Deep Learning Based Depth Estimation and Reconstruction of Light Field Images

Abstract

Light field imaging is one of the most promising methods to capture realistic 3D scenes. In this paper, we propose a deep learning network composed of two sub-networks performing depth estimation and light field image reconstruction, respectively. We simultaneously train the two sub-networks by employing a loss function combining the reconstruction loss of the reconstruction network and the estimation loss of the depth estimation network. Experimental results demonstrate that the proposed method accurately estimates the disparity maps of light field images and also faithfully reconstructs light field images.