Automatic Reconstruction of Semantic 3D Models from 2D Floor Plans (2306.01642v1)

Abstract: Digitalization of existing buildings and the creation of 3D BIM models for them has become crucial for many tasks. Of particular importance are floor plans, which contain information about building layouts and are vital for processes such as construction, maintenance or refurbishing. However, this data is not always available in digital form, especially for older buildings constructed before CAD tools were widely available, or lacks semantic information. The digitalization of such information usually requires manual work of an expert that must reconstruct the layouts by hand, which is a cumbersome and error-prone process. In this paper, we present a pipeline for reconstruction of vectorized 3D models from scanned 2D plans, aiming at increasing the efficiency of this process. The method presented achieves state-of-the-art results in the public dataset CubiCasa5k, and shows good generalization to different types of plans. Our vectorization approach is particularly effective, outperforming previous methods.

• The paper introduces an automated pipeline that reconstructs semantic 3D models from 2D floor plans with improved vectorization accuracy on the CubiCasa5k dataset.
• It employs advanced deep learning techniques such as Faster-RCNN and FPN with ResNet to accurately detect symbols and segment walls for CAD-compatible outputs.
• The approach enhances architectural digitalization by converting legacy scanned plans into semantically rich, vectorized 3D models suitable for modern design and refurbishment.

Overview of Automatic Reconstruction of Semantic 3D Models from 2D Floor Plans

The paper introduces a novel approach for the reconstruction of 3D models from 2D floor plans, addressing the need for digitalization in existing buildings, particularly those lacking computer-aided design (CAD) data or semantic information. The approach aims to automate and improve the efficiency of creating vectorized 3D models from scanned 2D plans, and achieves state-of-the-art performance on the CubiCasa5k dataset, with robustness to various plan types.

Introduction

The digitalization of building information, particularly floor plans, is essential for contemporary construction, maintenance, and refurbishment projects. Many older buildings lack digital plans due to their construction prior to the widespread availability of CAD tools. Consequently, manual conversion is required, which is cumbersome and prone to error. The diversity of symbols and annotation systems in floor plans further complicates automation. Previous solutions have relied on traditional methods like morphological operations and Hough transforms, but recent advances in deep learning have shifted focus toward more effective neural network approaches, as evidenced by works like Liu et al. [2017] and Kalervo et al. [2019].

Methodology

The proposed pipeline for 3D model reconstruction is performed in several stages:

1. Preprocessing and Data Augmentation: To handle various image sizes, a sliding window technique is used, complemented by diverse data augmentation strategies like random scaling and rotation to increase model resilience.
2. Symbol Detection: For detecting discrete symbols such as doors and windows, a Faster-RCNN with a ResNet backbone is utilized due to its performance benefits in object detection tasks.
3. Wall Segmentation and Extraction: Employing an FPN with a ResNet backbone, the method uses an affinity-field loss along with binary cross-entropy to enhance wall segmentation. The resulting masks are then vectorized by segmenting walls into polygons using morphological operations and Hough transforms to detect line orientations. This step meticulously addresses the transformation of pixelwise masks into CAD-compatible vector representations.
4. 3D Reconstruction: Leveraging assumptions on the heights of structural elements, a 3D model is compiled using the extracted 2D data, translated into rectangular box primitives that account for architectural features like doors and windows.

Experimental Results

The method demonstrates significant efficacy on the CubiCasa5k dataset. The improved pipeline yields superior results, with an increase in the mean IoU for both segmentation masks and their vectorized outputs when compared to the baseline methods. Specifically, noteworthy improvements are observed in vectorization accuracy, underscoring the strength of the proposed approach in producing semantically rich 3D models from 2D source data.

Implications and Future Work

The authors highlight the practical applicability of this research in digitizing old building plans, thereby facilitating enhanced efficiency in architectural and construction practices. The theoretical implications extend towards advancements in automated semantic understanding from 2D graphical inputs using deep learning. Future work could explore integrating height information from textual plan annotations and extending dataset capabilities to accommodate diverse architectural styles beyond residential plans.

Conclusion

The paper makes a substantial contribution to the field of computational graphics and AI, presenting an automated pipeline that significantly advances the reconstruction of 3D models from 2D floor plans. This work not only achieves state-of-the-art results but also sets a benchmark for future research in automated architectural design and semantic understanding.