Exploring explainability and robustness of point cloud segmentation deep learning model by visualization

Modern deep learning techniques are very suitable for point cloud segmentation of catenary arches of railway systems. The downside of deep learning models is that they have a low explainability. In this paper we explore the explainability of a PointNet++ model that is used for segmenting point clouds of catenary arches. The exploring of explainability is done by creating a pipeline for adapting, segmenting and visualizing the point clouds. By adapting the point clouds the robustness of the model is also tested. From this research it follows that the PointNet++ model mainly relies on the location of the objects in order to segment them. Changing the shape of an object does not have a significant impact on the performance of the model.