Scholarlink

We propose a Frequency Filtering and Domain Adaptation Network for Cross-Domain Few-Shot Object Detection(CD-FSOD). Existing approaches mainly emphasize spatial feature alignment, yet they struggle to mitigate domain shifts caused by variations in texture style, illumination, and noise—differences that are often more evident in the frequency domain. To address this limitation, we introduce a frequency-centric framework comprising three key components: a Frequency Filter Module (FFM), an Amplitude-Phase Attention(APA) and a Dual-branch Domain Adapter (DDA). The FFM performs frequency decomposition and adaptive filtering at the feature level, suppressing domain-specific high-frequency noise and low-frequency distribution shifts to obtain more domain-invariant representations. The DDA conducts global-local domain alignment through parallel branches, facilitating structural reorganization of features and further enhancing the model’s cross-domain generalization. Experimental results demonstrate that our model outperforms previous state-of-the-art methods, achieving average mAP improvements of 1.7%, 1.9%, and 1.8% under 1/5/10-shot settings respectively.

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

Frequency Filtering and Domain Adaptation for Cross-Domain Few-Shot Object Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xuwei Cheng,
  • Hong Shi,
  • Mingxing Hou,
  • Xiaofang Mu

摘要

We propose a Frequency Filtering and Domain Adaptation Network for Cross-Domain Few-Shot Object Detection(CD-FSOD). Existing approaches mainly emphasize spatial feature alignment, yet they struggle to mitigate domain shifts caused by variations in texture style, illumination, and noise—differences that are often more evident in the frequency domain. To address this limitation, we introduce a frequency-centric framework comprising three key components: a Frequency Filter Module (FFM), an Amplitude-Phase Attention(APA) and a Dual-branch Domain Adapter (DDA). The FFM performs frequency decomposition and adaptive filtering at the feature level, suppressing domain-specific high-frequency noise and low-frequency distribution shifts to obtain more domain-invariant representations. The DDA conducts global-local domain alignment through parallel branches, facilitating structural reorganization of features and further enhancing the model’s cross-domain generalization. Experimental results demonstrate that our model outperforms previous state-of-the-art methods, achieving average mAP improvements of 1.7%, 1.9%, and 1.8% under 1/5/10-shot settings respectively.