Human body motion generation

Mathis Petrovich (Ph.D. Student)

Automatic analysis of people from visual data is of great importance for numerous applications in behavior analysis and prediction, autonomous driving, health care, surveillance, content search, and entertainment. More specifically, analyzing human body dynamics is an essential component of various research directions, such as action recognition, robotics, human-centric video generation, and 3D animation. Despite active research in this area for decades, many fundamental questions remain open. To this date, there is no method that is capable of generating a sequence of human shape and motion, conditioned on a wide range of action categories. This project will explore modeling complex human body motions using deep neural networks. The thesis will have a particular focus on generative models of 3D action sequences. The thesis will first concentrate on learning a generative model of actions from large motion capture collections, as well as from large video collections with action annotations. The thesis will then explore how such a generative model can be used to improve 3D motion estimation from videos and action recognition.

Primary Host:	Gül Varol (École des Ponts ParisTech)
Exchange Host:	Michael J. Black (Max Planck Institute for Intelligent Systems)
PhD Duration:	01 October 2020 - 30 September 2023
Exchange Duration:	01 April 2022 - 30 September 2023 - Ongoing