Scholarlink

<p>Rapid advancements in 3D acquisition technologies have significantly advanced the widespread application of point clouds in robotics and autonomous systems. However, the inherent unordered and unstructured nature of point clouds poses significant challenges for neural network design. Most methods neglect the potential of the features represented by the original positional information of the points. To capture comprehensive feature information to enhance local feature aggregation, we propose the Local Feature Aggregation Module (LFAM), which incorporates the original positional information and integrates it with edge features. To further enhance the classification accuracy and practical applicability of the model, we propose a novel point feature enhancement architecture, PointFEA, which initially employs LFAM and subsequently incorporates a channel-cluster-based attention mechanism. Our method optimally extracts local features while effectively integrating global contextual information. Extensive experimental evaluations demonstrate that PointFEA excels in point cloud classification, particularly on the ScanObjectNN dataset, achieving a 2.4% improvement over the baseline method.The code is available at <a href="https://github.com/2779705737/PointFEA/tree/master">https://github.com/2779705737/PointFEA/tree/master</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

PointFEA: beyond simple points, unveiling the power of raw positional features

  • 期刊论文

  • 发表日期：

    2025年5月15日

查看全文
  • Leyan Ren,
  • Jianming Wang,
  • Yukuan Sun,
  • Xiuyan Li

摘要

Rapid advancements in 3D acquisition technologies have significantly advanced the widespread application of point clouds in robotics and autonomous systems. However, the inherent unordered and unstructured nature of point clouds poses significant challenges for neural network design. Most methods neglect the potential of the features represented by the original positional information of the points. To capture comprehensive feature information to enhance local feature aggregation, we propose the Local Feature Aggregation Module (LFAM), which incorporates the original positional information and integrates it with edge features. To further enhance the classification accuracy and practical applicability of the model, we propose a novel point feature enhancement architecture, PointFEA, which initially employs LFAM and subsequently incorporates a channel-cluster-based attention mechanism. Our method optimally extracts local features while effectively integrating global contextual information. Extensive experimental evaluations demonstrate that PointFEA excels in point cloud classification, particularly on the ScanObjectNN dataset, achieving a 2.4% improvement over the baseline method.The code is available at https://github.com/2779705737/PointFEA/tree/master.