Experimental Evaluation of the MPEG-7 Visual Descriptors for Object Classification

In this paper we present an experimental study on the efficiency of the various MPEG-7 visual descriptors as far as the object classification is concerned. Seven different visual descriptors are investigated with the aid of three state of the art classifiers. Our data consist of 7686 manually annotated object instances corresponding to eight different object classes. A total of 1952 images from the athletics domain collected in the framework of the FP6 BOEMIE project were used for blob extraction. The aim of this study is to provide practical hints to the researchers working in the areas of automatic object detection and classification concerning the utilization of the MPEG-7 visual descriptors. The training and testing files we used along with the learned object models are provided online so as to be used as benchmark data. The results show there is a significant difference on the performance of the individual descriptors as well as on their combination with other descriptors irrespectively of the classifier used.