Computer Vision News - December 2019

Results The paper shows high quality results in generating random images froma single image. The authors also explore their model on several computer vision applications such as super resolution, paint to image, harmonization and more. We start by showing an empirical study presented in the paper, about the different components of the model. Below we can see the generation of images in different scales, captured at inference time. The figure shows the coarsest level, and its consecutive two up-sampled levels. It can be seen how the model refines the solution and, as it goes to the upper levels, the generated images become more realistic. Moreover, the shape and pose of the object are preserved while the texture is improved at each level. Research 6 Another nice demonstration in the paper is the effect of training with a different number of scales. It can be seen that the number of scales in the SinGAM strongly influences the results. A model with a small number of scales only captures texture, and as the number of scales increases, the model manages to capture larger structures as well as the global arrangement of objects in the scene. Best of ICCV 2019