Enhancing explainability of deep learning models for point cloud analysis: a focus on semantic segmentation

Semantic segmentation of point clouds plays a critical role in various applications, such as urban planning, infrastructure management, environmental analyses and autonomous navigation. Understanding the behaviour of deep neural networks (DNNs) in analysing point cloud data is essential for improving segmentation accuracy and developing effective network architectures and acquisition strategies. In this paper, we investigate the traits of some state-of-the-art neural networks using indoor and urban outdoor point cloud datasets. We compare PointNet, DGCNN, and BAAF-Net on specifically selected datasets, including synthetic and real-world environments. The chosen datasets are S3DIS, SynthCity, Semantic3D, and KITTI. We analyse the impact of different factors such as dataset type (synthetic vs. real), scene type (indoor vs. outdoor), and acquisition system (static vs. mobile sensors). Through detailed analyses and comparisons, we provide insights into the strengths and limitations not only of different network architectures in handling urban point clouds but also of their data structure. This study contributes to going beyond the mere and unconditional use of AI algorithms, trying to explain DNNs behaviour in point cloud analysis and paving the way for future research to enhance segmentation accuracy and develop possible guidelines both for network design and data acquisition in the geomatics field.