💻 computer science

Recognizing and Reconstructing a Multi-Unit Floor Plan

This paper proposes a novel pipeline combining improved MDA-Unet and MACU-Net segmentation architectures with a vectorization-based reconstruction module to automatically synthesize 3D digital twin models from 2D architectural floor plans, achieving state-of-the-art performance on the CubiCasa benchmark.

Original authors: Lukas Kratochvila, Gijs de Jong, Monique Arkesteijn, Simon Bilik, Tomas Zemcik, Karel Horak, Jan S. Rellermeyer

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lukas Kratochvila, Gijs de Jong, Monique Arkesteijn, Simon Bilik, Tomas Zemcik, Karel Horak, Jan S. Rellermeyer

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 a firefighter or an emergency responder rushing into a burning building. You have a flat, 2D paper map of the floor in your hand. It's cluttered, the ink might be faded, and it's hard to instantly visualize where the walls, doors, and stairs are in 3D space while you're running. Now, imagine if you could instantly turn that flat paper map into a floating, 3D hologram of the building in your mind (or on a tablet), showing you exactly where to run and where the exits are.

That is the problem this paper tries to solve.

The Problem: The "Flat Map" vs. The "Real World"

Cities are growing, and emergencies like fires or floods are getting more complex. While emergency teams have 2D floor plans, these drawings often look different depending on who drew them or what software they used. Turning a flat, messy drawing into a clear, 3D model usually requires a human to manually build it, which takes forever. Most existing 3D models only exist for brand-new buildings or are created using expensive, noisy laser scans (like a point cloud), which are too heavy and slow for quick emergency use.

The Solution: The "Magic Translator" Pipeline

The authors built a digital pipeline that acts like a super-smart translator. It takes a messy 2D floor plan (the input) and automatically turns it into a clean 3D model (the output). They split this process into two main stages: Recognition and Reconstruction.

1. Recognition: The "Eagle-Eyed Detective"

First, the system needs to look at the flat drawing and understand what everything is. Is that a wall? A window? A sliding door? A staircase?

To do this, they built two new "detectives" (AI models named CAB1 and CAB2). Think of these detectives as having special glasses:

Asymmetric Convolution: Instead of looking at the image with a standard square lens, these detectives use "L-shaped" or "T-shaped" lenses. This helps them spot long, thin walls or corners much better than standard tools.
The Attention Mechanism: Imagine a detective who knows exactly where to look. This feature tells the AI, "Ignore the background noise; focus on the door hinges and the wall edges." It highlights the important parts of the drawing.
Multi-Scale Skip Connections: This is like having a team of detectives working together. One looks at the big picture (the whole floor layout), while another zooms in on tiny details (a small window). They share their notes instantly so the final picture is perfect.

The AI doesn't just guess; it learns by looking at thousands of examples, trying to get the "score" (how well it identified walls vs. doors) as high as possible.

2. Reconstruction: The "Architect's Sculptor"

Once the AI has identified every wall and door, it has a messy, pixelated mask (like a coloring book where every color is a different object). The second stage is the Sculptor.

Cleaning Up: The sculptor takes the messy pixels and smooths them out, removing little bumps and errors.
Vectorizing: It turns those fuzzy pixels into crisp, mathematical lines and shapes (polygons). It's like tracing over a rough sketch with a perfect ruler.
Building the 3D Model: Finally, it takes those 2D shapes and gives them height. It knows walls go up, doors have frames, and stairs go up in steps. It builds a 3D model that can be viewed in 3D software (like Blender).

The Results: Did It Work?

The authors tested their "Magic Translator" against other top-tier methods using a famous dataset of floor plans called CubiCasa.

The Score: Their new models (CAB1 and CAB2) achieved a score of 0.86 (on a scale where 1.0 is perfect).
The Competition: The other best methods only scored around 0.74.
The Verdict: Their system was significantly better at finding walls, doors, windows, and stairs, even when the floor plans were from different countries or drawn in different styles.

They also tested it on a new, very complex dataset of large buildings (MURF) and showed that it could handle big, messy plans that other systems struggled with.

What's Next? (According to the Paper)

The authors admit their system isn't perfect yet.

Stairs: Currently, the 3D stairs look a bit like blocky steps because the system turns curves into straight lines. It doesn't yet know which way the stairs are facing.
Round Objects: If a building has a round tower, the system turns it into a polygon (a many-sided shape) that looks a bit like a stop sign instead of a perfect circle.
Old Paper Plans: They want to train their system to read old, hand-drawn paper plans that haven't been digitized yet.

The Bottom Line

This paper presents a new, automated way to turn flat, 2D building drawings into 3D models. By using advanced AI "detectives" to recognize the parts and a "sculptor" to build the shape, they created a system that is faster and more accurate than current methods. The goal is to help emergency responders understand buildings better and faster, potentially saving lives during crises. The code for this system is now available for others to use and improve.

Technical Summary: Recognizing and Reconstructing a Multi-Unit Floor Plan

Problem Statement
Digital twins (DTs) hold significant potential for urban management, particularly in emergency planning, escape route design, and rescue interventions. However, the creation of these twins remains largely manual due to a scarcity of existing 3D representations, which are often available only for new buildings. While 2D architectural floor plans are widely accessible from municipal authorities, they present challenges: they vary in rendering (textual and graphical differences between paper and electronic versions), lack clarity compared to 3D spaces, and often exist in raster formats rather than vectors. Furthermore, existing automated approaches frequently assume vector inputs (unsuitable for older buildings) or focus only on small, single-unit apartments rather than multi-unit floor plans. The core challenge addressed is the reliable recognition of critical structural elements (walls, windows, stairs, railings) in 2D raster floor plans to automatically synthesize a 3D representation (termed a 2.5D model as it estimates floor heights) for integration into digital twins.

Methodology
The authors propose a novel pipeline divided into two main stages: Recognition and Reconstruction.

Recognition (Semantic Segmentation):
The recognition phase utilizes a custom Convolutional Neural Network (CNN) based on MDA-UNet, MACU-Net, and U-Net3+ architectures. The proposed models, named CAB1 and CAB2, incorporate several specific enhancements:
- Asymmetric Convolution (AC) Blocks: Replaces standard square convolutions with an aggregation of $k \times k$ , $1 \times k$ , and $k \times 1$ kernels to better capture skeleton features.
- Attention Mechanism: Combines channel and spatial attention modules to weigh feature maps based on their importance, inferring maps from intermediate features.
- Multi-Scale Feature Skip Connections: The architectures utilize skip connections to mitigate disadvantages found in standard U-Net3+ designs. CAB1 uses regular connections, while CAB2 employs fully connected multi-scale skip connections, linking decoder layers to all previous encoder and decoder layers.
- Multi-Task Training Objective: The loss function is a weighted sum of three components:
  - Asymmetric Unified Focal Loss: Handles class imbalance by suppressing background loss and enhancing foreground loss.
  - Affinity Field Loss: Captures semantic structure in labels.
  - Heatmap Regression Loss: Derived from segmentation masks to predict endpoints of openings (doors, windows).
- Input Handling: The models are trained with varying input image sizes to function as multi-scale detectors.
Reconstruction (Vectorization and Visualization):
The reconstruction phase converts the pixel-wise segmentation masks into a 3D model through:
- Post-processing: Refines connected components in the mask until an uncertainty threshold is reached.
- Vectorization: Transforms refined components into polygons by deriving endpoints of minimum-area rotated rectangles. Vertices are merged if they are sufficiently close or collinear.
- Visualization: The resulting polygon set is visualized using Blender. Stairs are visualized as polygons of a fixed, elevated height, though the current method does not account for stair orientation.

Key Contributions

Novel Pipeline: Development of an end-to-end framework for generating 3D building representations from conventional 2D raster floor plans, specifically targeting multi-unit structures.
Architectural Innovations: Proposal of two new segmentation models (CAB1 and CAB2) featuring asymmetric convolutions, attention mechanisms, and improved multi-scale skip connections.
Comprehensive Evaluation: The framework was tested on multiple datasets (CubiCasa, R3D, CVC-FP, and a proprietary MURF dataset) and compared against state-of-the-art (SOTA) techniques, specifically DFPR and CubiCasa5K.
Open Source: The authors have made their code publicly available to support further research in the domain.

Experimental Results
The proposed methods were evaluated using Recall, F1 score, and Intersection over Union (IoU) metrics on the CubiCasa benchmark dataset.

Performance: The CAB1 and CAB2 models achieved an average F1 score of 0.86 and an average IoU of 0.76.
Comparison: These results outperformed the modified SOTA baselines, DFPR $\beta$ (F1: 0.74, IoU: 0.61) and CubiCasa5k $\beta$ (F1: 0.74, IoU: 0.60).
Generalization: The models demonstrated consistent performance across different datasets (CubiCasa, R3D, CVC-FP), with CAB2 showing slightly better results on the CVC-FP dataset.
Specific Class Performance: The models achieved high recall and F1 scores for walls (0.90/0.90), windows (0.91/0.90), and stairs (0.87/0.89).

Significance and Claims
The paper claims that the proposed pipeline successfully bridges the gap between available 2D floor plans and the need for 3D digital twins, offering a scalable and economical alternative to computationally demanding point cloud methods. By automating the extraction of structural elements from diverse and often older 2D drawings, the method supports safer and more efficient emergency services in growing urban environments.

The authors modestly note limitations, including the current inability to visualize stair orientation and artifacts generated when converting circular objects into polygons. They acknowledge that newer SOTA frameworks (e.g., from [18], [23], [24]) could not be directly compared due to unavailable source codes, though their wall recognition recall (0.94) compares favorably to reported results in [24] (0.89). Future work is identified as focusing on multi-floor alignment, collecting datasets for large-scale public and medical buildings, and improving segmentation for older, paper-based plans.