Keynote: Overview of AI methods for mixed reality: content acquisition and representation, editing and transmission.

Pierre Hellier

InterDigital (France)

In this talk, I will present an overview of recent AI methods to help content creation, edition and transmission for mixed reality applications. Concerning content acquisition and representation, state-of-the-art methods are now based on implicit neural representations or neural radiance field. There has been a recent explosion of Nerf for scene modeling. I will also decribe how a Generative model can be inverted to model particular observations. Content editing has recently benefited from AI, since machine learning techniques offer a signal representation that is disantengled, hence potentially suited for complex non-linear semantic editing. That has concerned style transfer, semantic facial editing, and human motion representation. Finally, I will review some AI techniques recently developed for image and video compression.

Pierre Hellier received his PhD in medical imaging in 2000, and moved as a postdoctoral researcher to Utrecht university. He was appointed as INRIA research scientist in 2001, focusing on medical image processing and image-guided surgery. In 2006, he was a visiting professor at McGill university in Montreal. In 2009, he co-founded a startup to industralize a neuronavigation system for transcranial magnetic stimulation, treating over 10 000 patients so far. In 2011, he moved to Technicolor, working on UGC content synchronization, enhancement, and professional postproduction (film grain, color grading, color restauration, unsharp masking, mosaicing). Since 2019, he is with InterDigital, focusing on machine learning techniques for image and video compression.