Scholarlink

<p>Point cloud upsampling is essential for enhancing the quality and utility of 3D data in applications such as autonomous driving, robotic navigation, and augmented/virtual reality. However, existing methods often neglect the impact of downsampling on upsampling, leading to loss of local structural information and generation of outlier points. In this paper, we propose Optimizing Point Cloud Upsampling through Downsampling Refinement (PU-DR), a point cloud upsampling network based on a generative adversarial network (GAN) that introduces the Multi-DenseCNN and Up-Refinedown-Up modules. The Multi-DenseCNN feature extractor efficiently encodes multi-scale information, capturing local interdependencies between point features. The Up-Refinedown-Up module integrates a Refinedown module, which combines feature-optimized sampling with residual attention, and a dynamic graph convolution upsampling module, significantly enhancing the correction capability of upsampling. Experimental results on both synthetic and real-world scanned data demonstrate that PU-DR outperforms previous approaches, achieving lower Chamfer distance (0.56) and Hausdorff distance (5.89) while preserving local details and reducing outliers. Our approach offers a new perspective on point cloud processing, contributing to the generation of high-quality upsampled point clouds. The source code and datasets are available online at <a href="https://github.com/6Mrzhu/PU-DR">https://github.com/6Mrzhu/PU-DR</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

Optimizing point cloud upsampling through downsampling refinement: a generative adversarial network approach

  • 期刊论文

  • 发表日期：

    2025年12月7日

查看全文
  • Yilong Zhu,
  • Fengjiao Yang,
  • Riming Sun

摘要

Point cloud upsampling is essential for enhancing the quality and utility of 3D data in applications such as autonomous driving, robotic navigation, and augmented/virtual reality. However, existing methods often neglect the impact of downsampling on upsampling, leading to loss of local structural information and generation of outlier points. In this paper, we propose Optimizing Point Cloud Upsampling through Downsampling Refinement (PU-DR), a point cloud upsampling network based on a generative adversarial network (GAN) that introduces the Multi-DenseCNN and Up-Refinedown-Up modules. The Multi-DenseCNN feature extractor efficiently encodes multi-scale information, capturing local interdependencies between point features. The Up-Refinedown-Up module integrates a Refinedown module, which combines feature-optimized sampling with residual attention, and a dynamic graph convolution upsampling module, significantly enhancing the correction capability of upsampling. Experimental results on both synthetic and real-world scanned data demonstrate that PU-DR outperforms previous approaches, achieving lower Chamfer distance (0.56) and Hausdorff distance (5.89) while preserving local details and reducing outliers. Our approach offers a new perspective on point cloud processing, contributing to the generation of high-quality upsampled point clouds. The source code and datasets are available online at https://github.com/6Mrzhu/PU-DR.