AI Seminar sponsored by Fortive

Victor Lempitsky -- Deep Generative Models for Avatars and Landscapes

Abstract: Deep Generative Models, in particular those based on adversarial learning, have achieved remarkable results at synthesizing and editing 2D photographs. In this talk, I will discuss how a very successful model of this class (StyleGAN) can be extended to two new domains. First, I will discuss an extension that allows to model realistic landscape timelapse videos. After training, the system can synthesize new landscape videos. The resulting model can also reenact static landscape photographs, and we show that such reenactments outperform previous approaches to this task. The second extension allows to model the fullbody appearance of humans. Here, we combine StyleGAN with a modern deformable body model (SMPL-X) and a neural rendering approach into a system that can synthesize 3D fullbody avatars from scratch or create such avatars from one or few photographs.

Bio: Victor Lempitsky leads the Samsung AI Center in Moscow as well as the Vision, Learning, Telepresence (VIOLET) Lab at this center. He is also an associate professor at Skolkovo Institute of Science and Technology (Skoltech). In the past, Victor was a researcher at Yandex, at the Visual Geometry Group (VGG) of Oxford University, and at the Computer Vision group of Microsoft Research Cambridge. He has a PhD ("kandidat nauk") degree from Moscow State University (2007). Victor's research interests are in various aspects of computer vision and deep learning, in particular, generative deep learning. He has served as an area chair for top computer vision and machine learning conferences (CVPR, ICCV, ECCV, ICLR, NeurIPS) on multiple occasions. His recent work on neural head avatars was recognized as the most-discussed research publication of 2019 by Altmetric Top 100 rating.