Scholarlink

<p>Real-world deep learning training data often follow a long-tailed (LT) distribution, where a few classes (head classes) have the most samples and many classes (tail classes) have very few samples. Models trained on LT datasets typically achieve high accuracy on head classes, but suffer from poor performance on tail classes. To address this challenge, strategies based on supervised contrastive learning have been explored. However, existing methods often focus on either reducing the dominance of head class features or expanding the feature space of tail classes, but rarely achieve a balanced feature distribution across both. In this paper, we propose Balanced clustering contrastive learning (BCCL) to balance the feature space between the head and tail classes more effectively. The proposed approach introduces two main components. First, we employ queue-based clustering to extract multiple centroids. This addresses the intra-minibatch class absence issue and maintains intra-class balance. Second, we expand the feature space of tail classes based on class frequency to enhance their expressiveness. An evaluation of four LT datasets, CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018, demonstrates that BCCL consistently outperforms the existing methods. These results establish the ability of BCCL to maintain a balanced feature space in diverse environments. Our code is available at <a href="https://github.com/GGTINE/BCCL">https://github.com/GGTINE/BCCL</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

Balanced clustering contrastive learning for long-tailed visual recognition

  • 期刊论文

  • 发表日期：

    2025年1月15日

查看全文
  • Byeong-il Kim,
  • Byoung Chul Ko

摘要

Real-world deep learning training data often follow a long-tailed (LT) distribution, where a few classes (head classes) have the most samples and many classes (tail classes) have very few samples. Models trained on LT datasets typically achieve high accuracy on head classes, but suffer from poor performance on tail classes. To address this challenge, strategies based on supervised contrastive learning have been explored. However, existing methods often focus on either reducing the dominance of head class features or expanding the feature space of tail classes, but rarely achieve a balanced feature distribution across both. In this paper, we propose Balanced clustering contrastive learning (BCCL) to balance the feature space between the head and tail classes more effectively. The proposed approach introduces two main components. First, we employ queue-based clustering to extract multiple centroids. This addresses the intra-minibatch class absence issue and maintains intra-class balance. Second, we expand the feature space of tail classes based on class frequency to enhance their expressiveness. An evaluation of four LT datasets, CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018, demonstrates that BCCL consistently outperforms the existing methods. These results establish the ability of BCCL to maintain a balanced feature space in diverse environments. Our code is available at https://github.com/GGTINE/BCCL.