EAGER: Learning Upsampling Operators for Animation of Cloth and Fluids

University Of Utah, Salt Lake City UT

Investigators

Abstract

The PI's goal in this exploratory research is to tackle the fundamental obstacle preventing high quality, interactive animation of natural phenomena, namely the enormous number of degrees of freedom involved. Because hand animating a typical mesh is a tedious and time consuming task, computer graphics has turned to physics and simulation to animate most natural phenomena. In this context, the promise of simulation is generality; an infinite space of material properties and initial conditions can be explored. This generality is also simulation's greatest limitation; the space of possible animations is vast, while the space of desirable animations is a great deal smaller. The PI's approach is to use simulation's strength (its ability to create rich animation data under a variety of conditions) to combat its greatest limitations (high dimensionality and computational expense). To this end, he will develop machine learning tools for finding new and more expressive low-dimensional representations, which do not describe all possible animations but rather succinctly describe the space of desirable animations. Previous attempts to apply machine learning to the animation of natural phenomena have shown promise, but also significant limitations. These approaches have suffered from over-fitting, have sacrificed locality, and have not allowed artistic control over the space of possible animations. Furthermore, these approaches have been too data-driven, failing to allow for the input of valuable human knowledge and intuition or mathematical and physical models. Until these limitations are addressed, the promise of high-quality interactive computer animation of natural phenomena will remain out of reach. For concreteness the PI will focus on cloth and fluids as test bed domains (initially assuming an algorithmic paradigm of coarse simulation enhanced by data-driven upsampling operators), for which he will explore questions of sparseness, expanded feature sets, combining operators, and artistic control. Broader Impacts: Simulation is a powerful technique whose usefulness is not limited to computer animation. So while the test bed domains fall within the realm of traditional computer graphics, project outcomes will allow for high-quality, interactive computer animation of natural phenomena across all of science and engineering, with particular applicability to film, video games, virtual reality, medical training, etc. Moreover, the unique context of computer animation will necessarily require new machine learning algorithms that will feed back into that community as well. The PI plans to develop and release the majority of his source code under free BSD licenses.