AI-Assisted 3D Reconstruction and Assembly Diagram Generation for Building Components from Catalog and Smartphone Field Images

How does an AI-generated three-dimensional (3D) model fit into BIM as a diagram to support fieldwork? This research addresses this question by decomposing a device or construction component into parts based on images, assembly diagrams and 3D models. A number of AI application programs and platforms were applied for 2D assembly catalogs and 3D reconstruction of 3DGS models. The AI-assisted diagram interpretation is generated mainly based on field imagery of assembled or disassembled devices or fixtures. The illustrated structure, which is subject to future updates, contributes to the composition usually required for field reference. Field imagery is also AI-reconstructed in 3D to validate the diagram. The result combines the advantages of image-based 3D part decomposition and the regeneration of 2D assembly diagrams. In total, about 29 architectural fixtures and MEP devices were regenerated, creating 190 models of supporting formats and 87 Gemini® or ChatGPT® diagrams. Two of the seven sets of prompt types were revised to emphasize orientation, leading to about 44.4% of all errors in the diagrams. The comparative evaluation of 3DGS models demonstrated standard deviations ranging from 0.165503 mm at an 80% sensitivity level to 3.786150 mm across various 3D applications. The novelty of AI-assisted image-to-BIM lies in reinterpreting a subject from its as-built form in a number of AI-assisted, open, and accessible approaches. The 3DGS model offers important advantages in terms of verification by combining the visual and structural details of an object. Combining diagrams and 3D modeling enables an evolving conversion to a BIM IFC structure with a part-mapping table. Supported by cloud access and smartphone interaction, 3D catalogs and AR present an imagery-to-AR approach for potential field assembly assistance.