Kernel Point Non-local Networks for LiDAR Semantic Segmentation

Abstract

LiDAR point cloud semantic segmentation based on convolutional neural networks has become an effective way to understand traffic scenes. Previous works mainly focus on projecting point clouds onto a plane and then use efficient 2D CNN to achieve efficient feature extraction. However, the projection process is accompanied by 3D information loss, challenging to adapt to the complex traffic environment. In this paper, we propose a point-based segmentation network based on three-dimensional convolution, which directly takes the point cloud as input, integrates a variety of distributed kernel point convolutions and introduces an attention mechanism to learn 3d point features efficiently. To evaluate our algorithm, we conducted sufficient experiments on the widely used public dataset SemanticKITTI [1]. The results show that our proposed Kernel Point Non-local module improving the accuracy of KPConv [2] from 58.8% to 61.5%, leading to new state-of-the-art among point-based methods.