Abstract

In this paper, we present a novel method to synthe-size dynamic texture sequences from extremely few samples,e.g., merely two possibly disparate frames, leveraging bothMarkov Random Fields (MRFs) and manifold learning. De-composing a textural image as a set of patches, we achievedynamic texture synthesis by estimating sequences of tem-poral patches. We select candidates for each temporal patchfrom spatial patches based on MRFs and regard them assamples from a low-dimensional manifold. After mappingcandidates to a low-dimensional latent space, we estimatethe sequence of temporal patches by finding an optimal tra-jectory in the latent space. Guided by some key proper-ties of trajectories of realistic temporal patches, we derivea curvature-based trajectory selection algorithm. In con-trast to the methods based on MRFs or dynamic systemsthat rely on a large amount of samples, our method is able to deal with the case of extremely few samples and requires no training phase. We compare our method with the state of the art and show that our method not only exhibits superiorperformance on synthesizing textures but it also producesresults with pleasing visual effects.