Scholarlink

<p>We tackle the task of domain-adaptive crowd counting, which aims to improve the accuracy of crowd counting models when applied to different domains or environments. Existing methods often fail to adequately address the various differences between source and target domains, such as different crowd density, image style, and data imbalance during training. Typically, these methods only focus on one or two of these differences, resulting in less-than-optimal performance. Our approach introduces a comprehensive optimization strategy that fully considers the diverse differences between source and target domains. First, we introduce a sample weighting strategy to tackle the data imbalance between source and target domains, enabling the model to effectively utilize data from both domains for better alignment. Second, to account for the variation in crowd density, we employ a conditional domain adversarial network, which can better adapt the model to the different crowd density of the target domain. Additionally, we incorporate style contrastive learning to reduce the bias of the model to the source domain with changed image style. Together, these techniques ensure closer alignment with the target domain and enhance overall performance. Experimental evaluations on several public datasets show that our method significantly outperforms previous approaches. Code is available at <a href="https://github.com/yjj511/cdascc">https://github.com/yjj511/cdascc</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

Advancing crowd counting accuracy in diverse environments via comprehensive domain alignment strategies

  • 期刊论文

  • 发表日期：

    2025年4月30日

查看全文
  • Junjie Ye,
  • Wenxiao Zhang,
  • Xun Yang

摘要

We tackle the task of domain-adaptive crowd counting, which aims to improve the accuracy of crowd counting models when applied to different domains or environments. Existing methods often fail to adequately address the various differences between source and target domains, such as different crowd density, image style, and data imbalance during training. Typically, these methods only focus on one or two of these differences, resulting in less-than-optimal performance. Our approach introduces a comprehensive optimization strategy that fully considers the diverse differences between source and target domains. First, we introduce a sample weighting strategy to tackle the data imbalance between source and target domains, enabling the model to effectively utilize data from both domains for better alignment. Second, to account for the variation in crowd density, we employ a conditional domain adversarial network, which can better adapt the model to the different crowd density of the target domain. Additionally, we incorporate style contrastive learning to reduce the bias of the model to the source domain with changed image style. Together, these techniques ensure closer alignment with the target domain and enhance overall performance. Experimental evaluations on several public datasets show that our method significantly outperforms previous approaches. Code is available at https://github.com/yjj511/cdascc.