Scholarlink

<p>As a vital component of cultural heritage, accurate classification of Terracotta Warrior fragments is essential for the protection and transmission of cultural relics. Conventional manual feature-design approaches, however, are insufficient to tackle the complexity and diversity of fragments. To overcome this, we propose a point cloud classification method based on a residual Mamba architecture, PointRM, specifically designed to improve the accuracy of fragment classification. To capture rich local representations, the method incorporates a geometry-guided feature aggregation module that effectively integrates geometric coordinates and point-wise feature information within local regions. Meanwhile, a feature extraction module constructed from residual Mamba blocks facilitates the acquisition of higher-level global representations, strengthening the expressive capacity for irregular point cloud fragments. Experimental results on multiple benchmarks demonstrate that PointRM achieves competitive performance, particularly in fragments. This work is expected to open new avenues for the virtual restoration and digital research of cultural relics.</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

A residual mamba point cloud classification framework for 3D Terracotta Warrior fragments

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月14日

查看全文
  • Yajing Li,
  • Mingquan Zhou,
  • Pengbo Zhou,
  • Yong Wang,
  • Yinhan Liao,
  • Li An

摘要

As a vital component of cultural heritage, accurate classification of Terracotta Warrior fragments is essential for the protection and transmission of cultural relics. Conventional manual feature-design approaches, however, are insufficient to tackle the complexity and diversity of fragments. To overcome this, we propose a point cloud classification method based on a residual Mamba architecture, PointRM, specifically designed to improve the accuracy of fragment classification. To capture rich local representations, the method incorporates a geometry-guided feature aggregation module that effectively integrates geometric coordinates and point-wise feature information within local regions. Meanwhile, a feature extraction module constructed from residual Mamba blocks facilitates the acquisition of higher-level global representations, strengthening the expressive capacity for irregular point cloud fragments. Experimental results on multiple benchmarks demonstrate that PointRM achieves competitive performance, particularly in fragments. This work is expected to open new avenues for the virtual restoration and digital research of cultural relics.