Point Cloud Completion for MEP Components Using Deep Learning Techniques

Point clouds are increasingly leveraged for as-built model reconstruction of facilities, and these reconstructed models play a key role in applications for digital twins and sustainable building. However, point clouds of Mechanical, Electrical, and Plumbing (MEP) systems often experience extensive occlusions, which heavily affect the performance of model reconstruction. To tackle this problem, this study adopts deep learning (DL)-based point cloud completion algorithms to complete occluded MEP point clouds. To overcome the scarcity of datasets, this study utilizes parametric BIM modeling and occlusion simulations to generate point cloud datasets for MEP components. Based on generated datasets, the effectiveness of PoinTr DL algorithms and five distinct training strategies for point cloud completion are investigated. The results indicate that: 1) The PoinTr model with pre-train strategy achieved the best CD-L2 score of 0.330, demonstrating effective completion even with 75% missing of point clouds. 2) the completion of point clouds for real-world MEP components showed favorable results, underscoring the practical applicability of this approach.