Scholarlink

<p>Recently, Cross-Domain Adaptive Crowd Counting (CDACC) has received extensive attention due to its independence from large numbers of expensive annotation samples. CDACC aims to train a counting model using the low-cost labeled data as the source domain and generalize it to the target domain. At present, typical methods seek to minimize domain discrepancies by aligning the data, features, and parameters of the source domain with those of the target domain, significantly alleviating the performance degradation because of huge domain shifts. However, they ignore the inherent diversity within the source domain data and consistency learning strategy, when applied in realistic scenarios. This paper proposes a novelty CDACC method, termed joint perturbation consistency (JPC), to improve the generalization from the source domain to the target domain via a self-training framework. Specifically, our method is designed by refining pre-training procedure to fully leverage the diverse data available in the source domain, aiming to boost the model's generalization capabilities. Furthermore, we introduce a targeted perturbation selection strategy for consistency learning, which identifies appropriate perturbations to improve the model robustness. We also adapt the consistency learning framework to better align with the model's learning objectives, thereby extending the domain generalization capabilities and facilitating effective CDACC. The experimental results evaluated on the four target datasets demonstrate that our proposed framework outperforms some CDACC methods, highlighting its effectiveness in realistic scenarios. The code is available at: <a href="https://github.com/jax0619/JPC">https://github.com/jax0619/JPC</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

Joint perturbation consistency across image and feature levels for cross-domain adaptive crowd counting

  • 期刊论文

  • 发表日期：

    2025年3月5日

查看全文
  • Xie Chengjie,
  • Lu Shuhua,
  • Shi Yangyu,
  • Zheng Diwen

摘要

Recently, Cross-Domain Adaptive Crowd Counting (CDACC) has received extensive attention due to its independence from large numbers of expensive annotation samples. CDACC aims to train a counting model using the low-cost labeled data as the source domain and generalize it to the target domain. At present, typical methods seek to minimize domain discrepancies by aligning the data, features, and parameters of the source domain with those of the target domain, significantly alleviating the performance degradation because of huge domain shifts. However, they ignore the inherent diversity within the source domain data and consistency learning strategy, when applied in realistic scenarios. This paper proposes a novelty CDACC method, termed joint perturbation consistency (JPC), to improve the generalization from the source domain to the target domain via a self-training framework. Specifically, our method is designed by refining pre-training procedure to fully leverage the diverse data available in the source domain, aiming to boost the model's generalization capabilities. Furthermore, we introduce a targeted perturbation selection strategy for consistency learning, which identifies appropriate perturbations to improve the model robustness. We also adapt the consistency learning framework to better align with the model's learning objectives, thereby extending the domain generalization capabilities and facilitating effective CDACC. The experimental results evaluated on the four target datasets demonstrate that our proposed framework outperforms some CDACC methods, highlighting its effectiveness in realistic scenarios. The code is available at: https://github.com/jax0619/JPC.