Scholarlink

<p>Cross-domain few-shot semantic segmentation faces significant challenges due to intra-class individual differences. This paper proposes a novel query self-similarity matching (QSM) framework that pays close attention to the query image itself. QSM introduces a foreground self-similarity module to capture representative foreground features, generating a more accurate foreground prototype through self-similarity learning. An adaptive prototype fusion module is then employed to flexibly fuse prototypes based on their contribution and global importance, resulting in a more robust query prototype. Finally, an adjacent channel attention module is designed to enhance cross-channel information exchange and highlight important query feature channels. Experiments on four datasets (Deepglobe, ISIC2018, Chest X-ray, and FSS-1000) demonstrate the effectiveness of QSM, achieving state-of-the-art performance with mIoU scores of 63.92% and 67.67% under 1-shot and 5-shot settings, respectively. The source code and datasets are publicly available at <a href="https://github.com/tch284994/QSM">https://github.com/tch284994/QSM</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

QSM: self-similarity guided query prototyping for robust cross-domain few-shot semantic segmentation

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Jianbing Yi,
  • Chenghua Tang,
  • Xin Wu,
  • Wenwu Xiong,
  • Xin Peng,
  • Yanzhen Chen

摘要

Cross-domain few-shot semantic segmentation faces significant challenges due to intra-class individual differences. This paper proposes a novel query self-similarity matching (QSM) framework that pays close attention to the query image itself. QSM introduces a foreground self-similarity module to capture representative foreground features, generating a more accurate foreground prototype through self-similarity learning. An adaptive prototype fusion module is then employed to flexibly fuse prototypes based on their contribution and global importance, resulting in a more robust query prototype. Finally, an adjacent channel attention module is designed to enhance cross-channel information exchange and highlight important query feature channels. Experiments on four datasets (Deepglobe, ISIC2018, Chest X-ray, and FSS-1000) demonstrate the effectiveness of QSM, achieving state-of-the-art performance with mIoU scores of 63.92% and 67.67% under 1-shot and 5-shot settings, respectively. The source code and datasets are publicly available at https://github.com/tch284994/QSM.