Component-aware 3D damage localization on Gaussian splatting for bridge inspection

Abstract

Structural damage assessment in civil infrastructure increasingly relies on visual data, while conventional image-based approaches remain limited by viewpoint inconsistency and the lack of explicit, structured geometric representations for reasoning. To address these challenges, recent advances in 3D scene representations such as 3D Gaussian Splatting enable efficient and high-quality reconstruction, but their application in infrastructure inspection remains limited due to the lack of integrated semantic understanding. This study presents a probabilistic framework that augments 3D Gaussian primitives with component-aware damage fields, allowing semantic information to be embedded without replacing or overwriting the underlying appearance representation. To improve robustness under multi-view inconsistencies, a stochastic multi-view learning strategy is combined with a local cross-attention mechanism to effectively fuse geometric and appearance cues. In addition, a structural gating formulation enforces consistency between component and damage predictions. Experimental results on three real-world bridge scenes demonstrate improved segmentation accuracy and cross-view consistency, providing initial evidence for consistent and interpretable 3D damage representations.