Tornado Motion Synthesis from Video Footage

In this paper we present a novel method to generate 3D animation of tornados from video clips taken from a single camera. Our source of input data is a variety of tornado video clips of widely-varying fidelity available on the world wide web. By extracting the silhouette of the foreground figure and using a model-based approach, the problem is reformulated as a local, optimized search of the shape space. The silhouette of the figure from the captured video is compared against the silhouette of a synthetic revolute using a pixel-by-pixel cost function. By using temporal coherence, physical constraints, and knowledge of tornado motion, the motion can be reconstructed. Additionally, we use this reconstruction to synthesize animations of the captured phenomena. Our main contribution is a method to synthesize a motion of highly dynamic objects like the tornado from a single video source. In addition, we employ simple yet robust 2D shape descriptors to reconstruct the shape and form of the tornado.