirements, an entropy‑based pruning strategy is introduced: ambiguous points receive multiple type hypotheses, but entropy measures filter out low‑certainty candidates, reducing variable count and cutting solving time while improving recall by 3.2 % and accuracy by 0.9 %.

5. Discussion

The R2V approach, though more complex than pure end‑to‑end heatmap methods, offers better generalization across diverse floor‑plan styles and handles near‑neighbor point interference by separating point types into distinct channels. However, it still relies on the Manhattan assumption and incurs non‑trivial optimization overhead.

Future directions include exploring graph‑convolutional networks and message‑passing techniques to embed relational reasoning directly into the network, potentially eliminating the heavyweight post‑processing step.

6. Conclusion and Future Work

R2V demonstrates a practical 2D floor‑plan standardization pipeline that bridges deep learning perception and combinatorial optimization, enabling high‑quality vector outputs for downstream AI services. Continued research will aim to relax geometric assumptions and further accelerate the optimization stage.

One more thing: Recruitment

Beike’s Business Intelligence Department is hiring senior/lead algorithm engineers and senior data analysts. Responsibilities include AI model development, data analysis, and building reporting systems. Applicants should have at least a bachelor’s degree, relevant experience, and strong analytical skills. Contact: [email protected].

References

Chen Liu, Jiajun Wu, Pushmeet Kohli, and Yasutaka Furukawa. Raster‑to‑Vector: Revisiting floorplan transformation. ICCV 2017.

Zhiliang Zeng. Deep Floor Plan Recognition Using a Multi‑Task Network with Room‑Boundary‑Guided Attention. ICCV 2019.

Yu, V. Koltun, and T. Funkhouser. Dilated Residual Networks. CVPR 2017.

Tsung‑Yi Lin et al. Focal Loss for Dense Object Detection. ICCV 2017.

Autocad. http://www.autodesk.com/products/autocad/overview.

Sketchup. https://www.sketchup.com.

Jiacheng Chen, Chen Liu, Jiaye Wu, Yasutaka Furukawa. Floor‑SP: Inverse CAD for Floorplans by Sequential Room‑wise Shortest Path. ICCV 2019.