Scholarlink

<p>Existing point cloud registration methods suffer from limitations in learning rotation-invariant features, such as insufficient hierarchical modeling, parameter redundancy, and limited robustness in low-overlap scenarios, which constrain their efficiency in high-performance computing (HPC) applications. To overcome these challenges, we propose HEFPA, a framework that integrates hyperbolic geometry with rotation-aware feature learning. In the feature extraction stage, we introduce a hyperbolic space mapping module that embeds rotation-invariant features into a constant negative curvature space via the Poincaré ball model, thereby explicitly capturing the hierarchical structure of point clouds. Additionally, we design a lightweight module, termed Fang-Layer, which, when combined with farthest point sampling (FPS), optimizes the superpoint distribution and reduces the computational complexity of matching by 37%. Extensive experiments on the 3DMatch, 3DLoMatch, and KITTI datasets demonstrate that HEFPA achieves a Registration Recall of 93.7%/78.5% with a model size of only 6.45 MB, outperforming PEAL by 0.5%/0.7%, while maintaining an inference speed of 0.16 s per frame. Thanks to its lightweight design, HEFPA exhibits superior scalability and energy efficiency in HPC environments, offering a promising solution for large-scale, low-overlap point cloud registration.</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

HEFPA: hyperbolic embedding and fast position-aware network for point cloud registration

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Yang Zhou,
  • Wenlin Huang,
  • Shuting Cai,
  • Jing Guo

摘要

Existing point cloud registration methods suffer from limitations in learning rotation-invariant features, such as insufficient hierarchical modeling, parameter redundancy, and limited robustness in low-overlap scenarios, which constrain their efficiency in high-performance computing (HPC) applications. To overcome these challenges, we propose HEFPA, a framework that integrates hyperbolic geometry with rotation-aware feature learning. In the feature extraction stage, we introduce a hyperbolic space mapping module that embeds rotation-invariant features into a constant negative curvature space via the Poincaré ball model, thereby explicitly capturing the hierarchical structure of point clouds. Additionally, we design a lightweight module, termed Fang-Layer, which, when combined with farthest point sampling (FPS), optimizes the superpoint distribution and reduces the computational complexity of matching by 37%. Extensive experiments on the 3DMatch, 3DLoMatch, and KITTI datasets demonstrate that HEFPA achieves a Registration Recall of 93.7%/78.5% with a model size of only 6.45 MB, outperforming PEAL by 0.5%/0.7%, while maintaining an inference speed of 0.16 s per frame. Thanks to its lightweight design, HEFPA exhibits superior scalability and energy efficiency in HPC environments, offering a promising solution for large-scale, low-overlap point cloud registration.