Data Preparation Impact on Semantic Segmentation of 3D Mobile LiDAR Point Clouds Using Deep Neural Networks

Currently, 3D point clouds are being used widely due to their reliability in presenting 3D objects and accurately localizing them. However, raw point clouds are unstructured and do not contain semantic information about the objects. Recently, dedicated deep neural networks have been proposed for the semantic segmentation of 3D point clouds. The focus has been put on the architecture of the network, while the performance of some networks, such as Kernel Point Convolution (KPConv), shows that the way data are presented at the input of the network is also important. Few prior works have studied the impact of using data preparation on the performance of deep neural networks. Therefore, our goal was to address this issue. We propose two novel data preparation methods that are compatible with typical density variations in outdoor 3D LiDAR point clouds. We also investigated two already existing data preparation methods to show their impact on deep neural networks. We compared the four methods with a baseline method based on point cloud partitioning in PointNet++. We experimented with two deep neural networks: PointNet++ and KPConv. The results showed that using any of the proposed data preparation methods improved the performance of both networks by a tangible margin compared to the baseline. The two proposed novel data preparation methods achieved the best results among the investigated methods for both networks. We noticed that, for datasets containing many classes with widely varying sizes, the KNN-based data preparation offered superior performance compared to the Fixed Radius (FR) method. Moreover, this research allowed identifying guidelines to select meaningful downsampling and partitioning of large-scale outdoor 3D LiDAR point clouds at the input of deep neural networks.