基于多尺度边缘概率指导的单目绝对深度估计算法

打开文本图片集

中图分类号：TP391 文献标识码：A

Abstract： Conventional monocular metric depth estimation methods often exhibit limited accuracy in boundary regions，leading to blurred object edges and structural distortions. To address this issue，a novel monocular metric depth estimation model based on Multi-scale Edge Probability Guidance （MEPG） is proposed. Specifically，the model incorporates a multi-scale edge detection module that extracts and integrates edge features across different scales in a probabilistic manner. This design enhances the network’s sensitivity and representation capability for boundary regions. Furthermore，an absolute depth estimation module is designed to transform relative depth predictions into metric depth values with physical scale. The MEPG model is evaluated on the KITTI and NYU Depth V2 datasets. Experimental results demonstrate that it achieves reductions of 6.3% and 3.2% respectively，in the AbsRel metric compared with the baseline model， validating the effectiveness and robustness of the proposed approach.

Keywords：monocular metric depth estimation； edge guidance；multi-scale representation； com-puter vision; deep learning

单目深度估计任务分为单目绝对深度估计（Monocular Metric Depth Estimation，MMDE）和单目相对深度估计（Monocular Relative Depth Estimation，MRDE），绝对深度估计方法直接估计物体绝对物理单位的深度，在计算机视觉和机器人技术的许多下游应用中具有实用价值;相对深度估计则估计每个像素与其它像素的相对深度差异，无尺度信息，可以满足各种类型环境，具有更好的泛化能力。（剩余8020字）