SVGA-Net: Sparse Voxel-Graph Attention Network for 3D Object Detection from Point Clouds

Accurate 3D object detection from point clouds has become a crucial component in autonomous driving. However, the volumetric representations and projection methods previous works fail to establish relationships between local sets. In this paper, we propose Sparse Voxel-Graph Attention Network (SVGA-Net), novel end-to-end trainable network which mainly contains voxel-graph module sparse-to-dense regression achieve comparable tasks raw LIDAR data. Specifically, SVGA-Net constructs complete graph within each divided spherical voxel global KNN through all voxels. The graphs serve as attention mechanism enhance extracted features. addition, enhances box estimation accuracy feature maps aggregation at different levels. Experiments on KITTI benchmark Waymo Open dataset demonstrate efficiency of extending representation proposed can decent accuracy.