Dynamic Texture with Fourier Descriptors

Dynamic textures encompass the class of video sequences that exhibit some stationary properties in time. Examples of such textures include ripples on water, smoke out of a chimney, waving flags etc. Recently, it has been shown that such textures can be expressed as the output of a linear dynamic system that can be identified from a sample image sequence. In this paper, we propose to identify the underlying linear dynamic system using a set of Fourier descriptors of the frames rather than the raw sequence. It is shown that this approach enhances the performance time of the learning process and only requires a much smaller learning data set. The proposed algorithm improves upon published algorithms by exploiting the inter pixel dependency, while producing similar visual quality for the synthesized sequences, as shown through several experiments conducted on different dynamic texture sequences.