Scholarlink

<p>Crowd counting has played a significant role in video surveillance, social security and other practical applications. However, the counting performance encounters a sharp degradation when the testing and training scenarios of the model are different, the so-called “domain shift/gap” problem. Although domain adaptive crowd counting approaches achieves a large progress in reducing domain gap, they typically need to adapt or update the pre-trained model with some target domain data. In this paper, we aim to generalize the trained model on a single source domain to any unseen target domain, without requiring target data or model updates. To this end, a new adversarial domain generalization framework is proposed to learn domain-invariant feature representations. In this framework, we tackle the domain shift problem in crowd counting by means of single domain generalization (SDG) technique. First, we conduct data augmentation at both pixel level and feature level to construct a simulated target domain. Next, with the source domain and its augmented domain, we design a feature extractor and a domain discriminator which are trained in an adversarial relationship to learn a generalized feature space. Last, we develop a distribution aligner to minimize the distribution difference between the source domain and its augmented domain. Thus, the learned feature space is also domain-agnostic in terms of feature distributions, contributing to enhance the domain generalization capability of the trained model. Extensive experiments on four public crowd counting datasets, ShanghaiTech Part_A &amp; Part_B, UCF-QNRF and NWPU-Crowd, demonstrate that our proposed framework can achieve superior generalization performance to the state-of-the-art methods.</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

Adversarial domain generalization for crowd counting in unseen scenarios

  • 期刊论文

  • 发表日期：

    2026年2月1日

查看全文
  • De Zhang,
  • Zishan Liang

摘要

Crowd counting has played a significant role in video surveillance, social security and other practical applications. However, the counting performance encounters a sharp degradation when the testing and training scenarios of the model are different, the so-called “domain shift/gap” problem. Although domain adaptive crowd counting approaches achieves a large progress in reducing domain gap, they typically need to adapt or update the pre-trained model with some target domain data. In this paper, we aim to generalize the trained model on a single source domain to any unseen target domain, without requiring target data or model updates. To this end, a new adversarial domain generalization framework is proposed to learn domain-invariant feature representations. In this framework, we tackle the domain shift problem in crowd counting by means of single domain generalization (SDG) technique. First, we conduct data augmentation at both pixel level and feature level to construct a simulated target domain. Next, with the source domain and its augmented domain, we design a feature extractor and a domain discriminator which are trained in an adversarial relationship to learn a generalized feature space. Last, we develop a distribution aligner to minimize the distribution difference between the source domain and its augmented domain. Thus, the learned feature space is also domain-agnostic in terms of feature distributions, contributing to enhance the domain generalization capability of the trained model. Extensive experiments on four public crowd counting datasets, ShanghaiTech Part_A & Part_B, UCF-QNRF and NWPU-Crowd, demonstrate that our proposed framework can achieve superior generalization performance to the state-of-the-art methods.