Scholarlink

<p>Visible-infrared person re-identification (VI-ReID) is a challenging retrieval task due to modality gaps, which further causes large intra-class and small inter-class distances. However, unpaired training samples and substantial modality discrepancies constrains model’s ability of cross-modality alignment. To handle this issue, we propose the dual-level joint align network (DJANet), which integrates a multi-modal alignment augmentation (MAA) module and a multi-scale feature aggregation (MFA) module. MAA employs bidirectional alignment to convert visible images to infrared and vice versa, enhancing data diversity and reducing modality differences at the image-level. MFA leverages a multi-head attention mechanism and a state space model to refine and aggregate multi-scale features, preserving critical information and strengthening cross-modal feature correlations, thereby reducing the modality differences at the feature-level. Experimental results on three benchmark datasets demonstrate the superiority of DJANet over state-of-the-art methods, achieving impressive performance gains. This study not only advances the field of VI-ReID but also highlights the potential of dual-level alignment strategies in cross-modality tasks. The code and related datasets are available at <a href="https://github.com/Hmn-92/DJANet">https://github.com/Hmn-92/DJANet</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

Bridging visible and infrared modalities: a dual-level joint align network for person re-identification

  • 期刊论文

  • 发表日期：

    2025年1月20日

查看全文
  • Mengnan Hu,
  • Qianli Zhou,
  • Rong Wang

摘要

Visible-infrared person re-identification (VI-ReID) is a challenging retrieval task due to modality gaps, which further causes large intra-class and small inter-class distances. However, unpaired training samples and substantial modality discrepancies constrains model’s ability of cross-modality alignment. To handle this issue, we propose the dual-level joint align network (DJANet), which integrates a multi-modal alignment augmentation (MAA) module and a multi-scale feature aggregation (MFA) module. MAA employs bidirectional alignment to convert visible images to infrared and vice versa, enhancing data diversity and reducing modality differences at the image-level. MFA leverages a multi-head attention mechanism and a state space model to refine and aggregate multi-scale features, preserving critical information and strengthening cross-modal feature correlations, thereby reducing the modality differences at the feature-level. Experimental results on three benchmark datasets demonstrate the superiority of DJANet over state-of-the-art methods, achieving impressive performance gains. This study not only advances the field of VI-ReID but also highlights the potential of dual-level alignment strategies in cross-modality tasks. The code and related datasets are available at https://github.com/Hmn-92/DJANet.