Scholarlink

<p>The cross-modal crowd counting methods introduce additional modal features to more effectively address the challenges of counting tasks in complex lighting scenes. To fully encode the contextual semantic information from different modalities, we propose a Cross-attention Multi-scale Fusion Network (CMFNet) by combining attention mechanism. Specifically, Cross-modal feature Extraction-Aggregation Module (CEAM) has been developed with the purpose of fusing complementary features during the process of cross-modal information interaction. The module employs cross-attention in order to achieve feature alignment in an adaptive manner. Furthermore, the Multi-scale Context Aggregation Module (MCAM) has been proposed to perform multi-level feature extraction on the fusion features obtained from CEAM, thereby facilitating the efficient mining of disparate levels of information from the cross-modal fusion features. The high-quality crowd density map is calculated through a regression head based on the feature map aggregated from MCAM. Experimental results on three popular datasets demonstrate that our approach not only shows competitive performance, but also indicates the good generalizability for RGB-T and RGB-D cross-modal crowd counting tasks. The source code is available at <a href="https://github.com/Ping134/CMFNet">https://github.com/Ping134/CMFNet</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

CMFNet: Cross-attention multi-scale fusion network for cross-modal crowd counting

  • 期刊论文

  • 发表日期：

    2025年11月10日

查看全文
  • Shihui Zhang,
  • Ping Li,
  • Jiawei Zhang,
  • Zhigang Huang,
  • Sheng Zhan

摘要

The cross-modal crowd counting methods introduce additional modal features to more effectively address the challenges of counting tasks in complex lighting scenes. To fully encode the contextual semantic information from different modalities, we propose a Cross-attention Multi-scale Fusion Network (CMFNet) by combining attention mechanism. Specifically, Cross-modal feature Extraction-Aggregation Module (CEAM) has been developed with the purpose of fusing complementary features during the process of cross-modal information interaction. The module employs cross-attention in order to achieve feature alignment in an adaptive manner. Furthermore, the Multi-scale Context Aggregation Module (MCAM) has been proposed to perform multi-level feature extraction on the fusion features obtained from CEAM, thereby facilitating the efficient mining of disparate levels of information from the cross-modal fusion features. The high-quality crowd density map is calculated through a regression head based on the feature map aggregated from MCAM. Experimental results on three popular datasets demonstrate that our approach not only shows competitive performance, but also indicates the good generalizability for RGB-T and RGB-D cross-modal crowd counting tasks. The source code is available at https://github.com/Ping134/CMFNet.