Scholarlink

<p>RGB-T is a multimodal method for crowd counting. This method uses fused features from RGB images and thermal maps to enhance counting precision. Most existing methods rely on convolutional neural networks or attention-based networks. However, the differences between RGB and thermal modality features limit the accuracy of crowd counting. To better address these differences and leverage the information from both modalities, a Deep Fusion Mamba Network (DFMNet) is proposed to enhance crowd counting performance. Specifically, a Deep Fusion Mamba Module (DFMM) is designed to map cross-modal features into a hidden state space and facilitate high-frequency feature interactions between RGB features and thermal map features. These frequent feature interactions reduce modality differences to some extent and achieve more effective feature fusion. DFMM consists of two submodules: the Channel Swapping Module (CSM) for shallow feature interactions and the Cross-Modal Feature Interaction Module (CFMM) for frequent cross-modal feature exchanges. The DFMM module is integrated into the network layers to construct DFMNet. This investigation into Mamba’s potential for cross-modal fusion has made significant progress and demonstrates robust performance in multimodal crowd counting tasks. The performance of our method is assessed using the RGBT-CC dataset and the ShanghaiTech RGBD dataset. The experimental results highlight the robustness and effectiveness of the model in various scenarios. The source code will be publicly available at <a href="https://github.com/c1299877959/DFMNet.">https://github.com/c1299877959/DFMNet.</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

DFMNet: deep fusion mamba network for multimodal crowd counting

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Yubo Yang,
  • Liangjun Huang,
  • Yangjian Chen,
  • Wencan Kang,
  • Jianwei Zhang

摘要

RGB-T is a multimodal method for crowd counting. This method uses fused features from RGB images and thermal maps to enhance counting precision. Most existing methods rely on convolutional neural networks or attention-based networks. However, the differences between RGB and thermal modality features limit the accuracy of crowd counting. To better address these differences and leverage the information from both modalities, a Deep Fusion Mamba Network (DFMNet) is proposed to enhance crowd counting performance. Specifically, a Deep Fusion Mamba Module (DFMM) is designed to map cross-modal features into a hidden state space and facilitate high-frequency feature interactions between RGB features and thermal map features. These frequent feature interactions reduce modality differences to some extent and achieve more effective feature fusion. DFMM consists of two submodules: the Channel Swapping Module (CSM) for shallow feature interactions and the Cross-Modal Feature Interaction Module (CFMM) for frequent cross-modal feature exchanges. The DFMM module is integrated into the network layers to construct DFMNet. This investigation into Mamba’s potential for cross-modal fusion has made significant progress and demonstrates robust performance in multimodal crowd counting tasks. The performance of our method is assessed using the RGBT-CC dataset and the ShanghaiTech RGBD dataset. The experimental results highlight the robustness and effectiveness of the model in various scenarios. The source code will be publicly available at https://github.com/c1299877959/DFMNet.