Scholarlink

<p>Few-shot semantic segmentation aims to address the dependence of traditional models on large-scale datasets, hoping to learn a model that can accurately segment unseen classes in a query image using only a few annotated samples (support images). Existing methods still face challenges in handling intra-class variations and in utilizing information. To alleviate these issues, we propose the Prototype Reorganization Network (PRNet). It includes innovative designs in three aspects: (1) A simple and efficient network framework that not only retains generalization ability but also improves the utilization of high-level information through a novel prior generation method. (2) A novel Prototype Reconstruction Transformer that overcomes the information loss issue inherent in traditional prototype generation methods. (3) A Learnable Prototype Fusion (LPF) strategy that can fully utilize category information and effectively mitigate the interference of intra-class variations and noise. Extensive experiments demonstrate PRNet’s superior performance on benchmark datasets such as PASCAL-<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(5^i\)</EquationSource> </InlineEquation> and COCO-<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(20^i\)</EquationSource> </InlineEquation>. Significantly, our model achieves mIoU scores of 69.6% and 74.4% on the PASCAL-<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(5^i\)</EquationSource> </InlineEquation> dataset for the 1-shot and 5-shot settings, respectively, surpassing the state-of-the-art method by 0.2% and 2.6%. The code is available at <a href="https://github.com/shixiaolong-522/PRNet">https://github.com/shixiaolong-522/PRNet</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

PRNet: prototype reorganization few-shot semantic segmentation network

  • 期刊论文

  • 发表日期：

    2026年2月27日

查看全文
  • Shaojun Qu,
  • Xiaolong Shi,
  • Donglin Xie

摘要

Few-shot semantic segmentation aims to address the dependence of traditional models on large-scale datasets, hoping to learn a model that can accurately segment unseen classes in a query image using only a few annotated samples (support images). Existing methods still face challenges in handling intra-class variations and in utilizing information. To alleviate these issues, we propose the Prototype Reorganization Network (PRNet). It includes innovative designs in three aspects: (1) A simple and efficient network framework that not only retains generalization ability but also improves the utilization of high-level information through a novel prior generation method. (2) A novel Prototype Reconstruction Transformer that overcomes the information loss issue inherent in traditional prototype generation methods. (3) A Learnable Prototype Fusion (LPF) strategy that can fully utilize category information and effectively mitigate the interference of intra-class variations and noise. Extensive experiments demonstrate PRNet’s superior performance on benchmark datasets such as PASCAL- \(5^i\) and COCO- \(20^i\) . Significantly, our model achieves mIoU scores of 69.6% and 74.4% on the PASCAL- \(5^i\) dataset for the 1-shot and 5-shot settings, respectively, surpassing the state-of-the-art method by 0.2% and 2.6%. The code is available at https://github.com/shixiaolong-522/PRNet.