Scholarlink

<p>Domain generalization aims to train models on multiple source domains to perform well on unseen target domains. Large vision-language backbones such as CLIP demonstrate strong zero-shot transfer and have sparked many parameter-efficient ways to leverage pre-trained features. However, ambiguous labels are universal in real-world data but largely ignored, which not only shrinks out-of-distribution robustness but also generalization. To address this challenge, we introduce a concise two-stage remedy: <b>LA</b>bel disambiguation first, then <b>D</b>omain-<b>A</b>ware learning (LADA). First, we convert hard labels into calibrated soft labels, down-weighting low-confidence samples. Then, we capture domain-related information via a domain-aware prompt and a prototype-based projection head with the enriched soft labels. These two coupled steps encode more informative cues for each domain, unleashing CLIP’s potential. Our LADA delivers a +5.7% gain on average and markedly improves performance on the most distribution-shifted datasets for the ResNet-50 CLIP (OpenAI) backbone. Furthermore, extensive model analysis demonstrates the superiority of our method in effectiveness, robustness, efficacy, flexibility, and efficiency.</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

LADA: Label Disambiguation and Domain-Aware Learning for Domain Generalization

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Zhiqing Xiao,
  • Haobo Wang,
  • Yali Ye,
  • Wentao Ye,
  • Hao Chen,
  • Gang Chen,
  • Junbo Zhao,
  • Rex Ying

摘要

Domain generalization aims to train models on multiple source domains to perform well on unseen target domains. Large vision-language backbones such as CLIP demonstrate strong zero-shot transfer and have sparked many parameter-efficient ways to leverage pre-trained features. However, ambiguous labels are universal in real-world data but largely ignored, which not only shrinks out-of-distribution robustness but also generalization. To address this challenge, we introduce a concise two-stage remedy: LAbel disambiguation first, then Domain-Aware learning (LADA). First, we convert hard labels into calibrated soft labels, down-weighting low-confidence samples. Then, we capture domain-related information via a domain-aware prompt and a prototype-based projection head with the enriched soft labels. These two coupled steps encode more informative cues for each domain, unleashing CLIP’s potential. Our LADA delivers a +5.7% gain on average and markedly improves performance on the most distribution-shifted datasets for the ResNet-50 CLIP (OpenAI) backbone. Furthermore, extensive model analysis demonstrates the superiority of our method in effectiveness, robustness, efficacy, flexibility, and efficiency.