As-built Building Information Modeling (BIM) of bridges is still constrained by a fragile transition from reality capture to information, in which segmentation outputs do not consistently yield geometrically reliable, semantically connected Industry Foundation Classes (IFC) models suitable for downstream engineering use. This paper presents a human-in-the-loop Scan-to-BIM workflow for reinforced-concrete beam bridges and shows, through comparison with fully automated and deterministic baselines, that a post-refinement stage remains necessary between segmentation and IFC reconstruction. The workflow combines a dynamic graph convolutional neural network (DGCNN)-based background removal with a strategic, protocol-driven, geometry-guided human-in-the-loop partitioning step that separates the structure into slabs, bearings, piers, and abutments in a canonical aligned frame, after which adaptive geometric routines automatically parameterize each component and assemble a connectivity-aware IFC4 model. The framework is designed to remain effective under repeated bridge-scanning conditions, including varying point densities across slab regions, residual points from in-contact objects, piers partially submerged in the ground at an angle, and bearings that are out of the scanner's line of sight due to settlement or constrained visibility. The method was evaluated on three unseen bridge case studies with varying span counts, point densities, pier geometries, and bearing occlusion conditions. IFC-versus-as-built evaluation demonstrated low-centimeter geometric fidelity, with average mean absolute deviation/root mean square error (MAD/RMSE) values of 0.7/1.2 cm for slabs, 1.4/2.0 cm for piers, 1.9/2.6 cm for bearings, and 1.6/2.2 cm for abutments. The results show that strategically placed human intervention, when constrained by a repeatable protocol and geometric guidance, optimizes accuracy and time more effectively than fully automated semantic pipelines or deterministic parametric modeling alone, and enables robust, documentation-grade, connectivity-aware IFC bridge reconstruction. • This paper balances automation and manual intervention to transform a raw bridge point cloud into BIM. • The framework isolates non-structural elements from structural points with DGCNN. • Unsupervised clustering and computational geometry-based techniques characterize parameters. • Profiles are generated in IFC based on spatial location, true geometry, and element connectivity. • A unified end-to-end pipeline is tested on benchmark datasets of full-scale bridges.
Sakr et al. (Wed,) studied this question.