Advancing MEP Semantic Segmentation with Deep Learning and BIM-derived Synthetic Point Clouds

Abstract: This paper proposes the Ray-Based Laser Scanning and Intersection Algorithm (RBLSIA) to generate synthetic point clouds for Mechanical, Electrical, and Plumbing (MEP) systems using BIM models, addressing the lack of MEP datasets for deep learning-based semantic segmentation. Twenty comparative experiments were conducted to assess the performance across different training datasets, synthetic point cloud generation methods.The results show that RBLSIA-generated synthetic point clouds outperform those from uniform sampling by 3.32% in mean Intersection over Union (mIoU). Additionally, increasing the volume of synthetic samples improves overall accuracy (OA) and mIoU, surpassing the performance of models trained with real point clouds.