Scholarlink

<p>To address the challenges of pose variation sensitivity and limited detail recovery in human body point cloud completion, we propose PoseNorm-PCN, a pose normalization completion network designed for human body data. The method employs a three-stage modeling strategy. First, the pose normalization network transforms input point clouds with arbitrary poses into a canonical space. Second, the geometric offset inference network predicts <i>z</i>-axis displacements from the deformed front viewpoints to recover the back-side geometry. Finally, the recovered back point cloud in the canonical space is mapped back to the original pose through local coordinate transformations, yielding a complete human body point cloud consistent with the input pose. Extensive experiments on multiple real-world human scan datasets demonstrate that PoseNorm-PCN outperforms state-of-the-art completion methods, including PCN, PoinTr, and AdaPoinTr, achieving the lowest Chamfer Distance and Earth Mover’s Distance. The model exhibits strong robustness and generalization under complex poses and diverse clothing conditions, making it highly suitable for three-dimensional human body reconstruction, virtual try-on, and human–computer interaction applications. The related code and resources have been released at: <a href="https://github.com/BestLiMing/PoseNorm_PCN.git">https://github.com/BestLiMing/PoseNorm_PCN.git</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

PoseNorm-PCN: pose-normalized human point cloud completion from a single front view

  • 期刊论文

  • 发表日期：

    2025年12月12日

查看全文
  • Ming Li,
  • Guiqin Li,
  • Xihang Li

摘要

To address the challenges of pose variation sensitivity and limited detail recovery in human body point cloud completion, we propose PoseNorm-PCN, a pose normalization completion network designed for human body data. The method employs a three-stage modeling strategy. First, the pose normalization network transforms input point clouds with arbitrary poses into a canonical space. Second, the geometric offset inference network predicts z-axis displacements from the deformed front viewpoints to recover the back-side geometry. Finally, the recovered back point cloud in the canonical space is mapped back to the original pose through local coordinate transformations, yielding a complete human body point cloud consistent with the input pose. Extensive experiments on multiple real-world human scan datasets demonstrate that PoseNorm-PCN outperforms state-of-the-art completion methods, including PCN, PoinTr, and AdaPoinTr, achieving the lowest Chamfer Distance and Earth Mover’s Distance. The model exhibits strong robustness and generalization under complex poses and diverse clothing conditions, making it highly suitable for three-dimensional human body reconstruction, virtual try-on, and human–computer interaction applications. The related code and resources have been released at: https://github.com/BestLiMing/PoseNorm_PCN.git.