Scholarlink

<p>The success of existing image classification models largely depends on fitting labeled samples for every image category. This raises a critical issue: conventional classification models trained on limited categories can only identify those seen categories included in the training set, but fail to recognize unknown categories. Generalized Category Discovery aims to classify data into known and unknown categories. However, two major challenges hinder performance: (1) the lack of supervisory signals for unknown classes, making their discovery difficult, and (2) high inter-class similarity between known and unknown categories, leading to ambiguity in class differentiation. To address these challenges, we propose Class Alignment and Boundary Calibration (CABC) for Generalized Category Discovery. Specifically, CABC enforces class alignment to improve within-class compactness, while boundary calibration enhances between-class and known–unknown separability, thereby yielding more discriminative feature spaces. Additionally, we train the classifier using a distance-weighted pseudo-label loss, which takes into account the proximity of unlabeled samples to class centers, reducing the influence of noisy pseudo-labels. CABC effectively improves the discrimination of known and unknown categories, achieving strong performance across a variety of benchmarks and demonstrating its robustness in complex tasks.</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

Class alignment and boundary calibration for generalized category discovery

  • 期刊论文

  • 发表日期：

    2026年5月11日

查看全文
  • Xiao Li,
  • Min Fang,
  • Guizhi Wang,
  • Haixiang Li

摘要

The success of existing image classification models largely depends on fitting labeled samples for every image category. This raises a critical issue: conventional classification models trained on limited categories can only identify those seen categories included in the training set, but fail to recognize unknown categories. Generalized Category Discovery aims to classify data into known and unknown categories. However, two major challenges hinder performance: (1) the lack of supervisory signals for unknown classes, making their discovery difficult, and (2) high inter-class similarity between known and unknown categories, leading to ambiguity in class differentiation. To address these challenges, we propose Class Alignment and Boundary Calibration (CABC) for Generalized Category Discovery. Specifically, CABC enforces class alignment to improve within-class compactness, while boundary calibration enhances between-class and known–unknown separability, thereby yielding more discriminative feature spaces. Additionally, we train the classifier using a distance-weighted pseudo-label loss, which takes into account the proximity of unlabeled samples to class centers, reducing the influence of noisy pseudo-labels. CABC effectively improves the discrimination of known and unknown categories, achieving strong performance across a variety of benchmarks and demonstrating its robustness in complex tasks.