Scholarlink

<p>The key challenge in Semi-Supervised Domain Adaptation (SSDA) lies in transferring knowledge from a source domain with abundant labeled data to a target domain characterized by scarce labeled and plentiful unlabeled samples. While distribution shifts already complicate this task, the challenge increases substantially when source and target data come from different sensing modalities, such as RGB and thermal images, resulting in a severe form of domain shift commonly referred to as Semi-Supervised Heterogeneous Domain Adaptation (SSHDA). Standard SSHDA approaches typically rely on pre-trained feature extractors for each modality followed by trainable projectors that align representations in a shared latent space, leading to a two-step pipeline. However, the requirement of a dedicated pre-trained encoder for every modality hinders their applicability, particularly in cross-sensor scenarios where such models may not exist or may fail to capture sensor-specific characteristics. To address these limitations, we propose <i>HEADS</i> (Heterogeneous End-to-end Adversarial Domain adaptation in a Semi-supervised manner), an end-to-end neural framework equipped with fully trainable modality-specific branches that jointly implements feature extraction and knowledge transfer. The model promotes the separation of domain-invariant features, crucial for the downstream task, from domain-specific components that impede cross-modality adaptation. Furthermore, we enhance feature invariance through an adversarial learning objective grounded in the Wasserstein distance. We validate our framework across three multi-modal benchmarks: SUN RGB-D, TRISTAR, and HANDS. Extensive experiments show that <i>HEADS</i> consistently outperforms both standard baselines and state-of-the-art SSHDA approaches, especially under minimal target supervision and across diverse backbone architectures. Our code is available at <a href="https://github.com/giu-guarino/HEADS">https://github.com/giu-guarino/HEADS</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

HEADS: An End-to-End Adversarial Framework for Heterogeneous Semi-Supervised Domain Adaptation

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Giuseppe Guarino,
  • Cássio F. Dantas,
  • Dino Ienco,
  • Raffaele Gaetano,
  • Gemine Vivone,
  • Matteo Ciotola,
  • Giuseppe Scarpa

摘要

The key challenge in Semi-Supervised Domain Adaptation (SSDA) lies in transferring knowledge from a source domain with abundant labeled data to a target domain characterized by scarce labeled and plentiful unlabeled samples. While distribution shifts already complicate this task, the challenge increases substantially when source and target data come from different sensing modalities, such as RGB and thermal images, resulting in a severe form of domain shift commonly referred to as Semi-Supervised Heterogeneous Domain Adaptation (SSHDA). Standard SSHDA approaches typically rely on pre-trained feature extractors for each modality followed by trainable projectors that align representations in a shared latent space, leading to a two-step pipeline. However, the requirement of a dedicated pre-trained encoder for every modality hinders their applicability, particularly in cross-sensor scenarios where such models may not exist or may fail to capture sensor-specific characteristics. To address these limitations, we propose HEADS (Heterogeneous End-to-end Adversarial Domain adaptation in a Semi-supervised manner), an end-to-end neural framework equipped with fully trainable modality-specific branches that jointly implements feature extraction and knowledge transfer. The model promotes the separation of domain-invariant features, crucial for the downstream task, from domain-specific components that impede cross-modality adaptation. Furthermore, we enhance feature invariance through an adversarial learning objective grounded in the Wasserstein distance. We validate our framework across three multi-modal benchmarks: SUN RGB-D, TRISTAR, and HANDS. Extensive experiments show that HEADS consistently outperforms both standard baselines and state-of-the-art SSHDA approaches, especially under minimal target supervision and across diverse backbone architectures. Our code is available at https://github.com/giu-guarino/HEADS.