GVnet: Gaussian model with voxel-based 3D detection network for autonomous driving

This paper proposed a two-stage Voxel-based 3D Object detector which named GVnet. Voxel-based method mainly relies on sampling and Grouping point in voxel and the feature map generated by subsequent 3D CNN to control the quality of detection. Moreover, traditional voxel feature encoder (VFE) methods cannot adjust the quality of feature map through reasonable sampling. Therefore, the method we propose is an improvement to the existing VFE. The specific operations are: First calculate the corresponding Gaussian distribution of the original point cloud data, and then sampling any number of points by controlling the confidence value to improve the performance of voxel encoder and further improve the quality of the feature map output by the 3D CNN. In addition, a voxel ROI pooling method is proposed in stage two. In ROI Pooling, the receptive field in the original space and the corresponding raw point are obtained through the mapping relationship between feature and ROI, then change the raw point to adjust the receptive field to improve the performance of classification and regression. Finally, the experimental results on the KITTI, nuScenes and Waymo dataset show that the performance of GVnet under most of the evaluation indexes is better than the current detection methods, at the cost of only a small amount of inference time.