Scholarlink

<p>Domain adaptation has achieved remarkable progress in computer vision. However, in scenarios involving multiple target domains—each exhibiting distinct characteristics and distribution patterns—traditional single-target domain adaptation methods often fall short. Although research on multi-target domain adaptation is growing, many existing approaches rely on strong assumptions about the target domains. Moreover, the substantial distributional discrepancies among different target domains often limit the effectiveness of alignment strategies. To address these problems, we propose an active dual alignment strategy that integrates active learning with all-way domain and conditional discriminative alignment to optimize cross-domain feature adaptation. Specifically, informative target samples are first selected for annotation via active learning. Then, conditional discrimination is employed to enhance category-aware feature alignment, while all-way domain discrimination ensures consistency across domains at the holistic feature level. Additionally, we incorporate a multilevel feature fusion module to further improve the model’s generalization capability. Extensive experiments on benchmark datasets such as Office-Home and Office-31 demonstrate that our approach significantly outperforms existing methods, particularly under limited annotation budgets, thereby validating the effectiveness of the proposed ADA strategy in multi-target domain adaptation settings. Notably, both the multistage active sampling and multi-domain parallel feature alignment in MT-ADA impose rigid computing power requirements on high-performance computing. Therefore, this paper conducts core verification using a high-performance standalone machine, and the proposed modular design demonstrates favorable scalability. Our code is available at <a href="https://github.com/JSJ515-Group/MT-ADA">https://github.com/JSJ515-Group/MT-ADA</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

MT-ADA: active dual alignment for multi-target domain adaptation

  • 期刊论文

  • 发表日期：

    2026年7月22日

查看全文
  • Yu-e Lin,
  • Xiuhe Deng,
  • Xingzhu Liang

摘要

Domain adaptation has achieved remarkable progress in computer vision. However, in scenarios involving multiple target domains—each exhibiting distinct characteristics and distribution patterns—traditional single-target domain adaptation methods often fall short. Although research on multi-target domain adaptation is growing, many existing approaches rely on strong assumptions about the target domains. Moreover, the substantial distributional discrepancies among different target domains often limit the effectiveness of alignment strategies. To address these problems, we propose an active dual alignment strategy that integrates active learning with all-way domain and conditional discriminative alignment to optimize cross-domain feature adaptation. Specifically, informative target samples are first selected for annotation via active learning. Then, conditional discrimination is employed to enhance category-aware feature alignment, while all-way domain discrimination ensures consistency across domains at the holistic feature level. Additionally, we incorporate a multilevel feature fusion module to further improve the model’s generalization capability. Extensive experiments on benchmark datasets such as Office-Home and Office-31 demonstrate that our approach significantly outperforms existing methods, particularly under limited annotation budgets, thereby validating the effectiveness of the proposed ADA strategy in multi-target domain adaptation settings. Notably, both the multistage active sampling and multi-domain parallel feature alignment in MT-ADA impose rigid computing power requirements on high-performance computing. Therefore, this paper conducts core verification using a high-performance standalone machine, and the proposed modular design demonstrates favorable scalability. Our code is available at https://github.com/JSJ515-Group/MT-ADA.