Scholarlink

<p>Mirror detection remains a significant challenge in computer vision due to misleading reflections and the lack of distinctive visual cues. While the segment anything model (SAM) demonstrates strong segmentation capabilities, its performance on mirrors is limited by its reliance on RGB features alone. We introduce SAMirror, a novel framework that integrates rich visual information from SAM with crucial depth cues for robust mirror detection. Our core innovation lies in the depth-infusion adapter, which injects and contextualizes auxiliary scene context. Extensive experiments on mirror detection datasets demonstrate that SAMirror outperforms state-of-the-art methods, achieving an IoU of 0.902 and an F-measure of 0.930 on the MSD dataset, showcasing superior performance in both accuracy and generalization. By effectively leveraging depth information, SAMirror advances the field of mirror detection, offering new possibilities for applications in robotics, autonomous navigation, and beyond..</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

SAMirror: enhancing mirror detection via integrated visual-depth cues in segment anything model

  • 期刊论文

  • 发表日期：

    2025年9月16日

查看全文
  • Qiushi Meng,
  • Yunxun Liu,
  • Ran Hu,
  • Mengmeng Liang,
  • Jixue Yan,
  • Lei Zhu

摘要

Mirror detection remains a significant challenge in computer vision due to misleading reflections and the lack of distinctive visual cues. While the segment anything model (SAM) demonstrates strong segmentation capabilities, its performance on mirrors is limited by its reliance on RGB features alone. We introduce SAMirror, a novel framework that integrates rich visual information from SAM with crucial depth cues for robust mirror detection. Our core innovation lies in the depth-infusion adapter, which injects and contextualizes auxiliary scene context. Extensive experiments on mirror detection datasets demonstrate that SAMirror outperforms state-of-the-art methods, achieving an IoU of 0.902 and an F-measure of 0.930 on the MSD dataset, showcasing superior performance in both accuracy and generalization. By effectively leveraging depth information, SAMirror advances the field of mirror detection, offering new possibilities for applications in robotics, autonomous navigation, and beyond..