Exploiting Sequence Analysis for Accurate Light-Field Depth Estimation

Depth estimation for light field (LF) images is the cornerstone of many applications cameras, such as 3D reconstruction, defects inspection, face liveness detection, and so forth. In recent years, convolutional neural network (CNN) has dominated primary workhorse depth estimation. However, interpretability accuracy results still need to be improved. This paper uses conditional random (CRF) theory explain model LF Further, from perspective sequence analysis, we extract features epipolar plane image (EPI) patches with recurrent (RNN) serve unary term energy function in CRF. Then, a unified (called LFRNN) designed solve CRF get map. Our LFRNN builds upon two-stage architecture, involving local refinement. first part, design an RNN analyze vector sequences EPI obtain values. There are two thinking behind this part. The general principle that slope straight line inversely proportional depth; second our unique observation those lines distributed sequences. continuous used optimize output We train on synthetic dataset test it both real-world datasets. Quantitative qualitative validate superior performance over state-of-the-art methods.