Segmentation of 3D Point Clouds Using Deep Learning: Comparative Study of Modern Point-Based Techniques

Abstract Semantic segmentation of three-dimensional (3D) point clouds plays a vital role in construction management, facility operations, and the broader built environment. With the increasing use of LiDAR and photogrammetry technologies on job sites, the ability to accurately and efficiently classify elements within 3D scans is essential for progress monitoring, as-built modeling, and digital twin development. However, challenges persist due to data imbalance, occlusions, and variability in real-world construction environments. This study conducts a comparative analysis of three state-of-the-art deep learning architectures, PointNet++, ResPointNet++, and point neighbor aggregation with transformer (PointNAT), for point-based semantic segmentation, specifically in construction and industrial contexts. We introduce a novel expert-annotated data set featuring point clouds captured across an educational institution’s building infrastructure, representing real-world construction elements such as pipes, HVAC units, structural frames, and machinery. The data set mimics typical built environments and supports tasks central to facility and construction management. To address issues of class imbalance and improve model performance, we incorporate focal loss and data augmentation strategies. Model evaluations emphasize segmentation accuracy, computational efficiency, and inference speed, key metrics for real-time applications in construction workflows. The results demonstrate that modern point-based models, particularly ResPointNet++, offer high segmentation accuracy and efficiency, making them suitable for deployment in practical in architecture, engineering, construction, and facilities management (AEC/FM) scenarios. By bridging the gap between computer vision methods and construction-focused applications, this research provides actionable insights for applying deep learning to automate 3D scene understanding in construction and facility environments. The findings contribute to ongoing efforts in integrating intelligent automation into reality capture workflows for the built environment.