Full-Viewpoint Depth Recovery of Light Field Image via Deep Neural Network

Recovery depth from a lenslet light field image can facilitate lots of applications including super-resolution and 3D reconstruction. However, current works mainly focus on the central sub-aperture image but pay little attention to full-viewpoint light field images. In this paper, we propose a deep learning-based method to recovery full-viewpoint depth by estimating the disparity map from the given light field image. We employ the ResNet to extract multi-dimensional features from the given light field image which is encoded as a 3D epipolar plane image, establish dense connections to enable the neural network to calculate the cost volume from extracted features, and use an AutoEncoder to convert the cost volume to a disparity map of the given light field image. We show several experimental results and two comparisons with the related works to demonstrate the effect and performance of our method.