Synthetic-to-Real Domain Adaptation with Virtual Laser Scanning and Self-Training–Based Category-Aware Cuboid Mixing for Semantic Segmentation of Bridge Point Clouds
Year: 2025
Type: article
Abstract: A scarcity of real-world point clouds poses a considerable challenge in training a bridge semantic segmentation model. Although virtual point cloud synthetization has emerged as a promising solution, the domain gap between synthetic and real-world data has been a major roadblock. We therefore present a synthetic-to-real domain adaptation method integrated with virtual laser scanning (VLS) and self-training-based category-aware cuboid mixing (ST-CACM). We evaluate the method’s effectiveness for bridge semantic segmentation, comparing to the model trained only with real-world point clouds. Results show that our method can achieve 95.34% of an overall accuracy, 94.31% of a mean class accuracy, and 89.83% of a mean intersection over union. These metrics demonstrate that our method is comparable to the real-world dataset-trained model, while reducing the quantity of real-world training data. Also, the two core components, VLS and ST-CACM, are shown to be effective at alleviating the domain gap between synthetic and real-world data.
Authors: Xiaofei Yang, Yuguang Fu, Jinwoo Kim
Institutions: Nanyang Technological University, Hanyang University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4416569279