Scholarlink

<p>Point cloud normal estimation is a fundamental task in 3D computer graphics, essential for downstream applications such as surface reconstruction and semantic segmentation. While recent advances in deep learning have significantly improved normal estimation accuracy, existing methods often struggle with capturing fine-grained geometric details. In this study, we propose a novel encoder that integrates a local gradient attention module and positional encoding to better capture subtle geometric variations. By introducing the gradient attention module, we effectively capture fine-grained information along the z-axis, while positional encoding using sine and cosine functions further amplifies these variations. Extensive experiments on both synthetic and real-world datasets demonstrate that our approach outperforms state-of-the-art methods, achieving up to a 2.53% improvement in accuracy on PCPNet dataset. Our work not only advances normal estimation but also demonstrates its potential for surface reconstruction tasks. The code is available at <a href="https://github.com/ABc90/gam-net-normal-main">https://github.com/ABc90/gam-net-normal-main</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

Enhanced normal estimation of point clouds via fine-grained geometric information learning

  • 期刊论文

  • 发表日期：

    2025年2月22日

查看全文
  • Wei Jin,
  • Jun Zhou,
  • Mingjie Wang,
  • Nannan Li,
  • Weixiao Wang,
  • Xiuping Liu

摘要

Point cloud normal estimation is a fundamental task in 3D computer graphics, essential for downstream applications such as surface reconstruction and semantic segmentation. While recent advances in deep learning have significantly improved normal estimation accuracy, existing methods often struggle with capturing fine-grained geometric details. In this study, we propose a novel encoder that integrates a local gradient attention module and positional encoding to better capture subtle geometric variations. By introducing the gradient attention module, we effectively capture fine-grained information along the z-axis, while positional encoding using sine and cosine functions further amplifies these variations. Extensive experiments on both synthetic and real-world datasets demonstrate that our approach outperforms state-of-the-art methods, achieving up to a 2.53% improvement in accuracy on PCPNet dataset. Our work not only advances normal estimation but also demonstrates its potential for surface reconstruction tasks. The code is available at https://github.com/ABc90/gam-net-normal-main.