Scholarlink

<p>Person re-identification (Re-ID) aims to learn discriminative features for identifying individuals across non-overlapping cameras without manual annotations. Recent unsupervised methods rely on noisy pseudo labels, hindering generalization. We propose UPCA, a unified framework integrating three components to enhance pseudo-label quality and feature representation. The Part-based Pseudo Label Refinement (PPLR) module enforces global–local feature consistency to reduce label noise without extra models. Camera-aware Jaccard distance (CaJaccard) refines cross-camera similarity via camera-aware k-reciprocal neighbors (CKRNNs) and camera-level query expansion (CLQE), mitigating unreliable matches ignored by traditional Jaccard. Convolutional Block Attention Module (CBAM) enhances part-level features by emphasizing informative spatial and channel cues. Experiments on Market-1501, DukeMTMC-reID, MSMT17, and VeRi-776 show UPCA outperforms PPLR by + 2.0%/+1.3%, + 1.5%/+0.8%, and + 1.6%/+1.2% mAP/Rank-1 on Market-1501, MSMT17, and VeRi-776, and achieves 75.3% mAP and 86.4% Rank-1 on DukeMTMC-reID, demonstrating robustness and generalization in fully unsupervised Re-ID.</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

UPCA: unsupervised person re-identification with camera-aware distance and lightweight attention

  • 期刊论文

  • 发表日期：

    2025年10月25日

查看全文
  • Chenjie Zhang,
  • Xiangxin Kai,
  • Xuemei Bai,
  • Hanping Hu

摘要

Person re-identification (Re-ID) aims to learn discriminative features for identifying individuals across non-overlapping cameras without manual annotations. Recent unsupervised methods rely on noisy pseudo labels, hindering generalization. We propose UPCA, a unified framework integrating three components to enhance pseudo-label quality and feature representation. The Part-based Pseudo Label Refinement (PPLR) module enforces global–local feature consistency to reduce label noise without extra models. Camera-aware Jaccard distance (CaJaccard) refines cross-camera similarity via camera-aware k-reciprocal neighbors (CKRNNs) and camera-level query expansion (CLQE), mitigating unreliable matches ignored by traditional Jaccard. Convolutional Block Attention Module (CBAM) enhances part-level features by emphasizing informative spatial and channel cues. Experiments on Market-1501, DukeMTMC-reID, MSMT17, and VeRi-776 show UPCA outperforms PPLR by + 2.0%/+1.3%, + 1.5%/+0.8%, and + 1.6%/+1.2% mAP/Rank-1 on Market-1501, MSMT17, and VeRi-776, and achieves 75.3% mAP and 86.4% Rank-1 on DukeMTMC-reID, demonstrating robustness and generalization in fully unsupervised Re-ID.