open access

https://hdl.handle.net/2115/100467

Transfer Learning for Infrastructure-Oriented Semantic Segmentation of Urban 3D Point Clouds

英語

Road Surfaces

2026-07-01

Proceedings of the 10th International Symposium on Reliability Engineering and Risk Management (ISRERM 2026)

Accurate semantic segmentation of urban 3D point clouds is a prerequisite for infrastructure condition monitoring and maintenance risk assessment. However, infrastructure-orientated urban scenes present segmentation challenges due to unclear boundaries with geometrically similar classes, high scene variability, and class imbalance in training datasets. Furthermore, training deep learning models such as the Point Transformer framework from scratch requires substantial computational resources and large-scale annotated datasets, which are often unavailable in real-world applications. To address these issues, this study proposes a transfer learning approach that uses pre-trained weights from the nuScenes autonomous driving dataset. Since the nuScenes dataset contains extensive urban driving scenes, the pre-trained model learns general geometric and contextual features from complex urban environments. Fine-tuning these models on target datasets reduces training time and computational costs while improving segmentation accuracy, which makes deep learning techniques more accessible for civil engineering practitioners. The proposed method builds upon the Point Transformer V3 architecture and integrates multi-scale context aggregation with boundary-aware learning to improve segmentation performance in ambiguous regions. Experimental results on a urban 3D point clouds segmentation task demonstrate the potential of transfer learning to bridge the gap between autonomous driving and civil infrastructure applications, achieving higher segmentation accuracy than training from scratch with considerably less training time, particularly for difficult and underrepresented classes such as outdoor ground and indoor ground surface.

会議発表論文

Version of Record