Depth Estimation from Light Field Images via Convolutional Residual Network
Ji-Hun Mun, Yo‐Sung Ho · 2018
Estimating the depth map from multi-view images or light field images is an essential component in 3D geometry analysis. Conventionally, the depth map is estimated from integrated local and global information of stereoscopic images. However, the estimated depth map is not accurate due to the depth discontinuity and homogeneity. To solve this problem, we propose a light field depth estimation method based on the convolutional residual network. The discontinuity problem in the depth map is handled by computing depth cost maps in the residual network. In addition, we consider a phase shifted light field image in the loss function to acquire a robust depth map in the homogeneous region of the light field images. Experimental results demonstrate that our network outperforms other neural network architectures in terms of the depth map accuracy.