“To enable Cinematic photos on existing pictures that were not taken in multi-view stereo, we trained a convolutional neural network with encoder-decoder architecture to predict a depth map”

Link. Age of wonders.