BLPNet: An End-to-End Model Towards Voxelization Free 3D Object Detection

Abstract

Recent achievements in robotics application and autonomous navigation are accomplished through the fast progress in point cloud 3D object detection models. Many of the current single staged 3D object detection models rely heavily on the voxelization and PointNet based method for feature extraction. Although it provides an efficient way to process point cloud data, its lack of spatial and geometric relationships on both voxel and point level limited the detection accuracy of models. In this paper, we focus on these limitations of the voxelization based object detection pipeline and proposed a single-stage non-voxelization 3D object detection framework. This framework utilizes the bilateral convolution layer, region-based feature clustering and lattice to feature map layer to address the lack of spatial, and geometric relationships in both voxel and point level, which further improves 3D object detection accuracy on the KITTI dataset. Our method achieved 75.63 mAP in moderate difficulty and outperformed many influential object detection models on the KITTI benchmark leaderboard. Contribution-We proposed a voxelization free single-stage model to achieve a desirable performance on the 3D object detection task.