Parallelization Strategies for Hierarchical Density-Based Clustering Algorithm Using OpenMP for Scan-To-BIM Applications: Fid-Bau Portal

Parallelization Strategies for Hierarchical Density-Based Clustering Algorithm Using OpenMP for Scan-To-BIM Applications

Ma, J. / Leite, F.

Abstract

Clustering is an unsupervised learning method that provides insights by investigating unknown structures in a dataset without exploiting any ground truth target information. For constructing an as-built Building Information Models (BIM) from captured laser-scanned datasets, the segmentation process precedes modeling, which provides a baseline to be traced for obtaining 3D models from point clouds. For the segmentation process, a clustering algorithm can be effectively applied so that it can group the points having similar features without predefined criteria which, in turn, segments can be easily separated from the entire scene. Amongst various types of clustering algorithms, Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) was developed as a density-based and hierarchical clustering algorithm which provides a simplified tree of significant clusters. This algorithm has several distinct advantages over other clustering algorithms: (1) avoids “flat” (i.e. non-hierarchical) labeling of data objects, (2) automatically simplifies the hierarchy into the most significant clusters, and (3) requires a single input parameter (i.e. minimum number of points) for density threshold. However, this algorithm has an overall computation time complexity represented as a quadratic form (i.e., $O (d n^{2})$ \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{O}}\left( {dn^2 } \right)$$\end{document}) which suffers from the computational efficiency issue especially for massive amounts of data such as those found in 3D point clouds. To ease the applicability of HDBSCAN to Scan-to-BIM applications, this research aims to parallelize major time-consuming components of HDBSCAN algorithm. OpenMP interface was adopted for thread parallelization and parallel efficiency was measured by calculating speedup and efficiency from strong and weak scaling results.