Scholarlink

<p>The distribution of data in real-world applications often shows a long-tailed shape, making long-tailed recognition challenging as traditional models bias toward majority categories. Multi-expert ensemble methods have shown promise but often suffer from insufficient expert diversity and high model variance. To address these issues, we propose multiple experts with knowledge fusion (MEKF), which includes diversified fusion experts and dual-view self-distillation. MEKF enhances expert diversity by fusing features of different depths and introduces distribution diversity loss with distribution weights. Dual-view self-distillation reduces model variance by extracting semantic information from weakly augmented data predictions. Experiments on CIFAR100-LT, ImageNet-LT, and Places-LT benchmarks validate MEKF’s effectiveness, showing excellent performance.</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

MEKF: long-tailed visual recognition via multiple experts with knowledge fusion

  • 期刊论文

  • 发表日期：

    2025年1月20日

查看全文
  • Qian Zhang,
  • Chenghao Ji,
  • Mingwen Shao,
  • Hong Liang

摘要

The distribution of data in real-world applications often shows a long-tailed shape, making long-tailed recognition challenging as traditional models bias toward majority categories. Multi-expert ensemble methods have shown promise but often suffer from insufficient expert diversity and high model variance. To address these issues, we propose multiple experts with knowledge fusion (MEKF), which includes diversified fusion experts and dual-view self-distillation. MEKF enhances expert diversity by fusing features of different depths and introduces distribution diversity loss with distribution weights. Dual-view self-distillation reduces model variance by extracting semantic information from weakly augmented data predictions. Experiments on CIFAR100-LT, ImageNet-LT, and Places-LT benchmarks validate MEKF’s effectiveness, showing excellent performance.