Vision-guided Gaussian Splatting for robust sparse-view bridge reconstruction toward digital twins

High-fidelity 3D reconstruction of bridge structures is crucial for creating accurate bridge digital twins to support structural health monitoring, deformation analysis, and maintenance decision-making. However, in practical field environments, image acquisition is often limited to sparse viewpoints due to safety, accessibility, and cost constraints, posing significant challenges to conventional reconstruction methods. To address this issue, we propose a bridge-oriented vision-guided Gaussian Splatting framework with confidence-driven geometric optimization for robust sparse-view bridge reconstruction. Specifically, the proposed framework integrates the representation capability of a pre-trained vision foundation model with the strengths of existing sparse-view 3D Gaussian Splatting methods, providing reliable geometric guidance under limited-view observations. Based on the extracted geometric priors, we introduce a Confidence-Guided Voxel Filtering strategy to preserve reliable structural regions while suppressing uncertain observations for robust Gaussian initialization. Furthermore, we propose a Bridge Depth Regularization module that incorporates confidence-weighted depth supervision during Gaussian optimization to preserve the geometric consistency of bridge engineering components. Combined with existing optimization constraints, it further improves reconstruction stability and geometric completeness under sparse-view conditions. Experiments on the LLFF benchmark and a real-world UAV bridge dataset demonstrate that the proposed framework consistently improves reconstruction accuracy and geometric fidelity, providing reliable geometric inputs for bridge inspection and subsequent digital-twin workflows.