Scaled functional principal component analysis for human motion synthesis

Many of the existing data-driven human motion synthesis methods rely on statistical modeling of motion capture data. Motion capture data is a high dimensional time-series data, therefore, it is usually required to construct an expressive latent space through dimensionality reduction methods in order to reduce the computational costs of modeling such high-dimensional data and avoid the curse of dimensionality. However, different features of the motion data have intrinsically different scales and as a result we need to find a strategy to scale the features of motion data during dimensionality reduction. In this work, we propose a novel method called Scaled Functional Principal Component Analysis (SFPCA) that is able to scale the features of motion data for FPCA through a general optimization framework. Our approach can automatically adapt to different parameterizations of motion. The experimental results demonstrate that our approach performs better than standard linear and nonlinear dimensionality reduction approaches in keeping the most informative motion features according to human vision judgment.