A Voxel Classification and Deletion Strategy for LiDAR-based 3D Object Detection

LiDAR-based 3D object detection plays an important role in autonomous driving. Among them, voxel-based methods have found wide applications in industrial deployments. This paper proposes a strategy to alleviate the imbalance in the numbers of foreground points and background points with the help of voxel classification (VoxCls) and voxel deletion (VoxDel). It can significantly improve the performance of voxel-based models without changing the network inference structure. First, a binary classification is developed to guide the network in distinguishing foreground voxels from background voxels. By doing so, the features of voxels implicitly incorporate information about both foreground and background. Second, background voxels are filtered based on VoxCls to solve the imbalance issue between foreground and background. Finally, experimental results demonstrate the effectiveness of enhancing voxel-based 3D object detection performance.