Non-Linear Dynamic Texture Analysis and Synthesis Using Constrained Gaussian Process Latent Variable Model

Abstract

Linear dynamic system (LDS) has been proposed to model dynamic texture. However, the temporal evolution of dynamic texture is non-linear in general and is not fully captured by the linear model. In this paper, we formulate the dynamic texture learning and synthesis via nonlinear approach. Assuming that dynamic texture is sampled from a low dimensional manifold, the constrained Gaussian process latent variable model (CGPLVM) is proposed to model the dynamic texture as a set of latent states. The essence of dynamic texture is captured as the spatial relationship within the latent states. Moreover, Metropolis-Hastings sampling method is used to sample new states, which hold the spatio-temporal statistics of dynamic texture. Experimental results demonstrate that our approach can produce dynamic texture sequences with promising visual quality.