Deep learning-based point cloud completion for MEP components

Point clouds are increasingly leveraged for as-built model reconstruction of facilities. However, point clouds of Mechanical, Electrical, and Plumbing (MEP) systems often experience extensive occlusions, which heavily affect the performance of model reconstruction. To address this challenge, this paper explores deep learning (DL)-based point cloud completion algorithms to complete occluded MEP point clouds. Due to the limited availability of datasets, parametric BIM modeling and occlusion simulation are used to generate synthetic point cloud datasets of MEP components. Based on generated datasets, the effectiveness of five different DL algorithms and five distinct training strategies for point cloud completion are investigated. The results indicate that: (1) The PoinTr model with a pre-training strategy achieved the best Chamfer Distance (CD) and F-score, demonstrating effective completion even with 75 % missing point clouds. 2) Applying the proposed point cloud completion method to three practical tasks further demonstrates the algorithm's applicability.