3D bridge segmentation using semi-supervised domain adaptation

Understanding the scene is crucial for automated bridge inspection. Traditionally, bridges are measured using 3D sensors that produce large point clouds. Manually interpreting the captured data is time-consuming and error-prone. This paper proposes an unsupervised and semi-supervised domain adaptation approach for 3D bridge segmentation using labeled synthetically generated data and no or limited real-world data. To achieve this, a pipeline was developed for automatically generating artificial scenes of bridges and virtually scanning them with an artificial sensor. This data, along with real-world data, is utilized for the proposed methods. In the unsupervised approach, a deep feature alignment method integrates real-world data into the training procedure. Instead of feature alignment, a semi-supervised method is proposed to guide the training using only a small amount of annotated real-world data. The findings demonstrate that performance can be enhanced in an unsupervised manner. However, performance gains are significantly amplified when 10 % of real-world data is integrated with synthetic data and used in the proposed guided training. The approaches are validated using two distinct deep learning architectures.