Scholarlink

<p>Self-supervised monocular depth estimation often suffers from degraded performance in challenging scenarios such as low-texture regions, weak illumination, overexposure, and dynamic objects. In this paper, we propose an enhancement method that does not modify the network architecture. Instead, we introduce a novel direction mask mechanism that guides the model to focus on reliable regions by masking out areas with inaccurate projection based on the discrepancy between optical flow and reconstructed optical flow. This strategy reduces the influence of erroneous supervision and enables the network to learn more effectively during optimization. Building upon this, we adopt a teacher–student distillation framework, where a teacher network pre-trained with DepthAnything provides high-quality pseudo-labels to a student network. This transforms the self-supervised task into a supervised one, significantly improving depth estimation accuracy. Experimental results demonstrate that our method achieves lower errors compared to traditional self-supervised photometric loss-based approaches. On the KITTI benchmark, our method outperforms the current state-of-the-art method ManyDepth2 by 10.5%, 6.9%, and 2.9% in SqRel, RMSE, and RMSElog metrics, respectively. On Cityscapes, our method reduces AbsRel by 32.8%, SqRel by 56.2%, RMSE by 25.0% and RMSElog by 26.4% relative to ManyDepth2. We achieve state-of-the-art performance on the Cityscapes dataset, confirming the robustness of our approach in dynamic environments. Code and reproduction instructions are available at: <a href="https://github.com/1225735136/ORDMP">https://github.com/1225735136/ORDMP</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

ORDMP: self-supervised monocular depth estimation via optical-flow-reconstructed directional masks and large-model teacher pseudo-labels

  • 期刊论文

  • 发表日期：

    2025年12月3日

查看全文
  • Shuwen Jia,
  • Yongxiong Wang,
  • Han Chen,
  • Shuai Huang

摘要

Self-supervised monocular depth estimation often suffers from degraded performance in challenging scenarios such as low-texture regions, weak illumination, overexposure, and dynamic objects. In this paper, we propose an enhancement method that does not modify the network architecture. Instead, we introduce a novel direction mask mechanism that guides the model to focus on reliable regions by masking out areas with inaccurate projection based on the discrepancy between optical flow and reconstructed optical flow. This strategy reduces the influence of erroneous supervision and enables the network to learn more effectively during optimization. Building upon this, we adopt a teacher–student distillation framework, where a teacher network pre-trained with DepthAnything provides high-quality pseudo-labels to a student network. This transforms the self-supervised task into a supervised one, significantly improving depth estimation accuracy. Experimental results demonstrate that our method achieves lower errors compared to traditional self-supervised photometric loss-based approaches. On the KITTI benchmark, our method outperforms the current state-of-the-art method ManyDepth2 by 10.5%, 6.9%, and 2.9% in SqRel, RMSE, and RMSElog metrics, respectively. On Cityscapes, our method reduces AbsRel by 32.8%, SqRel by 56.2%, RMSE by 25.0% and RMSElog by 26.4% relative to ManyDepth2. We achieve state-of-the-art performance on the Cityscapes dataset, confirming the robustness of our approach in dynamic environments. Code and reproduction instructions are available at: https://github.com/1225735136/ORDMP.