Shift-Invariant Dynamic Texture Recognition

We address the problem of recognition of natural motions such as water, smoke and wind-blown vegetation. Such dynamic scenes exhibit characteristic stochastic motions, and we ask whether the scene contents can be recognized using motion information alone. Previous work on this problem has considered only the case where the texture samples have sufficient overlap to allow registration, so that the visual content of the scene is very similar between examples. In this paper we investigate the recognition of entirely non-overlapping views of the same underlying motion, specifically excluding appearance-based cues. We describe the scenes with time-series models—specifically multivariate autoregressive (AR) models—so the recognition problem becomes one of measuring distances between AR models. We show that existing techniques, when applied to non-overlapping sequences, have significantly lower performance than on static-camera data. We propose several new schemes, and show that some outperform the existing methods. 1 Recognition from motion Motion is a powerful cue for visual recognition of scenes and objects. Johansson’s moving dot displays [1] show that objects which are highly ambiguous from a single view are readily recovered once motion is supplied. In computer vision, the classification of scenes from motion information has seen considerable research, summarized in the recent survey of Chetverikov and Péteri [2]. In this paper, we focus on classification of objects using the class of state-space dynamic texture models introduced by Doretto and Soatto [3, 4] and Fitzgibbon [5]. Dynamic textures are image sequences of moving scenes which exhibit characteristic stochastic motion. Examples include natural scenes such as water, wind-blown flowers and fire. State-space models [5, 4] view a dynamic texture as a realization of a time-series model such as an autoregressive process. By determining the model parameters for such sequences, we can hope to recognize similar motions by comparing the models represented by the parameters. Our goal in this paper is to define a distance measure between pairs of image sequences which is low for models representing the same motion (or motion class), and high for models derived from motions of different classes. Such distance