Efficient Transfer of Contact-Point Local Deformations for Data-Driven Simulations

Abstract

We present a new approach for data-driven deformation enrichment, which requires a much smaller set of examples. The central idea is to reuse deformation samples. This is done by transferring pre-generated examples from one contact point to another, when the surrounding material as well as the induced deformation are similar. Our similarity notion is derived from two descriptors that use multivariate Hermite polynomials as a basis. The first descriptor comprehends information on the local material density near a contact point, which allows us to segment an object into regions with similar material neighborhoods. At each characteristic location, multiple samples are obtained for different interaction patterns. The obtained information is then encoded in the second descriptor – the deformation descriptor. At run-time, the two descriptors are evaluated at the current contact point. Based on the similarity to the example descriptors, suitable pre-generated data is selected, interpolated, and used to enrich an object surface. We demonstrate our method in several applications and provide quantitative evaluations.