Scholarlink

<p>In recent years, advancements in 3D acquisition technologies have enabled the large-scale capture of high-precision point clouds, crucial for applications demanding fine spatial detail. However, the unordered structure, sparsity, and non-uniform density of point clouds present significant challenges for deep feature learning: The irregular structure complicates long-range dependency modeling, while sparsity and non-uniform density lead to computationally expensive and complex local geometric representation. To address these, we propose GD-DGT, a novel architecture integrating geometric-aware dynamic graph convolution with the attention mechanism. Our approach dynamically constructs local graphs to encode spatial relationships and adaptively aggregates neighborhood features to handle geometric anisotropy, while simultaneously employing Transformer modules to capture global dependencies. A multi-scale feature fusion strategy is further incorporated to meet the substantial computational and memory demands of large-scale point cloud processing. Extensive evaluations on ModelNet40, ShapeNetPart, and S3DIS benchmarks demonstrate that GD-DGT delivers notable and consistent gains over established baseline methods. Ablation studies validate the complementary roles of individual components. Our work underscores the effectiveness of combining graph convolution with Transformers for enhanced 3D point cloud feature learning. Code is available at: <a href="https://github.com/AIyc-0101/GD-DGT">https://github.com/Alyc-0101/GD-DGT.</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 3D point cloud learning with geometric-aware dynamic graph convolution and Transformer networks

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Chengjun Wang,
  • Chang Zhao

摘要

In recent years, advancements in 3D acquisition technologies have enabled the large-scale capture of high-precision point clouds, crucial for applications demanding fine spatial detail. However, the unordered structure, sparsity, and non-uniform density of point clouds present significant challenges for deep feature learning: The irregular structure complicates long-range dependency modeling, while sparsity and non-uniform density lead to computationally expensive and complex local geometric representation. To address these, we propose GD-DGT, a novel architecture integrating geometric-aware dynamic graph convolution with the attention mechanism. Our approach dynamically constructs local graphs to encode spatial relationships and adaptively aggregates neighborhood features to handle geometric anisotropy, while simultaneously employing Transformer modules to capture global dependencies. A multi-scale feature fusion strategy is further incorporated to meet the substantial computational and memory demands of large-scale point cloud processing. Extensive evaluations on ModelNet40, ShapeNetPart, and S3DIS benchmarks demonstrate that GD-DGT delivers notable and consistent gains over established baseline methods. Ablation studies validate the complementary roles of individual components. Our work underscores the effectiveness of combining graph convolution with Transformers for enhanced 3D point cloud feature learning. Code is available at: https://github.com/Alyc-0101/GD-DGT.