Comparative study on automatic face edition

Abstract

[en] In 2018, a video of Barack Obama making a fake speech played by Jordan Pelee went viral [27], in the same year we have seen similar videos with different speeches and we have even seen how totally static images were animated, giving them a pattern to follow. That video and the other ones were made with Deep Learning technology which allowed to recreate the face of a person in the body of another or animate static faces. Since then techniques such as deep fake, face swapping, or face motion became more popular [14], we have seen it both in the movies and on social media, and thanks to this we can see how this impacts us on a social level, for both good and bad. All of this is because Computer Science has advanced exponentially lately, what we thought impossible a few years ago, is now possible and easy to do. Research in the field of computer vision has been very impressive and in a short time, we have gone from simply creating objects to generating artificial faces and making them replicate the patterns of behavior of other people. This project will consist of a state-of-the-art review of these different techniques. Three methods will be studied, all of them to edit the image that they obtain as input. It will detail the pipelines used and will show some results, with fixed parameters, to the public to obtain a qualitative assessment to define their faults and highlight the best of each method. And finally, a conclusion on the results obtained and the current status of this technology as well as its social impact and the applications that this involves.