SIENet: Spatial Information Enhancement Network for 3D Object Detection from Point Cloud

• DOI: 10.48550/arxiv.2103.15396
• Publication Date: 2021-01-01
• Authors: Ziyu Li; Yuncong Yao; Zhibin Quan; Wankou Yang; Jin Xie
• Keywords: Margin (machine learning); Benchmark (surveying); Representation

LiDAR-based 3D object detection pushes forward an immense influence on autonomous vehicles. Due to the limitation of intrinsic properties LiDAR, fewer points are collected at objects farther away from sensor. This imbalanced density point clouds degrades accuracy but is generally neglected by previous works. To address challenge, we propose a novel two-stage framework, named SIENet. Specifically, design Spatial Information Enhancement (SIE) module predict spatial shapes foreground within proposals, and extract structure information learn representative features for further box refinement. The predicted complete dense sets, thus extracted contains more semantic representation. Besides, Hybrid-Paradigm Region Proposal Network (HP-RPN) which includes multiple branches discriminate generate accurate proposals SIE module. Extensive experiments KITTI benchmark show that our elaborately designed SIENet outperforms state-of-the-art methods large margin.