CCPR 2008 Keynote Speech 3 and Keynote Speech 4

Abstract

Firstly, we consider the space of small image patches, say 7times7 pixels, and show that patches cropped from natural images are distributed in a wide variety of manifolds from low dimensional manifolds for regular textons and image primitives, to high dimensional manifolds for textures. Secondly, we introduce a learning and modeling method which pursues these manifolds by information projection, and derive statistical models, such as the active basis model for the low dimensional manifolds and the MRF models for the high dimensional manifolds. Thirdly we show how these two types of manifolds are integrated in large images to form the primal sketch model in early vision, and how they are mixed to form complex objects in the high level vision. Fourthly we show the transition of these manifolds through information scaling. The objective of this work is to study the structures of the image space, based on which we can understand the connections and transitions of a number of classic models used in computer vision and pattern recognition.