Scholarlink

<p>Camouflaged object detection (COD) remains a challenging task due to the low visual saliency and ambiguous boundaries of targets concealed within complex backgrounds. While recent fully supervised methods have achieved impressive performance, they heavily rely on expensive pixel-wise annotations, which are impractical in large-scale or domain-specific applications. To alleviate this issue, we propose a novel weakly supervised COD framework named Simple Labels, Rich Supervision (SLRS), which leverages a self-adaptive label expansion (SALE) strategy to transform sparse scribble annotations into high-quality pseudo-labels. Specifically, we design a content-aware expansion mechanism that adaptively adjusts the expansion region based on local density, edge complexity, and contextual semantic information, effectively mitigating the inherent sparsity and incompleteness of weak annotations. Additionally, an edge-aware cross-scale attention fusion (ECAF) module is introduced to enhance the model’s capability in capturing fine-grained edge details and multi-scale contextual dependencies. Extensive experiments conducted on three challenging COD benchmarks, including COD10K, CAMO, and CHAMELEON, demonstrate that our method significantly outperforms existing state-of-the-art weakly supervised approaches and achieves performance competitive with several fully supervised models. The proposed framework offers a promising solution for practical COD applications by reducing annotation costs while maintaining strong detection performance. The source code and datasets are publicly available at: <a href="https://github.com/mypaocaiyu/SLRS.">https://github.com/mypaocaiyu/SLRS.</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

Simple labels, rich supervision: a self-adaptive label expansion based weakly supervised camouflaged object detection

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Chen Wang,
  • Na Ma,
  • Xiaohui Li,
  • Xu Jia

摘要

Camouflaged object detection (COD) remains a challenging task due to the low visual saliency and ambiguous boundaries of targets concealed within complex backgrounds. While recent fully supervised methods have achieved impressive performance, they heavily rely on expensive pixel-wise annotations, which are impractical in large-scale or domain-specific applications. To alleviate this issue, we propose a novel weakly supervised COD framework named Simple Labels, Rich Supervision (SLRS), which leverages a self-adaptive label expansion (SALE) strategy to transform sparse scribble annotations into high-quality pseudo-labels. Specifically, we design a content-aware expansion mechanism that adaptively adjusts the expansion region based on local density, edge complexity, and contextual semantic information, effectively mitigating the inherent sparsity and incompleteness of weak annotations. Additionally, an edge-aware cross-scale attention fusion (ECAF) module is introduced to enhance the model’s capability in capturing fine-grained edge details and multi-scale contextual dependencies. Extensive experiments conducted on three challenging COD benchmarks, including COD10K, CAMO, and CHAMELEON, demonstrate that our method significantly outperforms existing state-of-the-art weakly supervised approaches and achieves performance competitive with several fully supervised models. The proposed framework offers a promising solution for practical COD applications by reducing annotation costs while maintaining strong detection performance. The source code and datasets are publicly available at: https://github.com/mypaocaiyu/SLRS.