Fast Visual Object Identification and Categorization

Recently recognition and categorization of objects from images using local features has become very popular. While similar approaches have been used for identification and categorization tasks they have not been treated in a common framework. In this paper we present a method that treats visual object identification, categorization in a common framework by exploiting ideas from image interclass transfer. We propose a hierarchically organized visual memory, where the high levels of the hierarchy represent generic classes and the leaves individual objects. The features used in the nodes of the hierarchy are learned using Adaboost on integral orientation histogram features (using these features makes a real time implementation possible). Learning the discrimination within a layer of the hierarchy is inspired by the work of Ferencz. Therefore, one can view our method as a hierarchical generalization of the interclass image transfer. First experiments demonstrate that the proposed method is able to learn meaningful object categories, as well as identification of individual objects.