Scholarlink

<p>In recent years, Transformer- and Mamba-based models have achieved remarkable progress in point cloud analysis, yet both exhibit notable limitations. Transformers suffer from quadratic complexity, while existing Mamba-based approaches, despite their efficiency, still suffer from several limitations, including the loss of spatial information during point cloud serialization, limited ability to capture fine-grained local geometric details, and reduced capacity to preserve spatial diversity when modeling complex global dependencies. To address these limitations, we propose OmniMamba, a Mamba-driven framework built upon three key designs. First, a Multi-Scale Feature Processor (MSFP) introduces topology-aware dimensional transformations to preserve spatial structure during serialization. Second, an Energy-influenced Parallel Mamba (EPM) module enhances local feature discrimination by incorporating energy-guided parallel state space modeling, improving the representation of fine-grained geometry without sacrificing efficiency. Third, a bidirectional state space fusion mechanism enables symmetric context propagation, effectively maintaining spatial diversity in global dependency modeling. Experimental results on ModelNet40 and PB_T50_RS demonstrate that OmniMamba achieves competitive classification accuracies of 95.1% and 93.34%. Furthermore, OmniMamba has demonstrated outstanding performance in few-shot classification tasks. On ShapeNetPart for point cloud segmentation, it achieves a <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(mIoU_I\)</EquationSource> </InlineEquation> of 86.6% and a <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(mIoU_C\)</EquationSource> </InlineEquation> of 84.8%. Furthermore, our method reduces memory usage by over 95.9% compared with the representative Transformer-based model, PointTransformer. These results highlight the strong potential of Mamba-driven frameworks for efficient and effective 3D point cloud processing. Our code is available at <a href="https://github.com/slc1112/OmniMamba">https://github.com/slc1112/OmniMamba</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

OmniMamba: a novel Mamba-driven framework for point cloud analysis

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Luchen Shen,
  • Junhai Zhai,
  • Hao Chen

摘要

In recent years, Transformer- and Mamba-based models have achieved remarkable progress in point cloud analysis, yet both exhibit notable limitations. Transformers suffer from quadratic complexity, while existing Mamba-based approaches, despite their efficiency, still suffer from several limitations, including the loss of spatial information during point cloud serialization, limited ability to capture fine-grained local geometric details, and reduced capacity to preserve spatial diversity when modeling complex global dependencies. To address these limitations, we propose OmniMamba, a Mamba-driven framework built upon three key designs. First, a Multi-Scale Feature Processor (MSFP) introduces topology-aware dimensional transformations to preserve spatial structure during serialization. Second, an Energy-influenced Parallel Mamba (EPM) module enhances local feature discrimination by incorporating energy-guided parallel state space modeling, improving the representation of fine-grained geometry without sacrificing efficiency. Third, a bidirectional state space fusion mechanism enables symmetric context propagation, effectively maintaining spatial diversity in global dependency modeling. Experimental results on ModelNet40 and PB_T50_RS demonstrate that OmniMamba achieves competitive classification accuracies of 95.1% and 93.34%. Furthermore, OmniMamba has demonstrated outstanding performance in few-shot classification tasks. On ShapeNetPart for point cloud segmentation, it achieves a \(mIoU_I\) of 86.6% and a \(mIoU_C\) of 84.8%. Furthermore, our method reduces memory usage by over 95.9% compared with the representative Transformer-based model, PointTransformer. These results highlight the strong potential of Mamba-driven frameworks for efficient and effective 3D point cloud processing. Our code is available at https://github.com/slc1112/OmniMamba.