Automatic 3D Model Construction from Bridge Drawings via Recursive Refinement of Generated Outputs with a Vision-Language Model

Abstract

Three-dimensional (3D) models are expected to improve the efficiency of bridge maintenance; however, many existing bridges still lack such models. This study proposes a method for constructing and automati-cally refining a 3D bridge model from drawings by leveraging a vision language model (VLM) to interpret the contents of multiple drawings. In the proposed method, the VLM extracts and organizes structural in-formation from several drawings and generates a script for 3D model construction. A self-verification and correction loop is then performed by detecting interferences and floating members in the constructed model and by evaluating consistency using comparison images between the original drawings and the generated 3D model. A system implementing the proposed method was developed and applied to actual bridge draw-ings, confirming that a 3D model reflecting the major dimensions can be constructed automatically.