Scholarlink

<p>Person re-identification (ReID) is pivotal in intelligent surveillance systems, yet challenges persist due to variations in illumination, viewpoint, and domain shifts. Large-scale vision-language models (VLMs) offer promise but struggle with fine-grained discrimination and cross-domain generalization. This paper introduces MTA-CLIP, a multi-task adaptive framework leveraging CLIP for ReID, which decouples shared and domain-specific semantics via Hierarchical Contextual Prompting. A lightweight visual adaptation module enhances identity-relevant features, while a progressive two-stage optimization strategy balances knowledge preservation and task adaptation. Experiments on Market-1501, MSMT17, and DukeMTMC-reID demonstrate MTA-CLIP’s superior robustness and generalization. Under the triple-domain joint training setting, it achieves 92.5% mAP and 96.9% Rank-1 accuracy on Market-1501, outperforming existing CLIP-based methods. The code is available at <a href="https://github.com/Jiasmiao/MTA-CLIP">https://github.com/Jiasmiao/MTA-CLIP</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

Multi-task adaptive CLIP for enhanced person re-identification across domains

  • 期刊论文

  • 发表日期：

    2026年6月21日

查看全文
  • Yuanyuan Li,
  • Simiao Jia,
  • Rongrong Li,
  • Dongyan Li,
  • Dong Wang,
  • Wenjie Zhou

摘要

Person re-identification (ReID) is pivotal in intelligent surveillance systems, yet challenges persist due to variations in illumination, viewpoint, and domain shifts. Large-scale vision-language models (VLMs) offer promise but struggle with fine-grained discrimination and cross-domain generalization. This paper introduces MTA-CLIP, a multi-task adaptive framework leveraging CLIP for ReID, which decouples shared and domain-specific semantics via Hierarchical Contextual Prompting. A lightweight visual adaptation module enhances identity-relevant features, while a progressive two-stage optimization strategy balances knowledge preservation and task adaptation. Experiments on Market-1501, MSMT17, and DukeMTMC-reID demonstrate MTA-CLIP’s superior robustness and generalization. Under the triple-domain joint training setting, it achieves 92.5% mAP and 96.9% Rank-1 accuracy on Market-1501, outperforming existing CLIP-based methods. The code is available at https://github.com/Jiasmiao/MTA-CLIP.