Scholarlink

<p>Vision-Language Pretraining (VLP) models have shown significant promise in scene understanding and representation learning. However, their application in fine-grained tasks like Cloth-Changing Person Re-Identification (CC-ReID) is challenging due to their reliance on unstable discriminative features such as clothing. Conversely, expert CC-ReID models possess exceptional fine-grained comprehension skills but struggle to obtain reliable cloth-agnostic representations, hindered by the time-consuming and labor-intensive process of obtaining precise annotations and the spurious data associations brought by the co-occurrence phenomenon of identity and clothing. This paper introduces the <b>Ca</b>usality-based <b>Pu</b>rification (CaPu) model. The CaPu constructs a clothing indication pipeline that leverages the unique strengths of multiple VLP models to efficiently capture clothing semantics. Utilizing these semantics, CaPu employs causality analysis to purify the relationship between learned visual features and intrinsic identity representation from two causal aspects: the Consistency Treatment Effect (CTE) and the Distinctiveness Treatment Effect (DTE). The CTE enhances feature consistency within each identity by simulating clothing changes. Meanwhile, the DTE enhances the model’s ability to perceive intrinsic identity representation. Extensive experiments on three standard CC-ReID datasets demonstrate that CaPu achieves state-of-the-art performance.</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

Clothing Purification with Causality Meets Vision-Language Pretraining Models

  • 期刊论文

  • 发表日期：

    2025年8月11日

查看全文
  • Zhengwei Yang,
  • Huilin Zhu,
  • Nan Lei,
  • Basura Fernando,
  • Zheng Wang

摘要

Vision-Language Pretraining (VLP) models have shown significant promise in scene understanding and representation learning. However, their application in fine-grained tasks like Cloth-Changing Person Re-Identification (CC-ReID) is challenging due to their reliance on unstable discriminative features such as clothing. Conversely, expert CC-ReID models possess exceptional fine-grained comprehension skills but struggle to obtain reliable cloth-agnostic representations, hindered by the time-consuming and labor-intensive process of obtaining precise annotations and the spurious data associations brought by the co-occurrence phenomenon of identity and clothing. This paper introduces the Causality-based Purification (CaPu) model. The CaPu constructs a clothing indication pipeline that leverages the unique strengths of multiple VLP models to efficiently capture clothing semantics. Utilizing these semantics, CaPu employs causality analysis to purify the relationship between learned visual features and intrinsic identity representation from two causal aspects: the Consistency Treatment Effect (CTE) and the Distinctiveness Treatment Effect (DTE). The CTE enhances feature consistency within each identity by simulating clothing changes. Meanwhile, the DTE enhances the model’s ability to perceive intrinsic identity representation. Extensive experiments on three standard CC-ReID datasets demonstrate that CaPu achieves state-of-the-art performance.