Probabilistic motion models for animation
Docentlecture by Gustav Eje Henter, Division of Speech, Music, and Hearing
Tid: Fr 2023-06-02 kl 14.00
Plats: Fantum, Lindstedtsvägen 24
Medverkande: Gustav Eje Henter
Data-driven character animation holds great promise for enhancing realism and creativity in games, film, virtual avatars and social robots. However, due to the high bar on visual quality, most existing AI animation solutions focus narrowly on a specific task, and do not generalise to different motion types.
This talk makes the case that 1) machine learning now has advanced far enough that strong, task-agnostic motion models are possible, and that 2) these models should be probabilistic in nature, to accommodate the great diversity in how behaviours can be realised. We present MoGlow, a new, award-winning deep-learning architecture that leverages normalising flows and satisfies our two desired criteria. Experiments show that MoGlow is competitive with comparable models in locomotion generation for both humans and dogs.
For a longer introduction showing our models in action, please see the following video: youtu.be/pe-YTvavbtA
If time allows, I will also introduce applications of normalising flows to other motion-generation problems such as co-speech gestures and dance, along with our recent work on diffusion models for these tasks.