Scholarlink

<p>With the rise of 3D vision, the learning of point cloud data has received increasing attention. Currently, many studies have achieved outstanding performance in tasks such as point cloud classification. However, these methods usually require large amounts of labeled data to train the model and have poor generalization performance for new categories never seen before. To alleviate this problem, this paper proposes a dual graph network for few-shot 3D point cloud classification (DGN), which consists of point and distribution graphs. DGN constructs the initial point graph using the embedding features of the support and query samples, and then establishes the instance-level and distribution-level relationships of each sample with other samples in a 1-vs-N manner. DGN processes these two relationships independently to alternately update the point and distribution graphs, and finally predict the class of query examples. Moreover, this paper also designs a geometric extension module to enhance the discrete point cloud by utilizing the internal geometric relationship of the point cloud, prompting the network to learn more representative feature representations in high-level space. Extensive experiments on several benchmark datasets demonstrate the superiority of DGN.</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

Dual graph network for few-shot 3D point cloud classification

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Yang Li,
  • Wei Xu,
  • Xiaoming Gao,
  • Haizhou Tan,
  • Hui Xue,
  • Hengxin Feng,
  • Baodi Liu

摘要

With the rise of 3D vision, the learning of point cloud data has received increasing attention. Currently, many studies have achieved outstanding performance in tasks such as point cloud classification. However, these methods usually require large amounts of labeled data to train the model and have poor generalization performance for new categories never seen before. To alleviate this problem, this paper proposes a dual graph network for few-shot 3D point cloud classification (DGN), which consists of point and distribution graphs. DGN constructs the initial point graph using the embedding features of the support and query samples, and then establishes the instance-level and distribution-level relationships of each sample with other samples in a 1-vs-N manner. DGN processes these two relationships independently to alternately update the point and distribution graphs, and finally predict the class of query examples. Moreover, this paper also designs a geometric extension module to enhance the discrete point cloud by utilizing the internal geometric relationship of the point cloud, prompting the network to learn more representative feature representations in high-level space. Extensive experiments on several benchmark datasets demonstrate the superiority of DGN.