2026 Volume 7 Issue 3 Pages 376-385
In Japan, the periodic inspection of approximately 177,000 road bridges requires substantial labor, with record creation and documentation tasks — including damage mapping and report preparation —accounting for approximately half of the inspection workload. This study focuses on generating reference 3D models to mitigate this task and compares the conventional SfM-MVS with recent methods, including 3DGS and DA3. The methods were evaluated using both quantitative metrics, namely number of images, computation time, dimensional error, and bending angle, and qualitative metrics, including surface quality, damage visibility, and failure patterns. In the full-scale bridge experiment, DA3 significantly reduced the required number of images to 1/2-1/5 and computation time to 1/30-1/90 compared with SfM-MVS, while limitations remained in dimensional accuracy, shape accuracy, visibility and stability. By incorporating low-cost smartphone-derived IMU and LiDAR data into foundation-model-based methods such as DA3 and MapAnything, the dimensional error was reduced from 10% to 5%, and the bending angle from 5° to 1°. The proposed pipelines also enabled 3D reconstruction under challenging conditions involving moving illumination, reflective surfaces, and low-texture objects, where conventional methods failed. These results suggest the potential of low-cost 3D reconstruction for bridge inspection documentation in environments such as box girders and tunnels.