Deep learning-based point cloud semantic segmentation methods require extensive labeled datasets, which involve a tedious manual labeling process. Although synthetic-based approaches have been developed to reduce labeling cost, they must be augmented with real-world data for practical performance. This study proposes an automated and fast point cloud labeling method for training data augmentation using a Scan-versus-BIM approach. The proposed approach consists of three main phases: member-based point cloud indexing, discrepancies computation, and optimal alignment of overlapping boundary points. Experimental results show mean intersection over union (mIoU) performance improvements of 39.27% and 33.49% over benchmark-based and synthetic-based approaches, respectively. The Augmented-Scan-versus-BIM-based approach, designed to explore the feasibility of the proposed method as a data augmentation solution for synthetic-based approaches, showed mIoU performance increases of 34.33% and 0.84% compared to synthetic-based and proposed approaches. Additionally, the method generates a labeled point cloud for a complex real-world plant site while reducing computational time by 63.10-fold.
Park et al. (Wed,) studied this question.