Scholarlink

<p>Point cloud classification is crucial in the processing and analysis of three-dimensional data. However, the irregularities and disorders inherent in point cloud data pose significant challenges when representing point cloud information. Most existing models utilize farthest point sampling to extract subsets of point clouds. However, this method tends to select spatially distant points, which may result in insufficient sampling in regions with high-density variation, leading to the loss of crucial geometric information. To address this issue, we design a hybrid model called MDCSNet, comprising two branches: a farthest point sampling branch and a criticality point sampling branch. The former enhances the perceptual capability of the model by dynamically integrating multi-scale spatial information based on a hierarchical structure. At the same time, the latter employs an optimized criticality point strategy to refine the extraction of crucial features in three-dimensional space. We conducted experiments on various datasets, achieving an overall accuracy of 93.4% and a mean accuracy of 91.8% on the ModelNet40 dataset, and an overall accuracy of 81.82% on the ScanObjectNN dataset, thereby demonstrating the effectiveness of our model.</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

Mdcsnet: multi-scale dynamic spatial information fusion with criticality sampling for point cloud classification

  • 期刊论文

  • 发表日期：

    2025年1月11日

查看全文
  • Pusen Xia,
  • Shengwei Tian,
  • Long Yu,
  • Xin Fan,
  • Zhezhe Zhu,
  • Hualong Dong,
  • Na Qu,
  • Tong Liu,
  • Xiao Yuan

摘要

Point cloud classification is crucial in the processing and analysis of three-dimensional data. However, the irregularities and disorders inherent in point cloud data pose significant challenges when representing point cloud information. Most existing models utilize farthest point sampling to extract subsets of point clouds. However, this method tends to select spatially distant points, which may result in insufficient sampling in regions with high-density variation, leading to the loss of crucial geometric information. To address this issue, we design a hybrid model called MDCSNet, comprising two branches: a farthest point sampling branch and a criticality point sampling branch. The former enhances the perceptual capability of the model by dynamically integrating multi-scale spatial information based on a hierarchical structure. At the same time, the latter employs an optimized criticality point strategy to refine the extraction of crucial features in three-dimensional space. We conducted experiments on various datasets, achieving an overall accuracy of 93.4% and a mean accuracy of 91.8% on the ModelNet40 dataset, and an overall accuracy of 81.82% on the ScanObjectNN dataset, thereby demonstrating the effectiveness of our model.