open access

https://hdl.handle.net/2115/100466

An Empirical Study of Bridge Point Cloud Semantic Segmentation Across Structural Spans

英語

Structural component recognition

2026-07-01

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

Semantic segmentation of bridge point clouds is important for automated inspection and digital asset management. Although Transformer-based methods perform well on benchmark datasets, the effect of bridge span scale on segmentation has not been systematically studied. This paper presents an empirical study of point cloud semantic segmentation across small-, medium-, and large-span bridges under a unified protocol. Using Point Transformer V3 (PTv3) with consistent preprocessing, training, and evaluation settings, we analyze segmentation performance for five structural components: background, piers/abutments, girders, decks, and superstructure accessories. Results show that performance varies across structural components and that the effect of span scale is non-uniform. Deck-related elements remain relatively stable across span scales, whereas piers and abutments show greater variability in large-span scenes. We also find that metrics computed on subsampled point clouds do not always match qualitative visual results, indicating that evaluation is influenced by sampling strategy and evaluation perspective. These findings show the importance of considering bridge span scale and structural characteristics when applying point cloud semantic segmentation to bridge inspection, and motivate further study of scale-aware modeling and evaluation for large infrastructure point clouds.

会議発表論文

Version of Record