Scholarlink

<p>Self-supervised monocular depth estimation only uses image sequences for training, which has been widely studied. Recent advances predominantly rely on Convolutional Neural Networks (CNNs) or Vision Transformers (ViTs), which treat images as grid or sequence structures for feature representation. However, both architectures lack the flexibility to effectively capture irregular and complex object geometric features. To overcome this challenge, we propose a novel framework called ViGDepth with two effective contributions: (1) Introduce Vision Graph Neural Networks (ViGs) into depth estimation to sufficiently model global topological relationships through graph-structured image representation. (2) Propose a Local Window Detail Enhancement (LWDE) module that leverages node feature similarity to construct a dynamic adjacency matrix within image patches, then performs graph reasoning to preserve fine-grained local details. Extensive experiments on the KITTI dataset demonstrate that the proposed method significantly improves performance over the baseline, achieving a <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\delta _1\)</EquationSource> </InlineEquation> accuracy of 0.904. Furthermore, ViGDepth exhibits excellent generalization performance on the Make3D dataset and the WeatherKITTI dataset, even under adverse weather conditions.</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

Vigdepth: self-supervised monocular depth estimation based on vision GNN and local graph reasoning

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Xin Chen,
  • Mingwen Wang

摘要

Self-supervised monocular depth estimation only uses image sequences for training, which has been widely studied. Recent advances predominantly rely on Convolutional Neural Networks (CNNs) or Vision Transformers (ViTs), which treat images as grid or sequence structures for feature representation. However, both architectures lack the flexibility to effectively capture irregular and complex object geometric features. To overcome this challenge, we propose a novel framework called ViGDepth with two effective contributions: (1) Introduce Vision Graph Neural Networks (ViGs) into depth estimation to sufficiently model global topological relationships through graph-structured image representation. (2) Propose a Local Window Detail Enhancement (LWDE) module that leverages node feature similarity to construct a dynamic adjacency matrix within image patches, then performs graph reasoning to preserve fine-grained local details. Extensive experiments on the KITTI dataset demonstrate that the proposed method significantly improves performance over the baseline, achieving a \(\delta _1\) accuracy of 0.904. Furthermore, ViGDepth exhibits excellent generalization performance on the Make3D dataset and the WeatherKITTI dataset, even under adverse weather conditions.