Scholarlink

<p>In this paper, we introduce GPT-ZSS, a unified zero-shot segmentation framework capable of performing instance, semantic, and panoptic segmentation for unseen categories without requiring any training samples. By leveraging GPT-generated fine-grained semantic embeddings, our approach achieves more nuanced semantic features. A feature fusion generator integrates these embeddings with hand-crafted features, thereby enhancing semantic understanding. Furthermore, we propose a multi-scale channel attention relationship alignment (MSCA-RA) strategy, which aligns inter-class relationships across visual and semantic spaces, significantly improving segmentation accuracy for unseen categories. Extensive experiments on benchmark datasets demonstrate that GPT-ZSS surpasses state-of-the-art results in zero-shot instance and semantic segmentation tasks while achieving promising performance in panoptic segmentation. The code is available at <a href="https://github.com/jjuh/GPT-ZSS.git">https://github.com/jjuh/GPT-ZSS.git</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

GPT-ZSS: a unified zero-shot segmentation framework leveraging GPT-generated semantic embeddings and relationship alignment

  • 期刊论文

  • 发表日期：

    2025年4月10日

查看全文
  • Dan Xu,
  • Li Ding,
  • Yang Zhou,
  • Qiang Qian,
  • Jinlong Shi

摘要

In this paper, we introduce GPT-ZSS, a unified zero-shot segmentation framework capable of performing instance, semantic, and panoptic segmentation for unseen categories without requiring any training samples. By leveraging GPT-generated fine-grained semantic embeddings, our approach achieves more nuanced semantic features. A feature fusion generator integrates these embeddings with hand-crafted features, thereby enhancing semantic understanding. Furthermore, we propose a multi-scale channel attention relationship alignment (MSCA-RA) strategy, which aligns inter-class relationships across visual and semantic spaces, significantly improving segmentation accuracy for unseen categories. Extensive experiments on benchmark datasets demonstrate that GPT-ZSS surpasses state-of-the-art results in zero-shot instance and semantic segmentation tasks while achieving promising performance in panoptic segmentation. The code is available at https://github.com/jjuh/GPT-ZSS.git.