Computer Vision News - December 2019

Research 8 The last application we show is the harmonization. This is the task of realistically blend a pasted object with the background image. To perform this task, the authors suggest to train the model on the background image, and then feeding the naively pasted object image into the coarsest level of the image. The results can be seen below (compared to Deep Paint Harmonization). Conclusion SinGAN is a novel framework which is able to generate images from a single image training. It uses a pyramid of generators that begins by generating an image from a random noise and ends in generating globally consistent image. The model learns the image's patch statistics across multiple scales, using the multiscale scheme. The authors demonstrate the applicability of the model to several computer vision tasks and show remarkable results. For this novel approach, the paper won the best paper prize at the latest ICCV. For more results and information we highly recommend you to read the paper. Best of ICCV 2019