Scholarlink

In recent years, transformers have significantly advanced 3D point cloud analysis by capturing long-range dependencies. Current methods focus on complex local feature processing, leading to substantial neighbor preparation and position embedding overhead, and fail to model remote context dependencies directly. Additionally, learned position embeddings struggle to integrate point cloud spatial information with deep features, impairing effective feature learning and target perception. We propose HAPE-Net, which excels in capturing remote context and demonstrates strong generalization and performance. Our method utilizes a Global Position Attention (GPA) module with Hierarchical Adaptive Position Encoding (HAPE), combined with an efficient Local Squeeze Aggregation (LSA) operator for local feature extraction, capturing shape context within and between point sets. HAPE adapts to network depth, promoting spatial and semantic information fusion. LSA and GPA effectively model short-range and long-range spaces, respectively. Our approach requires no additional operations like local position embedding or neighbor preparation. Experiments on ScanObjectNN, ShapeNetPart, and S3DIS demonstrate HAPE-Net’s effectiveness in point cloud classification and segmentation, surpassing previous methods.

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

Hierarchical Adaptive Position Encoding-Based Transformer for Point Cloud Analysis

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jiawei Yao,
  • Junfeng Yao,
  • Yong Yang,
  • Chunyang Huang

摘要

In recent years, transformers have significantly advanced 3D point cloud analysis by capturing long-range dependencies. Current methods focus on complex local feature processing, leading to substantial neighbor preparation and position embedding overhead, and fail to model remote context dependencies directly. Additionally, learned position embeddings struggle to integrate point cloud spatial information with deep features, impairing effective feature learning and target perception. We propose HAPE-Net, which excels in capturing remote context and demonstrates strong generalization and performance. Our method utilizes a Global Position Attention (GPA) module with Hierarchical Adaptive Position Encoding (HAPE), combined with an efficient Local Squeeze Aggregation (LSA) operator for local feature extraction, capturing shape context within and between point sets. HAPE adapts to network depth, promoting spatial and semantic information fusion. LSA and GPA effectively model short-range and long-range spaces, respectively. Our approach requires no additional operations like local position embedding or neighbor preparation. Experiments on ScanObjectNN, ShapeNetPart, and S3DIS demonstrate HAPE-Net’s effectiveness in point cloud classification and segmentation, surpassing previous methods.