Computer Vision News - December 2019

SinGan 5 The input to every generator G n is a random noise image z n , and the generated image from the previous scale is upsampled to the current resolution. The coarsest level generates an image from a pure random noise. The model is easily explained by the following figures: To train the model, the authors suggest a sequential scheme that goes from the coarsest level to the finest one. Once a level was trained, it is kept fixed during the training of the other levels. The loss of the SinGAN is composed by a linear combination of two losses, reconstruction and adversarial loss, so the objective at each level takes the form of: The adversarial loss L adv penalizes the distance between the distribution of patches in x n to the distribution of patches in . Specifically, when an image is generated, a Markovian discriminator D n discriminates between overlapping patches in the image, i.e. classifies them being real or not. The final discriminator score is the average over the patches score. The reconstruction loss Lrec is a bit different. This loss enforces the generator to generate images that are similar, in some sense, to the original image. This can be done by ensuring that it exists an input noise map that generates the original image at each level x n . Specifically, for n<N the generator needs to generate x n if z n =0, and for n=N, it needs to generate x N for a specific noise map z*. Mathematically speaking, when n<N the penalization has the form of: And for n=N it is simply Where z* is a fixed noise map. Best of ICCV 2019