Scholarlink

<p>Stereo matching plays a fundamental role in 3D perception; however, most existing methods mainly rely on spatial-domain features and often suffer from blurred boundaries and insufficient structural representation. To address these issues, we propose WTSNet, a novel stereo matching network that integrates wavelet transform and superpixel segmentation to enhance feature representation. Specifically, a Wavelet Transform-based attention Module (WTM) is designed to capture multi-frequency contextual information, while a Superpixel Segmentation Module (SSM) preserves local structural details for more accurate disparity estimation near object boundaries. By jointly exploiting frequency-domain cues and structure-aware priors, WTSNet effectively models both global scene geometry and fine-grained local features. Experimental results show that WTSNet achieves End-Point Errors (EPE) of 0.45 pixels on KITTI 2012 and 0.73 pixels on KITTI 2015 after 300 training epochs, outperforming representative methods such as TANet under the same setting. These results demonstrate the effectiveness and competitiveness of the proposed method on challenging stereo benchmarks.</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

WTSNet: An accurate stereo matching network based on wavelet transform and superpixel segmentation

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Haiming Qu,
  • Yunhui Luo,
  • Guiling Hu,
  • Chongbao Zhao,
  • Minxuan He,
  • Mingyu Shang,
  • Qianqian Xu

摘要

Stereo matching plays a fundamental role in 3D perception; however, most existing methods mainly rely on spatial-domain features and often suffer from blurred boundaries and insufficient structural representation. To address these issues, we propose WTSNet, a novel stereo matching network that integrates wavelet transform and superpixel segmentation to enhance feature representation. Specifically, a Wavelet Transform-based attention Module (WTM) is designed to capture multi-frequency contextual information, while a Superpixel Segmentation Module (SSM) preserves local structural details for more accurate disparity estimation near object boundaries. By jointly exploiting frequency-domain cues and structure-aware priors, WTSNet effectively models both global scene geometry and fine-grained local features. Experimental results show that WTSNet achieves End-Point Errors (EPE) of 0.45 pixels on KITTI 2012 and 0.73 pixels on KITTI 2015 after 300 training epochs, outperforming representative methods such as TANet under the same setting. These results demonstrate the effectiveness and competitiveness of the proposed method on challenging stereo benchmarks.