Application-Oriented Evaluation of Superpoint Transformer for Bridge Component Segmentation in Terrestrial LiDAR Point Clouds

Abstract

Automated segmentation of structural components from three-dimensional point clouds is critical for bridge inspection and maintenance, underpinning downstream applications like dimensional measurement and Building Information Modeling (BIM) generation. This study presents an application-oriented evaluation framework using the Superpoint Transformer (SPT). It learns at the superpoint level to segment structural components including decks, girders, parapets, abutments, and Others from laser-scanned three-dimensional point clouds. Existing deep learning approaches like PointNet++ and Point Transformer have difficulty capturing the large-scale structural contexts inherent to bridges, often producing spatially fragmented and structurally inconsistent predictions near component boundaries and geometrically complex regions. The proposed method leverages superpoint-level representation learning, systematic voxel size optimization, and a transfer learning strategy fine-tuning pre-trained parameters on new datasets. This enables efficient adaptation across different bridge configurations without explicit structural constraints. Evaluated on two datasets, the method achieved a mean Intersection over Union (mIoU) of 83.15% and an Overall Accuracy (OA) of 92.91% on Dataset A. Through fine-tuning from the model pre-trained on Dataset A, the best SPT model reached 74.69% mIoU and 86.20% OA on Dataset B. The approach achieved the highest overall mIoU among seven baseline methods and improved mIoU by approximately nine percentage points over the previous bridge-specific state of the art. It produced structurally coherent classifications preserving the spatial logic of bridge components, demonstrating practical utility for downstream civil engineering applications.