Curvature oriented clustering of sparse motion vector fields

We present an approach to unveil the underlying structure of dynamic scenes from a sparse set of local flow measurements. We first estimate those measurements at carefully selected locations, and subsequently group them into a finite set of different dense flow field hypotheses. These flow fields are represented as parametric functional models, and the number of flow models (=clusters) is determined by an information-theory based approach. Methodically, the grouping task is a two-step clustering scheme, whose intra-cluster modeling step exploits prior knowledge on real flow fields by enforcing low curvature, and the individual covariance matrices of the sparse local flow measurements are introduced in a principled way. The method has been tested successfully on both stereo and general motion sequences from the standard Middlebury database.