Video Retargeting Enhancement For An Interactive Exhibition

Abstract

The quality of computer-generated images and videos has progressed with technology to a point where it is difficult to tell the difference between what is real and what is not. An example is given by the video retargeting application, which consists in animating a target human subject with any desired motion. In recent years, several video retargeting methods has been developed. However, the majority of these methods share the big limitation of producing low-resolution results. To solve this issue, in this thesis, we propose a new framework called One-Shot- Video-Enhance, that allows to maintain high-resolution in synthesized retargeted videos, by transferring high-frequency details belonging to the target subject that otherwise would be lost. Compared to other deep quality-restoration approaches, our model not only proves to be faster, but also to outperform them in terms of details recovering, color and temporal consistency. In addition to that, One-Shot-Video-Enhance produces results that are extremely faithful to the target appearance, and this contributes to a realistic aspect of the subject in the animated video. Foremost, our model does not require any additional data than the one already needed for the synthesis of the retargeted video.