Visual recogntion in the real world

Prof. Jan-Olof Eklundh
KTH Stockholm, Sweden In recent years we have seen considerable efforts devoted to problems on visual object recognition. Approaches based on classical pattern recognition principles have been intermingled with methods extracting features and segmenting images and then using matching, or indexing. Experiments have been performed e.g. on images of faces, animals, everyday objects, and various kinds of indoor and outdoor scenes. Although, especially the latter type of images may represent rather general scenarios, limited attention has been given to problems on recognition in the real world. In this case there is some system that "sees", i.e. a system that is involved in a set of tasks (of which seeing is one), is embodied, and has some amount of knowledge. An example would be an assistive mobile robot capable of performing certain tasks, such as fetching and carrying things, while navigating and avoiding obstacles. In the talk we will consider the implications of such a scenario on visual recognition. We will first present a system for recognizing specific objects in a real indoor world based on learning using Support Vector Machines. We will then discuss what this system does and does not do in view of the sketched scenario, and delineate a more general system for it.