Scholarlink

<p>Cross-domain few-shot object detection (FSOD) faces significant challenges, including weak generalization ability (prone to false detection or missed detection), insufficient static prototype representation, and difficulties in detecting small or occluded objects. To address these issues, this paper proposes DICC-ViT, a novel framework leveraging Dynamic Information Coupling and Contrastive Learning built upon a shared-weight DINOv2 Vision Transformer backbone. First, we introduce a Region Proposal Re-screening (RPR) module that operates in the feature space, jointly evaluating Intersection-over-Union (IoU) and metric distances to category centers to correct semantic misclassifications inherent in traditional geometry-based RPNs. Second, a Dynamic Information Coupling (DIC) mechanism is designed to adaptively fuse support and query features via similarity-guided attention, generating task-aware dynamic prototypes that overcome the limitations of static averaging. Finally, a contrastive learning-based strategy is employed to iteratively optimize prototypes, explicitly enforcing intra-class compactness and inter-class separability. Extensive experiments across eight benchmark datasets, including PASCAL VOC, ArTaxOr, and DIOR, demonstrate that DICC-ViT significantly outperforms state-of-the-art methods such as DE-ViT, CD-ViTO and Meta-RCNN. Our approach achieves an average accuracy improvement of over 4% compared to the baseline, exhibiting superior robustness and generalization capabilities.</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

DICC-ViT: Improving cross-domain few-shot object detection via dynamic information coupling and contrastive learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月27日

查看全文
  • Man Wu,
  • Lili Wen,
  • Gaocai Wang

摘要

Cross-domain few-shot object detection (FSOD) faces significant challenges, including weak generalization ability (prone to false detection or missed detection), insufficient static prototype representation, and difficulties in detecting small or occluded objects. To address these issues, this paper proposes DICC-ViT, a novel framework leveraging Dynamic Information Coupling and Contrastive Learning built upon a shared-weight DINOv2 Vision Transformer backbone. First, we introduce a Region Proposal Re-screening (RPR) module that operates in the feature space, jointly evaluating Intersection-over-Union (IoU) and metric distances to category centers to correct semantic misclassifications inherent in traditional geometry-based RPNs. Second, a Dynamic Information Coupling (DIC) mechanism is designed to adaptively fuse support and query features via similarity-guided attention, generating task-aware dynamic prototypes that overcome the limitations of static averaging. Finally, a contrastive learning-based strategy is employed to iteratively optimize prototypes, explicitly enforcing intra-class compactness and inter-class separability. Extensive experiments across eight benchmark datasets, including PASCAL VOC, ArTaxOr, and DIOR, demonstrate that DICC-ViT significantly outperforms state-of-the-art methods such as DE-ViT, CD-ViTO and Meta-RCNN. Our approach achieves an average accuracy improvement of over 4% compared to the baseline, exhibiting superior robustness and generalization capabilities.