Learning Non-Gaussian Stochastic Systems for Dynamic Textures

Figure 1. A. Factor graph for an LDS with state-space potentials. B. Images corresponding to the training data and outputs of different algorithms. Scatterplots of the first two state component sequences (red = training, blue = simulated), along with the approximate scale. Only our method results in a realistic state simulation and correct clock structure. C. Other videos we are working on: smoke, traffic, flag, hallway.