Institut für Neuroinformatik, Ruhr-Universität Bochum, Germany
Wednesday 7 November 2007, 16.00
Seminar Room B10 (Basement)
Alexandra House, 17 Queen Square, London, WC1N 3AR
Organic Computing for face and object recognition
The talk starts by identifying hallmarks of Organic Computing, a new computer science philosophy with good potential for progress on artificial vision systems. Important parts of this methodology are learning from nature, discretizing continuous dynamics, and the integration of submodalities. The holy grail of automatizing vision systems is autonomous learning of the necessary routines from examples.
I will then describe correspondence-based techniques for face and object recognition and show where they need refinement through learning from examples. Tracking of facial points is an important prerequisite for video compression and animation. Tracking in general is a difficult problem, which requires global constraints to function anywhere near stable. In the first learning example, constraints for face tracking are learned automatically from bunch graph matching on a large number of frontal images. The second example is ongoing work on tracking of articulated human body movement. The third consists of the classification of genetic syndromes which can be diagnosed from the shape of the face by an expert. Graph matching together with standard classifiers yielded comparable recognition rates.
A major drawback of correspondence-based recognition methods is the time-consuming matching operation. I finally present a combination of rapid feature-based preselection with correspondence-based matching and a study of the relative importance of either component.