Scholarlink

<p>Unsupervised domain adaptation (UDA) for person re-identification (Re-ID) faces major challenges due to domain shifts and noisy pseudo-labels. To enhance these methods, we propose a new framework based on a dynamically multi-modal transformer called MAFormer, integrated with a Mutual Update Pseudo-Labelling strategy. Our approach uses attribute-based feature learning to improve person Re-ID by combining visual features with textual semantic embeddings from Sentence BERT. MAFormer captures both global and part-level features and models inter-part relationships through cross-attention, making it more robust against occlusions and misalignments. In person Re-ID, we introduce a mutual learning paradigm involving three specialized networks focused on global, attribute-consistent, and occlusion-aware features. Pseudo-labels are initially generated using DBSCAN clustering and further optimized through cross-network top-k consensus and a confidence-based memory bank. Extensive experiments on Market-1501 and DukeMTMC-ReID demonstrate that our approach outperforms current state-of-the-art UDA methods, significantly improving Rank scores and the mean average precision. These results confirm the effectiveness of our model in efficient UDA-based person 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

MAFormer: a multimodal transformer framework with dynamic pseudo-labeling for reliable UDA-based person Re-ID

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Naeem Ahmed,
  • Qing Tian,
  • Muhammad Saeed

摘要

Unsupervised domain adaptation (UDA) for person re-identification (Re-ID) faces major challenges due to domain shifts and noisy pseudo-labels. To enhance these methods, we propose a new framework based on a dynamically multi-modal transformer called MAFormer, integrated with a Mutual Update Pseudo-Labelling strategy. Our approach uses attribute-based feature learning to improve person Re-ID by combining visual features with textual semantic embeddings from Sentence BERT. MAFormer captures both global and part-level features and models inter-part relationships through cross-attention, making it more robust against occlusions and misalignments. In person Re-ID, we introduce a mutual learning paradigm involving three specialized networks focused on global, attribute-consistent, and occlusion-aware features. Pseudo-labels are initially generated using DBSCAN clustering and further optimized through cross-network top-k consensus and a confidence-based memory bank. Extensive experiments on Market-1501 and DukeMTMC-ReID demonstrate that our approach outperforms current state-of-the-art UDA methods, significantly improving Rank scores and the mean average precision. These results confirm the effectiveness of our model in efficient UDA-based person Re-ID.