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

Semantic segmentation of point clouds with deep learning (DL) heavily relies on large datasets for training. However, there is a significant scarcity of datasets for Mechanical, Electrical, and Plumbing (MEP) scenes. To address this gap, this study proposes a method namely the ray-based laser scanning and intersection algorithm (RBLSIA) for automatically generating synthetic point clouds for MEP from Building Information Modeling (BIM) models. Based on RBLSIA, this study conducted totally 25 groups of comparative experiments, investigating the semantic segmentation performance on MEP scenes with different training datasets and different generation approaches for synthetic point clouds. The results show that: 1) the mean Intersection over Union (mIoU) with synthetic point clouds produced by the RBLSIA method is on average 3.32% higher than that by the uniform sampling method; 2) increasing the number of synthetic point cloud samples further improved both the OA and mIoU for semantic segmentation, even surpassing the training accuracy achieved with real point clouds.