Stacked conditional adversarial networks and multi-patch loss for image depth estimation

Generative Adversarial Nets (GAN) were proposed as generative models estimated by an adversarial process between a generator and a discriminator. This allows the generator to learn a distribution of training data. Conditioning this network on additional input provides a conditional GAN which has been shown to be a good architecture for image translation tasks. We propose two changes to this network, with a focus on image depth estimation task. First, by using a stacked coarse to fine s tack of GANs, and second by using a multi-patch discriminator loss we show improved results in estimating depth from single images.