LessNet: Lightweight and efficient semantic segmentation for large-scale point clouds

With a wide range of applications in autonomous driving and robotics, semantic segmentation for large-scale outdoor point clouds is a critical and challenging issue. Due to the large number and irregular arrangement of point clouds, it is difficult to balance the efficiency and effectiveness. In this paper, we propose LessNet, a lightweight and efficient voxel-based method for LiDAR-only semantic segmentation, taking advantage of cylindrical partition and intra-voxel feature fusion. Specifically, we use a cylindrical partition method to distribute the outdoor point clouds more evenly in voxels. To better encode the voxel features, we adopt an intra-voxel aggregation method without querying neighbours. The voxel features are further input into a lightweight and effective 3D U-net to aggregate local features and dilate the receptive field. Extensive experiments prove the satisfied semantic segmentation performance and the improvement of each component in our proposed framework. Our method is capable of processing more than one million point clouds at a time while retaining low latency and few parameters. Moreover, our method achieves comparable performance with state-of-the-art approaches and outperforms all projection-based methods on the SemanticKITTI benchmark.