Encoding 3D Point Contexts for Self-Supervised Spall Classification Using 3D Bridge Point Clouds

Geometric features, such as normal and curvature, have been prominently used for point cloud-based unsupervised spall classification. In addition, some researchers use hand-crafted features (e.g., out-of-plane distance, eccentricity, principal curvatures in 2D slices). These features perform well in low noise settings; however, the performance tapers down significantly when the quality of point clouds is affected by factors such as higher noise and inconsistent point-to-point spacing. Instead of relying purely on handcrafted features, the research presented in this paper investigates the potential for combining domain knowledge with deep learning to automatically learn better quality defect-sensitive features for point cloud-based spall classification. Specifically, generic three dimensional (3D) shape and 3D neighborhood features have been encoded as inputs to a deep autoencoder network for self-supervised spall classification from point clouds. Overall, this approach only resulted in marginal improvement over classification results from the current state-of-the-art unsupervised approach that uses handcrafted features. However, significant improvement in the results were observed in datasets that had higher noise levels. Given that noise is pervasive in datasets from outdoor settings like civil infrastructure, this added robustness to noise improves the reliability of point cloud-based condition assessment for concrete bridges.