Abstract: Recent trends in large-scale image retrieval consider image representations based on the Bag of Visual Words model, because such representation enables efficient filtering using inverted files, while for more effective retrieval the representation can be enhanced by VLAD or Hemming embedding. In order to improve the effectiveness of the retrieval even more, some models try to relax from Bag of Visual Words model (i.e., from a shared feature space dictionary) and model content of multimedia objects by means of so-called feature signatures. Whereas the effectiveness of models based on feature signatures seems to be promising for several retrieval tasks, the efficiency of the models still represents a serious performance bottleneck, because the models employ expensive adaptive distance measures to compare two feature signatures. In this talk, we present several recent approaches that significantly improve the efficiency of the retrieval when using models based on feature signatures, making the models applicable also for large-scale image retrieval.
CV: Jakub Lokoc received the doctoral degree in software systems from the Charles University in Prague, Czech Republic. He is an assistant professor in the Department of Software Engineering at the Charles University in Prague, Faculty of Mathematics and Physics, Czech Republic. He is a member of siret research group and his research interests include metric access methods, multimedia retrieval and exploration, and similarity modeling.