Computer Vision News - December 2019

SinGan 7 We now dive into the different applications of the framework. SinGAN model is also able to perform super resolution from a single image. This can be done by iteratively upsampling the image and feeding it to the SinGAN's finest scale generator. The authors compared their results to five states of the art super resolution methods and demonstrated comparable results to SRGAN. Paint-to-image is another cool application of SinGAN. This is done by downsampling the clipart image and feeding it into one of the coarsest levels of the model. As can be seen below, this is a task that is well suited to this model. The SinGAN outperforms two other methods and generates an image that preserves the layout and general structure with realistic texture and fine details that match the training image. Best of ICCV 2019