Scholarlink

<p>Stereo matching, a pivotal task in computer vision, aims to extract depth information from images captured from different viewpoints. Despite its significance, traditional methods struggle with accuracy in regions with occlusions, texturelessness, and reflections. This paper introduces MambaStereo, a novel stereo matching network that leverages the Mamba model to construct a cost volume, enhancing correlation information between long-range pixels in ill-posed regions. To mitigate the limitations of Mamba’s local receptive field, we propose a channel and local enhancement module (CLEM) that integrates contextual information. Additionally, we replace some 3D convolutions in the aggregation network with partial convolutions, reducing computational cost while maintaining accuracy. A disparity refinement scheme based on feature error and dilation convolutions further enhances prediction accuracy. The experimental result shows MambaStereo significantly improves performance in ill-posed regions, with a 7.6% improvement in the px-2 error of all pixels compared to ACVNet on the KITTI2012 dataset and a 37.9% decrease in the number of parameters. This study highlights the potential of MambaStereo in advancing stereo matching technology for practical applications such as autonomous driving and 3D reconstruction. The code is available at <a href="https://github.com/linniao/MambaStereo">https://github.com/linniao/MambaStereo</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

MambaStereo: enhancing stereo matching accuracy in ill-posed regions via Mamba-based cost volume

  • 期刊论文

  • 发表日期：

    2025年12月23日

查看全文
  • Zhen Zhang,
  • Jinlong Liu,
  • Yaping Zhang

摘要

Stereo matching, a pivotal task in computer vision, aims to extract depth information from images captured from different viewpoints. Despite its significance, traditional methods struggle with accuracy in regions with occlusions, texturelessness, and reflections. This paper introduces MambaStereo, a novel stereo matching network that leverages the Mamba model to construct a cost volume, enhancing correlation information between long-range pixels in ill-posed regions. To mitigate the limitations of Mamba’s local receptive field, we propose a channel and local enhancement module (CLEM) that integrates contextual information. Additionally, we replace some 3D convolutions in the aggregation network with partial convolutions, reducing computational cost while maintaining accuracy. A disparity refinement scheme based on feature error and dilation convolutions further enhances prediction accuracy. The experimental result shows MambaStereo significantly improves performance in ill-posed regions, with a 7.6% improvement in the px-2 error of all pixels compared to ACVNet on the KITTI2012 dataset and a 37.9% decrease in the number of parameters. This study highlights the potential of MambaStereo in advancing stereo matching technology for practical applications such as autonomous driving and 3D reconstruction. The code is available at https://github.com/linniao/MambaStereo.