Automated material-aware BIM generation using deep learning for comprehensive indoor element reconstruction

Abstract

Automating 3D reconstruction of indoor environments is essential for scene understanding in Building Information Modeling (BIM). This paper addresses the challenge of integrating geometric and material attributes in scan-to-BIM processes. A deep learning-based framework is developed to automatically extract and integrate geometric and material attributes from point clouds, incorporating an enhanced instance segmentation network, a material classification model, and an automated BIM integration workflow for accurate indoor modeling. The proposed framework reconstructs accurate 3D BIM models of space-forming and space-occupying elements while preserving key attributes. Experimental results show significant improvements in instance segmentation accuracy, with reconstructed 3D BIM models achieving over 98 % $\text{correctness}$ and $\text{completeness}$, while the material classification model attains a point-based weighted $F1-\text{score}$ of 0.973 and an object-based accuracy of 94.70 %. These findings advance automated BIM generation, enhancing building planning, asset management, and sustainable design while inspiring further developments in scan-to-BIM automation.