2. G. Csurka, E. Gaussier, F. Pacull, F. Perronnin, J-M. Renders: Image and Multimedia Categorization

In the first part we present our Generic Visual Categorization (GVC) system based on visual adapted vocabularies. The idea is to build a universal vocabulary of image patch descriptors using Gaussian Mixture Models (GMM) and to adapt this vocabulary for each class independently with a Bayesian adaptation of the GMMs. Then the image is characterized by a set of histograms - one per class - where each histogram describes whether the image content is best modeled by the universal vocabulary or the corresponding class vocabulary. Experimental results are shown on a 35 class database created in the course of the European RevealThis project.

In the second part we present the multimedia categorization as a multiple-view categorization problem where documents have to be categorized into more than one category system. More particularly, we address the case where the two different categorizers (in our case text and image categorizers) have already been built previously based on non-necessarily identical training sets. On the top of these categorizers considered as black-boxes, we propose two re-weighting algorithms able to exploit dependencies between category systems using a third training set containing a few annotated examples. Experimental results are shown on web data and video sequences.