Class-Aware 3D Detector From Point Clouds With Partial Knowledge Diffusion and Center-Weighted IoU
Hao Peng, Guofeng Tong
Abstract
This paper focuses on point-based single-stage 3D object detection from point clouds and proposes a novel elegant detector CPC-3Det. Pyramid and confidence-guided backbones are widely used in point-based methods. However, the limitation of neighborhood points and negative sample construction bring obstacles to the discriminative feature learning and cost. Also, Scene-level spatial information loss should be noted. This paper presents the repository-based backbone consisting of a feature repository and partial knowledge to meet the issues. Additionally, explicit class-aware statistics are designed to raise robust features. Moreover, statistics-embedded detection heads through feature modulation and parameter control enhance CPC-3Det performance. Furthermore, The misalignment in IoU optimization caused by center offset is explored in this paper. The paper proposes a center-weighted IoU and designs hybrid losses to drive network parameter optimization. Extensive experiments on both the KITTI and Waymo Open datasets demonstrate the superiority of CPC-3Det over state-of-the-art methods.