Item

Abstract

In this work we propose a novel approach for realistic fire animation and manipulation. We apply a statistical learning method to an image sequence of a real-world flame to jointly capture flame motion and appearance characteristics. A low-dimensional generic flame model is then robustly matched to the video images. The model parameter values are used as input to drive an Expectation-Maximization algorithm to learn an {\em auto regressive process} with respect to flame dynamics. The generic flame model and the trained motion model enable us to synthesize new, unique flame sequences of arbitrary length in real-time.