Scholarlink

<p>Three-dimensional (3D) object detection based on dynamic LiDAR point clouds is crucial for industrial defect detection, autonomous driving system, and so on. However, the inherent sparsity and irregularity of point clouds pose significant challenges for detection models. PillarNeXt, a state-of-the-art pillar-based 3D object detection model, excels in scalability and generalization but has limitations in inference speed and capturing fine-grained geometric features for accurate bounding box prediction. To address these issues, we propose <i>R</i>eallocation and <i>F</i>usion Enhanced <i>PillarNeXt</i> (RFPillarNeXt), an improved version of PillarNeXt. RFPillarNeXt optimizes the backbone network by reallocating sparse convolution modules and fuses the BatchNorm2d layers in the detector head with preceding convolution layers. Experiments on the KITTI and nuScenes datasets demonstrate that RFPillarNeXt outperforms existing models, achieving a bbox mAP R40 of 87.969% and a 3D mAP R40 of 79.757% on the KITTI dataset, with a 10.3% increase in inference speed on a local GPU. These results highlight the effectiveness of our proposed improvements in enhancing geometric feature extraction and inference efficiency. Our code is available at <a href="https://github.com/whitewolf666/RFPillarNeXt">https://github.com/whitewolf666/RFPillarNeXt</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Rfpillarnext: geometry-aware 3D object detection for dynamic LiDAR point clouds

  • 期刊论文

  • 发表日期：

    2025年8月18日

查看全文
  • Wenhui Chen,
  • Xiangze Jiang,
  • Zhenyan Ji,
  • Hui Liu,
  • Guiping Zhu,
  • Jose Enrique Armendariz-Inigo,
  • Shen Yin

摘要

Three-dimensional (3D) object detection based on dynamic LiDAR point clouds is crucial for industrial defect detection, autonomous driving system, and so on. However, the inherent sparsity and irregularity of point clouds pose significant challenges for detection models. PillarNeXt, a state-of-the-art pillar-based 3D object detection model, excels in scalability and generalization but has limitations in inference speed and capturing fine-grained geometric features for accurate bounding box prediction. To address these issues, we propose Reallocation and Fusion Enhanced PillarNeXt (RFPillarNeXt), an improved version of PillarNeXt. RFPillarNeXt optimizes the backbone network by reallocating sparse convolution modules and fuses the BatchNorm2d layers in the detector head with preceding convolution layers. Experiments on the KITTI and nuScenes datasets demonstrate that RFPillarNeXt outperforms existing models, achieving a bbox mAP R40 of 87.969% and a 3D mAP R40 of 79.757% on the KITTI dataset, with a 10.3% increase in inference speed on a local GPU. These results highlight the effectiveness of our proposed improvements in enhancing geometric feature extraction and inference efficiency. Our code is available at https://github.com/whitewolf666/RFPillarNeXt.