Comparing GPLVM approaches for dimensionality reduction in character animation

Abstract

Gaussian Process Latent Variable Models (GPLVMs) have been found to allow dramatic dimensionality reduction in character animations, often yielding two-dimensional or three-dimensional spaces from which the animation can be retrieved without perceptible alterations. Recently, many researchers have used this approach and improved on it for their purposes, thus creating a number of GPLVM-based approaches. The current paper introduces the main concepts behind GPLVMs and introduces its most widely known variants. Each approach is then compared based on various criteria pertaining to the task of dimensionality reduction in character animation. In the light of our experiments, no single approach is preferred over all others in all respects. Depending whether dimensionality reduction is used for compression purposes, to interpolate new natural looking poses or to synthesize entirely new motions, different approaches will be preferred.