Towards Robust Perception and Model Integration

Many of today's vision algorithms are very successful in controlled environments. Real-world environments, however, cannot be controlled and are most often dynamic with respect to illumination changes, motion, occlusions, multiple people, etc. Since most computer vision algorithms are limited to a particular situation they lack robustness in the context of dynamically changing environments. In this paper we argue that the integration of information coming from different visual cues and models is essential to increase robustness as well as generality of computer vision algorithms. Two examples are discussed where robustness of simple models is leveraged by cue and model integration. In the first example mutual information is used as a means to combine different object models for face detection without prior learning. The second example discusses experimental results on multi-cue tracking of faces based on the principles of self-organization of the integration mechanism and self-adaptation of the cue models during tracking.