A novel uncertainty-aware point cloud approach for geometric quality monitoring in construction

Our research aims to address the following research question: How can we estimate and quantify the uncertainty in the predicted segmentation results to improve the accuracy of geometric quality monitoring? In addressing this research question, we develop a novel uncertainty-aware point cloud approach to evaluate segmentation quality without the access of ground truth labels. The developed approach consists of: (1) mask 3D-based instance segmentation; (2) estimation and quantification of uncertainty; (3) planar structures detection; and (4) geometric information extraction. A commercial office project is selected as a case to examine the effectiveness and feasibility of the proposed approach. The AP@25 for our Mask 3D segmentation was computed and found to be 0.83 (dropout 0), 0.87 (dropout 0.2) and 0.85 (dropout 0.4). The correlation coefficient for our proposed approach between the performance metric and the uncertainty of predicted results has demonstrated a strong relation. The findings presented in this research demonstrated that our developed approach can effectively capture and estimate model uncertainty and improve the segmentation performance.