Scholarlink

<p>Depth estimation is a critical component of computer vision, with self-supervised monocular depth estimation drawing significant attention due to its lack of reliance on ground truth depth data. To meet real-time application demands, designing lightweight and efficient architectures for resource-constrained devices is essential. This paper introduces a novel network called LEDepth, which incorporates a Dilated Convolution Transformer (DCT) module and a Cross-Attention Feature Interaction (CAFI) module. The DCT module efficiently captures multi-scale features through dilated convolution, while the CAFI module integrates local and global feature information via a cross-attention mechanism. This design enables LEDepth to enhance depth estimation accuracy and inference speed while reducing the number of parameters. Experiments on the KITTI dataset demonstrate that LEDepth is only 3.1 MB in size, with approximately 80% fewer parameters compared to Monodepth2, while achieving significantly higher accuracy. LEDepth also exhibits strong generalization performance on the Make3D dataset.</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

LEDepth: A Lightweight Self-Supervised Monocular Depth Estimation Network Combining CNN and Transformer

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Lei Zhang,
  • Chengzhi Lyu,
  • Pan Wang,
  • Hong He,
  • Fang Dai

摘要

Depth estimation is a critical component of computer vision, with self-supervised monocular depth estimation drawing significant attention due to its lack of reliance on ground truth depth data. To meet real-time application demands, designing lightweight and efficient architectures for resource-constrained devices is essential. This paper introduces a novel network called LEDepth, which incorporates a Dilated Convolution Transformer (DCT) module and a Cross-Attention Feature Interaction (CAFI) module. The DCT module efficiently captures multi-scale features through dilated convolution, while the CAFI module integrates local and global feature information via a cross-attention mechanism. This design enables LEDepth to enhance depth estimation accuracy and inference speed while reducing the number of parameters. Experiments on the KITTI dataset demonstrate that LEDepth is only 3.1 MB in size, with approximately 80% fewer parameters compared to Monodepth2, while achieving significantly higher accuracy. LEDepth also exhibits strong generalization performance on the Make3D dataset.