Unsupervised stereo image depth estimation of midan recovery structure

Abstract

The stereo image depth estimation is an important fileld in computer vision. At present, most of the researches on depth estimation based on deep learning make use of the mathematical characteristics of the image, such as left-right consistency and pixel distribution characteristics. In this paper, we pay attention to the relationship between the grid representation, which is a distance representation mode of biological system, and the depth representation. Inspired by the concept of Fourier decomposition, we propose a median recovery network structure. The Experiment shows that this network structure has a certain effect on depth estimation.