Scholarlink

<p>3D small objects detection is a challenging task. Most existing approaches based on self-attention mechanism to capture context features from grid representation, thereby improving the detection performance of small objects. However, it is difficult to obtain representative object features with a single grid representation. To address this problem, we propose a new 3D backbone, named Two-level Grid Context Propagation Network (TGCPN), which integrates two grid representations into a single framework. TGCPN consists of a Contextual Grid Grouping Module (CGGM) and a Grid-Point Dual Key attention (G-PDK). Specifically, we first dynamically partition voxel grids and pillar grids with CGGM to enable subsequent parallel processing. Then, G-PDK is used to capture the contextual information in the grids and the fine-grained details in the point clouds. Finally, feature propagation is performed on voxel grids and pillar grids to integrate features at different levels. Extensive experiments demonstrate that our TGCPN-based method exhibits excellent performance in 3D small object detection. In addition, our single-stage detector built on TGCPN exhibits excellent real-time performance compared with some mainstream 3D backbones. Our implementation is available at <a href="https://github.com/user-vv/TGCPN">https://github.com/user-vv/TGCPN</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

Tgcpn: two-level grid context propagation network for 3D small object detection

  • 期刊论文

  • 发表日期：

    2025年6月18日

查看全文
  • Yan Zhou,
  • Lei Pu,
  • Xuemiao Xu,
  • Chang’an Yi,
  • Xinwei He,
  • Yuexia Zhou,
  • Yewen Xu

摘要

3D small objects detection is a challenging task. Most existing approaches based on self-attention mechanism to capture context features from grid representation, thereby improving the detection performance of small objects. However, it is difficult to obtain representative object features with a single grid representation. To address this problem, we propose a new 3D backbone, named Two-level Grid Context Propagation Network (TGCPN), which integrates two grid representations into a single framework. TGCPN consists of a Contextual Grid Grouping Module (CGGM) and a Grid-Point Dual Key attention (G-PDK). Specifically, we first dynamically partition voxel grids and pillar grids with CGGM to enable subsequent parallel processing. Then, G-PDK is used to capture the contextual information in the grids and the fine-grained details in the point clouds. Finally, feature propagation is performed on voxel grids and pillar grids to integrate features at different levels. Extensive experiments demonstrate that our TGCPN-based method exhibits excellent performance in 3D small object detection. In addition, our single-stage detector built on TGCPN exhibits excellent real-time performance compared with some mainstream 3D backbones. Our implementation is available at https://github.com/user-vv/TGCPN.