Scholarlink

<p>Generalized category discovery (GCD) faces a primary challenge of simultaneously discovering new classes while maintaining the classification performance on known classes, which becomes more prominent in fine-grained recognition where inter-class differences could be very slight. To tackle this challenge, we propose a joint learning method based on class-aware clustering, which effectively promotes cluster consistency by introducing remapped true labels and robust pseudo-labels. Furthermore, we formulate individual identification as GCD by incorporating the concept of “unknown individual discovery" to make individual identification more aligned with real-world scenarios. Extensive experiments on multiple fine-grained image classification datasets and animal identification datasets demonstrate the superiority of our method over state-of-the-art methods in fine-grained recognition tasks. On fine-grained datasets (CUB-200, Stanford-Cars, and Herbarium19), it achieves 3.0%, 1.6%, and 0.5% higher accuracy on <i>all</i> classes, respectively; on animal identification datasets (OpenCows2020, Redpanda, and Ipanda-50), it achieves 1.8%, 0.3%, and 3.6% higher accuracy on <i>all</i> individuals, respectively.</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

Generalized category discovery in fine-grained recognition

  • 期刊论文

  • 发表日期：

    2025年10月14日

查看全文
  • Yifu Yuan,
  • Qijun Zhao,
  • Yue Yang,
  • Peng Chen

摘要

Generalized category discovery (GCD) faces a primary challenge of simultaneously discovering new classes while maintaining the classification performance on known classes, which becomes more prominent in fine-grained recognition where inter-class differences could be very slight. To tackle this challenge, we propose a joint learning method based on class-aware clustering, which effectively promotes cluster consistency by introducing remapped true labels and robust pseudo-labels. Furthermore, we formulate individual identification as GCD by incorporating the concept of “unknown individual discovery" to make individual identification more aligned with real-world scenarios. Extensive experiments on multiple fine-grained image classification datasets and animal identification datasets demonstrate the superiority of our method over state-of-the-art methods in fine-grained recognition tasks. On fine-grained datasets (CUB-200, Stanford-Cars, and Herbarium19), it achieves 3.0%, 1.6%, and 0.5% higher accuracy on all classes, respectively; on animal identification datasets (OpenCows2020, Redpanda, and Ipanda-50), it achieves 1.8%, 0.3%, and 3.6% higher accuracy on all individuals, respectively.