Synthetic-To-Real Domain Adaptation with Virtual Laser Scanning and Self-Training-Based Category-Aware Cuboid Mixing for Semantic Segmentation of Bridge Point Clouds

초록

A scarcity of real-world point clouds poses a considerable challenge in training a bridge semantic segmentation model. Although virtual point cloud synthetization offers a promising alternative, the persistent domain gap between synthetic and real-world data remains a critical obstacle. To address this, we present a synthetic-To-real domain adaptation method that integrates virtual laser scanning (VLS) and self-Training-based category-Aware cuboid mixing (ST-CACM). Our experimental evaluation demonstrates the method's effectiveness in bridge semantic segmentation through comparison with models trained on real-world point clouds and traditional synthetic point clouds. The proposed approach achieves an overall accuracy of 95.34%, a mean class accuracy of 94.31%, and a mean intersection over union of 89.83%, demonstrating performance comparable to that of the baseline models while significantly reducing the dependency on real-world training data. Notably, both core components, VLS and ST-CACM, effectively mitigated the domain gap between synthetic and real-world data, establishing a robust framework for synthetic-To-real domain adaptation in bridge segmentation tasks. These findings will advance the reconstruction of digital twins and the efficient operations and maintenance of in-service bridges.

키워드