Scholarlink

<p>Although self-supervised monocular depth estimation methods have made significant progress in computer vision, current real-time depth estimation methods based on visual features exhibit low accuracy in reconstructing fine structures in complex scenes. This paper introduces CT-Mono, a self-supervised monocular depth estimation framework that integrates convolutional neural networks and transformers. It enhances local representation through a multi-scale feature fusion mechanism and constructs a cross-region attention network to capture global contextual associations, thereby facilitating the accurate reconstruction of fine structures in complex scenes. First, leveraging the Depthwise Separable Dilated Convolution Block (DCB), multi-layer dilated convolution expands the receptive field, while the fusion of multi-scale pixel features enhances local details in key regions. Second, Enhanced Local–Global Features Fusion is employed alongside local convolution to capture fine-grained local image information, while the self-attention mechanism captures long-range dependencies. This synergistic modeling of local and global information significantly enhances feature representation. Finally, the reprojection loss is formulated by estimating the six-degree-of-freedom relative transformation between neighboring images using a lightweight encoder based on the pre-trained ResNet18, thereby optimizing depth prediction. Experiments on the KITTI dataset demonstrate that CT-Mono achieves an AbsRel of 0.102 and an RMSE of 4.430, significantly outperforming existing mainstream methods.</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

CT-Mono: Leveraging CNNs and Transformers for Self-Supervised Depth Estimation in Single-View Scenarios

  • 期刊论文

  • 发表日期：

    2025年9月9日

查看全文
  • Jiacheng Zhang,
  • Xichao Wang,
  • Xinyao Wang,
  • ChenXuan Zhu,
  • Liang Zhu

摘要

Although self-supervised monocular depth estimation methods have made significant progress in computer vision, current real-time depth estimation methods based on visual features exhibit low accuracy in reconstructing fine structures in complex scenes. This paper introduces CT-Mono, a self-supervised monocular depth estimation framework that integrates convolutional neural networks and transformers. It enhances local representation through a multi-scale feature fusion mechanism and constructs a cross-region attention network to capture global contextual associations, thereby facilitating the accurate reconstruction of fine structures in complex scenes. First, leveraging the Depthwise Separable Dilated Convolution Block (DCB), multi-layer dilated convolution expands the receptive field, while the fusion of multi-scale pixel features enhances local details in key regions. Second, Enhanced Local–Global Features Fusion is employed alongside local convolution to capture fine-grained local image information, while the self-attention mechanism captures long-range dependencies. This synergistic modeling of local and global information significantly enhances feature representation. Finally, the reprojection loss is formulated by estimating the six-degree-of-freedom relative transformation between neighboring images using a lightweight encoder based on the pre-trained ResNet18, thereby optimizing depth prediction. Experiments on the KITTI dataset demonstrate that CT-Mono achieves an AbsRel of 0.102 and an RMSE of 4.430, significantly outperforming existing mainstream methods.