Scholarlink

<p>Monocular depth estimation faces a core challenge: balancing the physical plausibility of depth distribution with the preservation of fine details. To address this, we propose MIDNet, a novel encoder–dual-decoder architecture leveraging depth discretisation and multi-scale feature interaction. The encoder employs a hierarchical dynamic windowing attention mechanism to extract both local textures and global structures; Secondly, the dual decoders separately reconstruct geometric details via deformable convolution and generate adaptive depth queries features via cross-scale attention mechanism; Finally, the Adaptive Interval Module (AIM) fuses multi-level features and decoder outputs to generate depth interval distributions that conform to physical constraints. Experiments on NYU Depth V2 and KITTI show that our method improves AbsRel by 10.2% and 16.0% over the current best, and exhibits strong generalisation ability, verifying its feasibility for monocular depth estimation.</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

Midnet: multi-scale interaction guided adaptive discretization network for monocular depth estimation

  • 期刊论文

  • 发表日期：

    2026年3月21日

查看全文
  • Long Yan,
  • Qiuyue Li,
  • Zhiyao Li,
  • Fuyang Yu

摘要

Monocular depth estimation faces a core challenge: balancing the physical plausibility of depth distribution with the preservation of fine details. To address this, we propose MIDNet, a novel encoder–dual-decoder architecture leveraging depth discretisation and multi-scale feature interaction. The encoder employs a hierarchical dynamic windowing attention mechanism to extract both local textures and global structures; Secondly, the dual decoders separately reconstruct geometric details via deformable convolution and generate adaptive depth queries features via cross-scale attention mechanism; Finally, the Adaptive Interval Module (AIM) fuses multi-level features and decoder outputs to generate depth interval distributions that conform to physical constraints. Experiments on NYU Depth V2 and KITTI show that our method improves AbsRel by 10.2% and 16.0% over the current best, and exhibits strong generalisation ability, verifying its feasibility for monocular depth estimation.