Scholarlink

<p>Point cloud analysis plays a pivotal role in various real-world applications, including autonomous driving, robotics, and augmented reality. Traditional methods often struggle to balance local feature extraction with global context modeling, particularly when dealing with unstructured and sparse point cloud data. To address this challenge, we introduce PointMSMamba, a novel multi-scale state-space model specifically designed for point cloud analysis. By employing a Trans-Z-order serialization method, our model effectively transforms unstructured point clouds into ordered sequences, preserving spatial proximity. Furthermore, a hierarchical multi-scale architecture and an Edge Graph Module are integrated within the Mamba layer to facilitate the efficient fusion of raw point features and local difference information. Experimental results on benchmark datasets, including ScanObjectNN, ModelNet40, and ShapeNetPart, demonstrate the superiority of PointMSMamba, achieving state-of-the-art performance with only 10.6M parameters. This not only verifies the model’s lightweight advantage but also highlights its efficacy in complex point cloud tasks. The code and pre-trained model have been made publicly available at <a href="https://github.com/1525898173/PointMSMamba">https://github.com/1525898173/PointMSMamba</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

Enhancing point cloud analysis with multi-scale state space modeling and edge graph augmentation

  • 期刊论文

  • 发表日期：

    2025年9月13日

查看全文
  • Weidong Zhu,
  • Qi Sun,
  • Kuifeng Luan,
  • Qidi Xie,
  • Bangyu Peng

摘要

Point cloud analysis plays a pivotal role in various real-world applications, including autonomous driving, robotics, and augmented reality. Traditional methods often struggle to balance local feature extraction with global context modeling, particularly when dealing with unstructured and sparse point cloud data. To address this challenge, we introduce PointMSMamba, a novel multi-scale state-space model specifically designed for point cloud analysis. By employing a Trans-Z-order serialization method, our model effectively transforms unstructured point clouds into ordered sequences, preserving spatial proximity. Furthermore, a hierarchical multi-scale architecture and an Edge Graph Module are integrated within the Mamba layer to facilitate the efficient fusion of raw point features and local difference information. Experimental results on benchmark datasets, including ScanObjectNN, ModelNet40, and ShapeNetPart, demonstrate the superiority of PointMSMamba, achieving state-of-the-art performance with only 10.6M parameters. This not only verifies the model’s lightweight advantage but also highlights its efficacy in complex point cloud tasks. The code and pre-trained model have been made publicly available at https://github.com/1525898173/PointMSMamba.