Overview
This project investigates how BIM-generated synthetic point clouds can reduce the labeling burden for semantic segmentation of civil mechanical, electrical, and plumbing (MEP) systems. Indoor MEP components are dense, occluded, and difficult to annotate manually, which limits deep-learning-based as-built BIM reconstruction. The project develops a BIM-to-point-cloud workflow that extracts MEP components, inherits semantic labels, aligns point density with terrestrial laser scanning data, and augments synthetic samples for model training. Experiments across representative point cloud segmentation networks show that enhanced synthetic data can substantially improve performance under limited real-label budgets. The work supports automated MEP recognition, as-built BIM updating, digital twin maintenance, and facility operation workflows.