Yosef Yomdin (Math and CS, Weizmann Institute of Science)
Tuesday, 16.11.2010, 11:30
In this talk I plan to discuss some results on image representation and processing via geometric models, in two extreme scales:
1. The finest scale representation is usually called "vectorization". Here the models are analytic aggregates mainly constructed from edges and ridges, equipped with their "extended color profiles". We provide some experimental data stressing the importance of accurate and flexible ``color profiles" of edges and ridges in our visual perception. Accordingly, the extended color profiles play the most central role in our vector representation and in its editing in "vector form". We describe our vector format which allows (for the first time, as we believe) for a "visually indistinguishable" reconstruction of high-resolution photo-images, and for their geometric processing.
2. Another extreme is the problem of capturing of global characters and objects on the image with global parametric models, like the model of an entire human figure. This is a very important step in "photo-animation". Recently motion data bases have been used in combination with model fitting in a work of Aachen Computer Vision group to provide high-quality animations of certain photo-characters. However, essential difficulties arise when the output 2D animation format is relatively rigid layers based one, like Flash. Another problem appears when the anatomic proportions and the pose of the character on the photo differ strongly from that used in the motion data-base. We show in this talk how to transform properly the data-base 3D animations into a 2D layer-based format, in order to preserve an illusion of a 3D motion, and how to correct it in order to make it appropriate for a specific pose and anatomic proportions of the photo-character.
Department of Mathematics, The Weizmann Institute of Science, Rehovot, Israel.