Mathematical Models in Computer Vision: Modeling Dynamic Scenes: An Overview of Dynamic Textures

ABSTRACT:Dynamic scenes with arbitrary radiometry and geometry present a challenge in that a physical model of their motion, shape, and reflectance cannot be inferred. Therefore, the issue of representation becomes crucial, and while there is no right or wrong representation, the task at hand should guide the modeling process. For instance, if the task is three-dimensional reconstruction, one can make assumptions on reflectance and illumination in order to recover shape and motion. If the task is synthesis, or reprojection, the correct shape is unimportant, as long as the model supports the generation of a valid view of the scene. If the task is detection or recognition, a physical model is not necessary as long as one can infer a statistical model that can be used to perform classification. We concentrate our attention on the two latter cases, and describe a modeling framework for dynamic scenes for the purpose of synthesis, detection and recognition. In particular, we restrict our attention to sequences of images of moving scenes that exhibit certain statistical stationarity properties, which have been called Dynamic Textures. They include sea-waves, smoke, foliage, whirlwind etc. In this chapter we describe a characterization of dynamic textures and pose the problems of modeling, learning, recognition and segmentation of dynamic textures using tools from time series analysis, and system identification theory.

	Main		Mathematical Models in Computer Vision: The Handbook, Springer (2005)
	Editors Preface Contents Contributors References Sample Chapter Order		Modeling Dynamic Scenes: An Overview of Dynamic Textures, G. Doretto and S. Soatto ABSTRACT:Dynamic scenes with arbitrary radiometry and geometry present a challenge in that a physical model of their motion, shape, and reflectance cannot be inferred. Therefore, the issue of representation becomes crucial, and while there is no right or wrong representation, the task at hand should guide the modeling process. For instance, if the task is three-dimensional reconstruction, one can make assumptions on reflectance and illumination in order to recover shape and motion. If the task is synthesis, or reprojection, the correct shape is unimportant, as long as the model supports the generation of a valid view of the scene. If the task is detection or recognition, a physical model is not necessary as long as one can infer a statistical model that can be used to perform classification. We concentrate our attention on the two latter cases, and describe a modeling framework for dynamic scenes for the purpose of synthesis, detection and recognition. In particular, we restrict our attention to sequences of images of moving scenes that exhibit certain statistical stationarity properties, which have been called Dynamic Textures. They include sea-waves, smoke, foliage, whirlwind etc. In this chapter we describe a characterization of dynamic textures and pose the problems of modeling, learning, recognition and segmentation of dynamic textures using tools from time series analysis, and system identification theory.

			Last Update: December20th, 2004, you can mail your comments to: nikos.paragios@computer.org