Reality Capture and Digital Twins

Automated Scan-to-BIM: A Deep Learning-Based Framework for Indoor Environments with Complex Furniture Elements

Mostafa Mahmoud, Zhebin Zhao, Wu Chen, Mahmoud Adham, Yaxin Li

Published 2025

Thumbnail for Automated Scan-to-BIM: A Deep Learning-Based Framework for Indoor Environments with Complex Furniture Elements

Abstract

Extensive 3D parametric datasets, such as Building Information Modeling (BIM) models, are crucial for reducing project costs, supporting planning, and enhancing operational efficiency in building management. However, conventional Scan-to-BIM methods rely heavily on manual or semi-automatic techniques, focusing on space-forming elements such as walls while often neglecting indoor space-occupying furniture. These methods struggle with incomplete point clouds, capturing shapes and orientations, and clustering inaccuracies. This paper presents an innovative and efficient deep learning-based framework to automatically reconstruct 3D models from point clouds. The framework accommodates diverse space-forming layouts and automatically generates parametric 3D BIM models for complex space-occupying elements like tables and chairs within the Revit platform. It also produces non-parametric 3D semantic representations of complete indoor scenes. Evaluation of publicly available and locally acquired datasets shows that the framework achieves over 98 % precision, recall, and F1-score, confirming its accuracy and effectiveness in generating complete 3D models. The reconstructed models preserve key real-world characteristics, including geometric fidelity, numerical attributes, spatial positioning, and various shapes and orientations of furniture. Seamless integration of deep learning and model-driven techniques overcomes the limitations of traditional Scan-to-BIM methods, providing an accurate and efficient solution for complex indoor space reconstruction.

Topics

2D floor plan3D floor planBIMFloor Plan Reconstruction and ParametrizationRevit automationdigital twinsindoor scene reconstructionlayout recoveryobject detectionpoint cloudsscan-to-BIMsemantic segmentation

Cite this paper

@article{9f1138e0,
title = {Automated Scan-to-BIM: A deep learning-based framework for indoor environments with complex furniture elements},
journal = {Journal of Building Engineering},
volume = {106},
pages = {112596},
year = {2025},
issn = {2352-7102},
doi = {https://doi.org/10.1016/j.jobe.2025.112596},
url = {https://www.sciencedirect.com/science/article/pii/S2352710225008332},
author = {Mostafa Mahmoud and Zhebin Zhao and Wu Chen and Mahmoud Adham and Yaxin Li},
keywords = {Building information modeling (BIM), Deep learning, Point clouds, Furniture, 3D models},
abstract = {Extensive 3D parametric datasets, such as Building Information Modeling (BIM) models, are crucial for reducing project costs, supporting planning, and enhancing operational efficiency in building management. However, conventional Scan-to-BIM methods rely heavily on manual or semi-automatic techniques, focusing on space-forming elements such as walls while often neglecting indoor space-occupying furniture. These methods struggle with incomplete point clouds, capturing shapes and orientations, and clustering inaccuracies. This paper presents an innovative and efficient deep learning-based framework to automatically reconstruct 3D models from point clouds. The framework accommodates diverse space-forming layouts and automatically generates parametric 3D BIM models for complex space-occupying elements like tables and chairs within the Revit platform. It also produces non-parametric 3D semantic representations of complete indoor scenes. Evaluation of publicly available and locally acquired datasets shows that the framework achieves over 98% precision, recall, and F1-score, confirming its accuracy and effectiveness in generating complete 3D models. The reconstructed models preserve key real-world characteristics, including geometric fidelity, numerical attributes, spatial positioning, and various shapes and orientations of furniture. Seamless integration of deep learning and model-driven techniques overcomes the limitations of traditional Scan-to-BIM methods, providing an accurate and efficient solution for complex indoor space reconstruction.}
}