Scholarlink

<p>Few-shot semantic segmentation (FSS) aims to segment objects of unseen categories using only a handful of annotated samples. A central challenge lies in learning transferable representations that generalize across large intra-class variations and high inter-class similarity under extreme data scarcity. In this work, we introduce a novel hybrid token learning framework equipped with bidirectional attention to tackle these issues. Our approach first extracts adaptive tokens from support images, encoding both target-specific details and background context, and then integrates them with a set of learnable target-agnostic tokens to form a hybrid token representation. These tokens are refined through a Symbiotic Attention Refinement Module, which employs bidirectional masked cross-attention to enable co-adaptive optimization between tokens and query features. Experiments on standard natural image benchmarks, PASCAL-<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(5^{i}\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mn>5</mn> <mi>i</mi> </msup> </math></EquationSource> </InlineEquation> and COCO-<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(20^{i}\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mn>20</mn> <mi>i</mi> </msup> </math></EquationSource> </InlineEquation>, demonstrate that our method achieves competitive performance with only 4.5M parameters. The results confirm that our hybrid token representation effectively mitigates distribution bias and enhances generalization in few-shot segmentation.</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

Hybrid token learning with bidirectional attention for few-shot semantic segmentation

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Liting Lei,
  • Yujie Zhang,
  • Yadang Chen,
  • Jianlin Qiu

摘要

Few-shot semantic segmentation (FSS) aims to segment objects of unseen categories using only a handful of annotated samples. A central challenge lies in learning transferable representations that generalize across large intra-class variations and high inter-class similarity under extreme data scarcity. In this work, we introduce a novel hybrid token learning framework equipped with bidirectional attention to tackle these issues. Our approach first extracts adaptive tokens from support images, encoding both target-specific details and background context, and then integrates them with a set of learnable target-agnostic tokens to form a hybrid token representation. These tokens are refined through a Symbiotic Attention Refinement Module, which employs bidirectional masked cross-attention to enable co-adaptive optimization between tokens and query features. Experiments on standard natural image benchmarks, PASCAL- \(5^{i}\) 5 i and COCO- \(20^{i}\) 20 i , demonstrate that our method achieves competitive performance with only 4.5M parameters. The results confirm that our hybrid token representation effectively mitigates distribution bias and enhances generalization in few-shot segmentation.