Learning parameterized models of image motion

摘要: A framework for learning parameterized models of optical flow from image sequences is presented. class motions represented by a set orthogonal basis fields that are computed training using principal component analysis. Many complex can be linear combination small number these flows. The learned motion may used estimation and model-based recognition. For we describe robust, multi-resolution scheme directly computing the parameters derivatives. As examples consider discontinuities, non-rigid human mouths, articulated motion.