Scholarlink

<p>To overcome the identical distribution constraint in unsupervised domain adaptation (UDA), most existing methods focus on cross-domain transferability while ignoring intra-domain semantic structures, leading to poor feature discriminability. Class-conditional alignment strategies rely heavily on model predictions and suffer from confirmation bias caused by noisy pseudo-labels. To address these issues, we propose a novel <b>D</b>iscriminative <b>P</b>rototypical <b>C</b>lustering <b>T</b>ransport (DPCT) framework for UDA, which centers on prototypical clustering transport and is equipped with a mean-teacher module to ensure stable and high-quality learning. Specifically, the mean-teacher component generates reliable soft pseudo-labels for unlabeled target data via information entropy maximization and semantic consistency regularization, which effectively alleviates confirmation bias. Built on these high-quality pseudo-labels, we design a discriminative prototypical clustering transport loss that jointly optimizes intra-domain feature aggregation (compactness) and inter-domain feature alignment (class structure consistency), enhancing both transferability and discriminability simultaneously. Theoretically, we derive a generalization error bound to verify the robustness of DPCT. Extensive experiments on five standard visual UDA benchmarks (Office-31, Office-Home, ImageCLEF-DA, VisDA-2017, DomainNet) demonstrate that DPCT consistently outperforms state-of-the-art methods. Specifically, DPCT obtains 93.0% mean accuracy on Office-31 (+2.0% over the best baseline), 75.1% mean accuracy on Office-Home (+2.4% over PCT), 92.2% mean accuracy on ImageCLEF-DA (+3.2% over PCT), 87.6% per-category accuracy on VisDA-2017 (+0.5% over DANN+MNI+COT), and 56.9% inductive accuracy on DomainNet (+2.7% over the second-best method). DPCT also shows strong practicality in single/multi-source adaptation, imbalanced domain adaptation, and source-free domain adaptation.</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

Discriminative prototypical clustering transport with mean teacher for visual domain adaptation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月27日

查看全文
  • Yufang Dan,
  • Di Zhou

摘要

To overcome the identical distribution constraint in unsupervised domain adaptation (UDA), most existing methods focus on cross-domain transferability while ignoring intra-domain semantic structures, leading to poor feature discriminability. Class-conditional alignment strategies rely heavily on model predictions and suffer from confirmation bias caused by noisy pseudo-labels. To address these issues, we propose a novel Discriminative Prototypical Clustering Transport (DPCT) framework for UDA, which centers on prototypical clustering transport and is equipped with a mean-teacher module to ensure stable and high-quality learning. Specifically, the mean-teacher component generates reliable soft pseudo-labels for unlabeled target data via information entropy maximization and semantic consistency regularization, which effectively alleviates confirmation bias. Built on these high-quality pseudo-labels, we design a discriminative prototypical clustering transport loss that jointly optimizes intra-domain feature aggregation (compactness) and inter-domain feature alignment (class structure consistency), enhancing both transferability and discriminability simultaneously. Theoretically, we derive a generalization error bound to verify the robustness of DPCT. Extensive experiments on five standard visual UDA benchmarks (Office-31, Office-Home, ImageCLEF-DA, VisDA-2017, DomainNet) demonstrate that DPCT consistently outperforms state-of-the-art methods. Specifically, DPCT obtains 93.0% mean accuracy on Office-31 (+2.0% over the best baseline), 75.1% mean accuracy on Office-Home (+2.4% over PCT), 92.2% mean accuracy on ImageCLEF-DA (+3.2% over PCT), 87.6% per-category accuracy on VisDA-2017 (+0.5% over DANN+MNI+COT), and 56.9% inductive accuracy on DomainNet (+2.7% over the second-best method). DPCT also shows strong practicality in single/multi-source adaptation, imbalanced domain adaptation, and source-free domain adaptation.