Performance boosting of conventional deep learning-based semantic segmentation leveraging unsupervised clustering: Fid-Bau Portal

Performance boosting of conventional deep learning-based semantic segmentation leveraging unsupervised clustering

Ma, Jong Won / Leite, Fernanda

Abstract In Scan-to-BIM, semantically understanding 3D point clouds is an essential process that must precede 3D BIM elements generation. The main idea of this research stemmed from an assumption where recognizing a group of points would be simpler than assigning label per point from the machine's perspective when using deep learning classifiers. To validate our assumption, conventional point-wise classification problem was formulated as segment-wise classification problem leveraging unsupervised clustering. A single parameter executable hierarchical density-based algorithm and PointNet-based algorithms are chosen for segmentation and classification purposes. Using same baseline architecture, our segment-wise classification framework showed performance boosts of 1.12% and 7.66% for S3DIS and 3DFacilities compared to point-wise classification approach. Solving the nature of class imbalance problem and dataset augmentation ability are determined to be the contributing factors for showing superior results. Given semantic segmentation deep learning algorithm, our proposed framework provides an opportunity to improve the performance leveraging unsupervised method.

Highlights • Formulating deep learning-based point-wise classification as segment-wise classification leveraging unsupervised clustering • The proposed segment-wise classification framework showed accuracy increases of 1.12% and 7.66% for S3DIS and 3DFacilities. • Resolving the nature of class imbalance problem by transforming uneven class distribution into a flatter shape • Minimization of the parameter optimization efforts of the model-driven methods for 3D point clouds semantic segmentation