Machine learning in general and deep neural networks in particular have been extensively used in a wide spectrum of applications ranging from automatic language translation to pattern recognition in machine vision. One of the popular applications of the so-called artificial intelligence applied to imaging has been in deepfakes.
Deepfakes refer to synthethic video of individuals who are animated at will, in such a photorealistic manner that they can fool untrained eyes into believing that they are real. Many such approaches require no more than a single picture of the targeted individual. Several solutions to create deepfakes are now available publicly without requiring any special skills or hardware to use them. Results obtained using the latter have shown to be at times so baffling that they compete and even occasionally surpass what can be produced by sophisticated special effects relying on computer graphics, expensive software and hardware and skilled operators. Social media are full of amazing and entertaining contents produced by regular people for the purpose of amusement and entertainment. Unfortunately, at the same time deepfakes are also being used to attack the dignity of celebrities and politicians, to potentially contribute to social unrest and to impact the general public opinion.
A new application of the technologies used in deepfakes is in immersive communication and telepresence such as those under the scope of AdMiRe project.
Imagine a remote person inserted into a television program using high resolution and high quality pictures of the latter transmitted off-line, and animated by the gesture and expressions of that individual through remote communication using an off-self smart phone. The latter is among many options that EPFL, a partner of the project, is exploring in order to allow for high quality, photo-realistic and easy to use telepresence of remote individuals in AdMiRe.
Prof. Dr. Touradj Ebrahimi / EPFL
Short video of a person in low light and low quality conditions that can animate the picture of the same person in higher quality, from a single picture of that person