Fully automated method for parametric modelling of industrial facility elements

Downstream applications of infrastructure models are highly reliant on accurate recovery of geometrical parameters of elements, especially in industrial facilities. However, current segmentation methods produce either point clusters, meshes or library templates as outputs, ignoring underlying geometry. This paper introduces an automated method for extracting geometrical parameters from output point clusters of instance segmentation methods. The paper proposes a two-stage pipeline consisting of regression of geometric parameters from point clusters, followed by iterative optimisation. This paper demonstrates that a robust objective function for measuring deviation between input and reconstructed models is critical for optimising geometries. The proposed solution is robust to noise and occlusions, and outperforms existing methods on primitive geometries, achieving around 30% reduction in chamfer and parameter errors in occluded scans compared to RANSAC. Crucially, it generalises to arbitrary parametric shapes, enabling modelling of non-primitive geometries. The method directly supports scan-to-BIM workflows by enabling reliable, parameter-consistent geometric modelling.