Light-Field Depth Estimation Using RNN and CRF

Convolutional Neural Networks (CNNs) have recently been successfully applied to depth estimation from light field. Different from those CNN-based methods, we utilize the sequence characteristics of Epipolar Plane Images (EPIs) and introduce a novel light-field depth estimation method based on the Recurrent Neural Network (RNN). Our network builds upon two-stages architectures, involving a local depth estimation and a depth refinement part. In the first part, we regard an EPI patch as a vector sequence which is fed into the RNN to obtain a local depth value. Then, guided by the theory of Conditional Random Field (CRF), we globally optimize the depth map in the second part. Our network was trained in the disparity truth values provided by the synthetic light-field dataset. Experimental results show that our method allows to estimate high-quality disparity results.