The estimated depth map provides valuable information in many computer vision applications such as autonomous driving, semantic segmentation and 3D object reconstruction. Since the light field camera capture both the spatial and angular light ray, we can estimate a depth map throughout
that properties of light field image. However, estimating a depth map from the light field image has a limitation in term of short baseline and low resolution issues. Even though many approach have been developed, but they still have a clear flaw in computation cost and depth value accuracy.
In this paper, we propose a network-based and epipolar plane image (EPI) light field depth estimation technique. Since the light field image consists of many sub-aperture images in a 2D spatial plane, we can stack the sub-aperture images in different directions to handle occlusion problem.
However, usually used many light field subaperture images are not enough to construct huge datasets. To increase the number of sub-aperture images for stacking, we train the network with augmented light field datasets. In order to illustrate the effectiveness of our approach, we perform the
extensive experimental evaluation through the synthetic and real light field scene. The experimental result outperforms the other depth estimation techniques.