Scholarlink

<p>Advancements in deep learning have significantly propelled point cloud analysis, particularly in point cloud classification. However, traditional methods often struggle with unstructured and sparse 3D point cloud data, posing challenges for effective feature extraction. Transformer-based models, leveraging self-attention, excel in capturing complex spatial information but suffer from quadratic computational complexity. On the other hand, Mamba modules offer linear complexity but primarily focus on global spatial structures, neglecting local interactions. To address these limitations, we introduce DTM, a hybrid model integrating Transformer and Mamba modules, enriched with density information. DTM facilitates the seamless integration of local and global features through a progressive feature learning process, incorporating local norm pooling, self-attention, and state space modeling (SSM). The incorporation of density information enhances the model’s capability to capture spatial structures, enabling more accurate feature extraction by differentiating regions. Extensive experiments demonstrate that DTM outperforms existing Transformer- and Mamba-based models across various benchmarks. Moreover, DTM maintains near-linear complexity, making it suitable for large-scale applications. The source code will be available at <a href="https://github.com/Nero-XD/DTM">https://github.com/Nero-XD/DTM</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

Dtm: Density embeded transformer mamba hybird network for point cloud analysis

  • 期刊论文

  • 发表日期：

    2025年6月14日

查看全文
  • Yaojie Chen,
  • Chengyu Deng

摘要

Advancements in deep learning have significantly propelled point cloud analysis, particularly in point cloud classification. However, traditional methods often struggle with unstructured and sparse 3D point cloud data, posing challenges for effective feature extraction. Transformer-based models, leveraging self-attention, excel in capturing complex spatial information but suffer from quadratic computational complexity. On the other hand, Mamba modules offer linear complexity but primarily focus on global spatial structures, neglecting local interactions. To address these limitations, we introduce DTM, a hybrid model integrating Transformer and Mamba modules, enriched with density information. DTM facilitates the seamless integration of local and global features through a progressive feature learning process, incorporating local norm pooling, self-attention, and state space modeling (SSM). The incorporation of density information enhances the model’s capability to capture spatial structures, enabling more accurate feature extraction by differentiating regions. Extensive experiments demonstrate that DTM outperforms existing Transformer- and Mamba-based models across various benchmarks. Moreover, DTM maintains near-linear complexity, making it suitable for large-scale applications. The source code will be available at https://github.com/Nero-XD/DTM.