Koopman operator based nonlinear dynamic textures

Dynamic texture (DT) is a simple yet powerful paradigm to model videos with repetitive spatiotemporal behavior. In this paper we propose a novel nonlinear approach for modeling complex DTs based on Koopman operator theoretic method. Koopman operator is linear but infinite dimensional operator, and captures full nonlinear behavior. We exploit this aspect to construct a linear stochastic system in Koopman mode space, and propose it as a model for DTs. Through various complex videos, we show superior performance of our modeling approach over other methods proposed in the literature.