Automatic Scan-to-BIM—The Impact of Semantic Segmentation Accuracy

Scan-to-BIM is the process of converting point cloud data into a Building Information Model (BIM) that has proven essential for the AEC industry. Scan-to-BIM consists of two fundamental tasks—semantic segmentation and 3D reconstruction. Deep learning has proven useful for semantic segmentation, and its integration into the Scan-to-BIM workflow can benefit the automation of BIM reconstruction. Given the rapid advancement of deep learning algorithms in recent years, it is crucial to analyze how their accuracy impacts reconstruction quality. In this study, we compare the performance of five deep learning models—PointNeXt, PointMetaBase, PointTransformer V1, PointTransformer V3, and Swin3D—and examine their influence on wall reconstruction. We propose a novel yet simple workflow that integrates deep learning and RANSAC for reconstructing walls, a fundamental architectural element. Interestingly, our findings reveal that even when semantic segmentation accuracy is lower, reconstruction accuracy may still be high. Swin3D consistently outperformed the other models in both tasks, while PointNeXt, despite weaker segmentation, demonstrated high reconstruction accuracy. PTV3, with its faster performance, is a viable option, whereas PTV1 and PointMetaBase delivered subpar results. We provide insights into why this occurred based on the architectural differences among the deep learning models evaluated. To ensure reproducibility, our study exclusively utilizes open-source software and Python 3.11 for processing, allowing future researchers to replicate and build upon our workflow.