2026 Volume 7 Issue 3 Pages 1005-1015
Bridge inspection and assessment require the ability to understand the spatial relationships among structural components and damage based on visual information. Although Vision-Language Models (VLMs) are promising for bridge maintenance applications, their ability to understand bridge-specific positional expressions has not been sufficiently investigated. This study constructed a prototype Visual Question Answering (VQA) benchmark for evaluating spatial understanding in bridge scenes. Bridge CAD models were developed to generate multi-view images, and question-answer pairs were automatically created from component attributes and geometric information. The constructed VQA benchmark was used to evaluate Qwen3.5-9B and GPT-5.6 Luna with respect to different task types and the number of input views. The results showed that the models could identify visible bridge components and damage and interpret positional and directional concepts when sufficient visual cues were available but sometimes produced definitive answers even for questions that were difficult to answer from the image s alone. The addition of slightly perturbed viewpoints reduced the truncation rate for Qwen3.5 and improved the answer accuracy for GPT-5.6. This benchmark captured differences in capabilities between the models.