HRNet: 3D object detection network for point cloud with hierarchical refinement

Abstract

Recently, 3D object detection from LiDAR point clouds has advanced rapidly. Although the second stage can improve the detection performance significantly, prior works concern little about the essential differences among different stages for the performance enhancement. To address this, this paper proposes a Hierarchical Refinement Network (HRNet) with two novel strategies. Firstly, we build the detection head on multi-scale voxel features to optimize the regression branch progressively with an effective Scale-aware Attentive Propagation (SAP) module. Then, we propose a Dynamic Sample Selection (DSS) module for the recalculation of the IoU during each stage to obtain more balanced positive and negative sample selections. Experiments over benchmark datasets show the effectiveness of our HRNet, particularly for car detection in the sparse point clouds.