Oles Petriv
14 Oct 2018
10:00 - 10:45
Stream B

Latest improvements in video super resolution using generative adversarial networks

In this talk we are going to discuss the following: 1) fundamental problems of single image super resolution (one to many mapping, local and non-local context dependency) 2) frame consistency problem in video super-resolution and other frame transformation tasks 3) super-resolution quality assessment issues (there is no ideal SR quality metric) 4) methods of training and best architectures of supervised 2x and 4x single image super-resolution 5) methods of video super-resolution with motion compensation and without it (pros and cons) 6) my approaches to train single image and video 4x SR models for unknown image degradations using generative adversarial networks for degradation modeling 7) tricks to make local patterns distribution lie on the manifold of natural images 8) fast, model free, method of getting good frame consistency for sequence of single image SR processed frames using correlation kernels.