Instance segmentation of reinforced concrete bridge point clouds with transformers trained exclusively on synthetic data

Bridges in the United States require element-level inspections every 24 months, typically relying on laborious manual assessments. Three-dimensional (3D) point clouds from LiDAR or photogrammetry can facilitate these inspections, but are difficult to leverage without automatically identifying individual structural elements. Existing research focuses on semantic segmentation, which classifies points into broader categories rather than identifying each element instance. A major bottleneck is the difficulty of producing instance-level annotations. To address this gap, the paper proposes and evaluates three synthetic data generation approaches to produce automatically labeled point clouds of bridges with element instance-level annotations. An occlusion technique is introduced to increase realism. The synthetic data is then evaluated for training Mask3D transformer model for instance segmentation of field-collected point clouds, achieving mean Average Precision (mAP) scores of 91.7 % on LiDAR data and 63.8 % on photogrammetry. These results demonstrate the potential to enhance element-level bridge inspections and improve overall infrastructure management.