Learning Visual Object Categories with Global Descriptors and Local Features

Different; types of visual object categories can be found in real-world applications. Some categories are very heterogeneous in terms of local features (broad categories) while others are consistently characterized by some highly distinctive local features (narrow categories). The work described in this paper was motivated by the need to develop representations and categorization mechanisms that: can be applied to domains involving different types of categories. A second concern of the paper is that these representations and mechanisms have potential for scaling up to large numbers of categories. The approach is based on combinining global shape descriptors with local features. A new shape representation is proposed. Two additional representations are used, one also capturing the object's shape and another based on sets of highly distinctive local features. Basic classifiers following the nearest-neighbor rule were implemented for each representation. A meta-level classifier, based on a voting strategy, was also implemented. The relevance of each representation and classifier to both broad and narrow categories is evaluated on two datasets with a combined total of 114 categories.