Recognizing real world objects using multiple views

Object recognition, i.e. the classification of objects into predefined categories is an important tool in many computer vision systems. Unlike other types of recognition, it must be quite generic in order to be able to handle the great variety of objects that can exist. An aid to the solution of this difficult problem can be the use of information from different camera views. Here, we have extended a robust object recognition method in order to be able to function with information from more than one camera, and from arbitrary viewpoints. This method uses local feature points to construct visual vocabularies which then form an input to Support Vector Machines. We have found the multi-camera variant to produce superior results to the single-camera one.