Latent nonuniform splines for animation approximation

Abstract

This paper presents a new method to approximate animation sequences through a nonlinear analysis of the spatiotemporal data. The main idea is to find a spline curve which best approximates a multivariate animation sequence in a reduced subspace. Our method first eliminates data redundancy among multiple animation channels using principal component analysis (PCA). The reduced sequence of latent variables is then approximated by a nonuniform spline with free knots. To solve the highly-nonlinear multimodal problem of the knot optimization, we introduce a stochastic algorithm called covariance matrix adaptation evolution strategy (CMA-ES). Our method optimizes the control points and the free knots using least-square method and CMA-ES, which guarantees the best approximation for arbitrary animation sequences such as mesh animations and motion capture data. Moreover, our method is applicable to practical production pipeline because both PCA-and CMA-based algorithms are computationally stable, efficient, and quasi manual parameter-free. We demonstrate the capability of the proposed method through comparative experiments with a common approximation technique.