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

Bin Lu, Yang Sun, Zhenyu Yang, Ran Song, Haiyan Jiang, Yonghuai Liu

Abstract: Highlights•The essential differences among different detection stages are examined.•Optimize the regression branch through the proposed Hierarchical Refinement Network with multi-scale voxel features gradually.•Adaptively select samples by using the statistical characteristics of IoUs in each stage.•Our HRNet achieves superior performance and works well for car detection at the easy difficulty level on the KITTI leaderboard.