Finding the Optimal Convolutional Kernel Size for Semantic Segmentation of Pole-like Objects in Lidar Point Clouds

Abstract. Pole-like objects (PLOs) are important street assets in urban environments, yet current deep learning methods often underperform in their segmentation compared to other objects. The main challenge is determining the right kernel size to effectively understand the unique structure of PLOs with an appropriate receptive field. In this study, we improve the segmentation performance of PLOs by optimizing the kernel size in a KPConv-based network. Our experiments show that kernel size of 9 yields an Intersection over Union (IoU) of 95.02% on the Parkville-3D dataset. We also develop a post-processing approach that transforms semantic segmentation outputs into panoptic segmentation results, enabling accurate detection of individual PLO instances. Furthermore, qualitative tests on an independent, unlabelled point cloud dataset from a different urban area demonstrate that our method consistently achieves accurate segmentation.