Unsupervised Learning of Depth and Depth-of-Field Effect From Natural Images With Aperture Rendering Generative Adversarial Networks

Figure 1. Unsupervised learning of depth and depth-of-field (DoF) effect from unlabeled natural images. (a) In training, we adopt only a collection of single-DoF images without any additional supervision (e.g., ground-truth depth, pairs of deep and shallow DoF images, and pretrained model). (b) Once trained, our model can synthesize tuples of deep and shallow DoF images and depths from random noise. The generated data are beneficial in training a shallow DoF renderer, which also requires no external supervision.