Depth estimation based on supervised learning with high precision has been achieved nowadays.But this methods rely on real depth labels.As for the method of unsupervised learning

it is effective but inaccurate.The newest approach to solve the problem of depth estimation is domain adaptation technology.However

these methods ignore the semantic difference between real image and synthetic image when performing domain transformation.To solve this problem

a domain adaption framework based on semantic perception is proposed to estimate depth.To improve stability when performing domain transformation.In addition

smooth loss function is introduced to overcome problem of the situation of discontinuous boundary and gradient disappearance in the process depth estimation

which improves the quality of the depth images.In order to verify the effectiveness of the proposed method

experiments are carried out on commonly used datasets

and the experimental results prove the effectiveness of the method.