Scholarlink

<p>The task of visible-infrared person re-identification (VI-ReID) involves matching images of pedestrians captured in visible light with their corresponding infrared representations. This paper proposes an innovative approach to enhance cross-modality person re-identification by leveraging an attention mechanism and asymmetric multi-granularity feature learning. We introduce a middle modality generator to bridge the gap between visible and infrared modalities, followed by a four-branch parameter sharing network (FBPN) that extracts features across these modalities. A lightweight channel attention module is incorporated to minimize the impact of background information. Furthermore, an asymmetric multi-granularity feature learning (AMFL) module is proposed to simultaneously capture global and local features, ensuring diversity in feature extraction. Experimental results on benchmark datasets demonstrate that our method significantly reduces modality disparities and achieves state-of-the-art performance in VI-ReID, outperforming existing approaches.</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

Enhancing cross-modality person re-identification through attention-guided asymmetric feature learning

  • 期刊论文

  • 发表日期：

    2025年3月16日

查看全文
  • Xuehua Song,
  • Junxing Zhou,
  • Hua Jin,
  • Xin Yuan,
  • Changda Wang

摘要

The task of visible-infrared person re-identification (VI-ReID) involves matching images of pedestrians captured in visible light with their corresponding infrared representations. This paper proposes an innovative approach to enhance cross-modality person re-identification by leveraging an attention mechanism and asymmetric multi-granularity feature learning. We introduce a middle modality generator to bridge the gap between visible and infrared modalities, followed by a four-branch parameter sharing network (FBPN) that extracts features across these modalities. A lightweight channel attention module is incorporated to minimize the impact of background information. Furthermore, an asymmetric multi-granularity feature learning (AMFL) module is proposed to simultaneously capture global and local features, ensuring diversity in feature extraction. Experimental results on benchmark datasets demonstrate that our method significantly reduces modality disparities and achieves state-of-the-art performance in VI-ReID, outperforming existing approaches.