An Automated Framework for Generating Synthetic Point Clouds from as-Built BIM with Semantic Annotation for Scan-to-BIM
Jong Won, Bing Han, Fernanda Leite
Abstract
Data scarcity is a major constraint which hinders Scan-to-BIM's generalizability in unseen environments. Manual data collection is not only time-consuming and laborious but especially achieving the 3D point clouds is in general very limited due to indoor environment characteristics. In addition, ground-truth information needs to be attached for the effective utilization of the achieved dataset which also requires considerable time and effort. To resolve these issues, this paper presents an automated framework which integrates the process of generating synthetic point clouds and semantic annotation from as-built BIMs. A procedure is demonstrated using commercially available software systems. The viability of the synthetic point clouds is investigated using a deep learning semantic segmentation algorithm by comparing its performance with real-world point clouds. Our proposed framework can potentially provide an opportunity to replace real-world data collection through the transformation of existing as-built BIMs into synthetic 3D point clouds.