Scholarlink

<p>Person re-identification (ReID) under challenging conditions such as occlusion and viewpoint variation requires robust and fine-grained feature representations. Although models based on Contrastive Language–Image Pre-training (CLIP) benefit from alignment in vision languages, they often overlook spatial structures and localized semantics, essential for accurate ReID. To overcome this limitation, we propose the Global Grouped Coordinate Attention (GGCA) module, which augments the CLIP visual encoder with spatially aware feature refinement. GGCA incorporates absolute spatial position embeddings into the attention mechanism and decomposes features into four semantically meaningful body regions: head, upper body, lower body, and accessories to enhance local discrimination. Furthermore, we employ structured textual prompts incorporating both identity and part-level semantics, whose embeddings are co-optimized with visual features to promote cross-modal consistency. Extensive experiments on four benchmark datasets validate the effectiveness of our approach. GGCA-CLIP achieves significant improvements over strong CLIP-based baselines, with a 2.2% mAP gain in Occluded-Duke and a 1.2% mAP increase in MSMT17. These results highlight the generality, efficiency, and interpretability of GGCA-CLIP for the spatially aware vision-language person ReID.</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

Global grouped coordinate attention for fine-grained person re-identification

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Zilong Li,
  • Jiashuai Xiao,
  • Jing Zhang

摘要

Person re-identification (ReID) under challenging conditions such as occlusion and viewpoint variation requires robust and fine-grained feature representations. Although models based on Contrastive Language–Image Pre-training (CLIP) benefit from alignment in vision languages, they often overlook spatial structures and localized semantics, essential for accurate ReID. To overcome this limitation, we propose the Global Grouped Coordinate Attention (GGCA) module, which augments the CLIP visual encoder with spatially aware feature refinement. GGCA incorporates absolute spatial position embeddings into the attention mechanism and decomposes features into four semantically meaningful body regions: head, upper body, lower body, and accessories to enhance local discrimination. Furthermore, we employ structured textual prompts incorporating both identity and part-level semantics, whose embeddings are co-optimized with visual features to promote cross-modal consistency. Extensive experiments on four benchmark datasets validate the effectiveness of our approach. GGCA-CLIP achieves significant improvements over strong CLIP-based baselines, with a 2.2% mAP gain in Occluded-Duke and a 1.2% mAP increase in MSMT17. These results highlight the generality, efficiency, and interpretability of GGCA-CLIP for the spatially aware vision-language person ReID.