Abstract: For the problem of inaccurate prediction of depth information caused by blurred,low-contrast and color distortion of complex weather scene images,previous studies have used the depth map of a standard scene as the a priori information for depth estimation of such scenes.However,this approach suffers from problems such as low accuracy of a priori information.This paper proposed a monocular depth estimation model TalentDepth based on a multiscale attention mechanism to realize the prediction of complex weather scenes.First,the multiscale attention mechanism was fused in the encoder to reduce the computational cost while retaining the information of each channel to improve the efficiency and capability of feature extraction.Second,to address the problem of unclear image depth,based on geometric consistency,a Depth Region Refinement(DSR) module was proposed to filter inaccurate pixel points in order to improve the reliability of depth information.Finally,the complex samples generated by the image translation model are input and the standard loss on the corresponding original images is calculated to guide the self-supervised training of the model.On the three datasets,NuScence,KITTI and KITTI-C,the error and accuracy are optimized compared to the baseline model.

Key words: Monocular depth estimation, Self-supervised learning, Multiscale attention, Knowledge distillation, Deep learning