SP-Net: A Sparse Convolution and Point-Encoding Enhanced Network for 3D Object Detection in LiDAR Point Clouds

LiDAR 3D object detection for autonomous driving is an important issue. To address this issue, this paper provides a two-stage anchor-based solution. Firstly, voxel feature encoding and sparse convolution networks were employed for proposal generation. Secondly, we performed voxel feature aggregation and hierarchical feature learning in combination with the proposals, multi-level and multiscale voxel features, point encoding and voxel-wise ROI pooling. The aggregated object-wise voxel features were utilized to refine the proposals. Extensive experiments were conducted on three challenging datasets: KITTI, nuScenes and Waymo. Experimental results demonstrated that our SP-Net can effectively achieve multi-category 3D object detection in diverse scenarios offering satisfactory accuracy and speed (28.6 FPS).