Voxel completion and 3D asymmetrical convolution networks for Lidar semantic segmentation

Abstract

The point cloud data collected by LiDAR is large in scale and contains rich spatial structure detail information, through the collection and labeling of LiDAR data, the automatic driving system can obtain detailed information about the environment around the vehicle. Due to lack of sufficient laser points, some methods transform the point cloud to dense representations such as multi-view or voxelized grids for processing, ignoring the information loss problem caused by the LiDAR imaging characteristics as well as the point cloud transformations, which leads to a degradation of the segmentation performance. In this work, We investigate a 3D semantic segmentation scheme with only LiDAR inputs, called voxel completion and 3D asymmetric convolution network. We propose a voxel completion sub-network to improve the feature extraction capability of the network by enlarging the receptive field and using multi-scale feature extraction to reduce the empty units in the voxels and obtain more complete voxel features. In addition, due to the presence of a large number of cubic objects in the autopilot scenario, to better match the autopilot scenario, we propose a 3D asymmetric convolution network that includes three components: a 3D residual block, an asymmetric convolution block, and a context module. These components are combined together to explore 3D geometric patterns, which can maintain their intrinsic properties and improve the performance of the network. Extensive experiments on the SemanticKITTI and nuScenes benchmark datasets demonstrate the superiority of the approach. For example, on the nuScenes validation set, our method outperforms the state-of-the-art method by 0.3% in mIoU.