Scholarlink

<p>Unsupervised domain adaptation (UDA) is an important task that transfers learned knowledge from the source domain to the unseen target domain. Domain shift is the major challenge faced by UDA methods for aerial data, which is caused by differences in appearance, distribution, decision boundary, sensor platforms, capturing conditions, etc. Mix representations combine inputs of both domains, which are helpful for the generation of domain-consistent features using self-training. Domain alignment at multiple levels is required for effective adaptive segmentation between domains. In this work, we proposed a Masked Domain Adversarial Adaptation Framework (MDAAF), which consists of the Masked Domain Adversarial Adaptation Network (MDAANet), Masked Domain Dual Adaptation (MDDA) approach, Joint Adversarial Alignment (JAA), Consistency Enforcement (CE), and Feature Dissimilarity-based Alignment (FDA) for effective UDA. The proposed MDAAF achieves the all-level adaptation (input, feature, and output) to handle domain shift issues using multi-level features generated by MDAANet with MDDA. FDA increased the inter-class feature dissimilarities, CE attained input-consistent output adaptation, and JAA achieved combined adaptation at the input–output level. MDAAF achieved state-of-the-art results on six benchmark domain adaptation tasks with performance improvement in the <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10044_2025_1473_Article_IEq1.gif" Format="GIF" Height="6" Rendition="HTML" Resolution="72" Type="Linedraw" Width="17" /> </InlineMediaObject> <EquationSource Format="TEX">\(\sim \)</EquationSource> </InlineEquation> 1.5–10% range from previous methods. The trained model weights and implementation code are available at <a href="https://github.com/chouhan-avinash/MDAAF/tree/master/">https://github.com/chouhan-avinash/MDAAF/tree/master/</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

MDAAF: masked domain adversarial adaptation framework for unsupervised domain adaptive semantic segmentation

  • 期刊论文

  • 发表日期：

    2025年5月13日

查看全文
  • Avinash Chouhan,
  • Arijit Sur,
  • Dibyajyoti Chutia,
  • Shiv Prasad Aggarwal

摘要

Unsupervised domain adaptation (UDA) is an important task that transfers learned knowledge from the source domain to the unseen target domain. Domain shift is the major challenge faced by UDA methods for aerial data, which is caused by differences in appearance, distribution, decision boundary, sensor platforms, capturing conditions, etc. Mix representations combine inputs of both domains, which are helpful for the generation of domain-consistent features using self-training. Domain alignment at multiple levels is required for effective adaptive segmentation between domains. In this work, we proposed a Masked Domain Adversarial Adaptation Framework (MDAAF), which consists of the Masked Domain Adversarial Adaptation Network (MDAANet), Masked Domain Dual Adaptation (MDDA) approach, Joint Adversarial Alignment (JAA), Consistency Enforcement (CE), and Feature Dissimilarity-based Alignment (FDA) for effective UDA. The proposed MDAAF achieves the all-level adaptation (input, feature, and output) to handle domain shift issues using multi-level features generated by MDAANet with MDDA. FDA increased the inter-class feature dissimilarities, CE attained input-consistent output adaptation, and JAA achieved combined adaptation at the input–output level. MDAAF achieved state-of-the-art results on six benchmark domain adaptation tasks with performance improvement in the \(\sim \) 1.5–10% range from previous methods. The trained model weights and implementation code are available at https://github.com/chouhan-avinash/MDAAF/tree/master/.