The Eff2 Project: Towards Efficient and Effective Support for Large-Scale High-Dimensional Indexing

Research in content-based multimedia retrieval needs to be a joint venture between the multimedia processing and database communities, in order to be usable in practice at a realistic scale. Unfortunately, however, while both communities have performed some excellent research, they have done so in isolation from each other. In this paper we describe the lessons learned from the Eff² project on Efficient and Effective image retrieval. In this project we have taken two major actions for bridging this gap. First, we have defined the CVDB workshop series in order to bring together researchers from the database and computer vision communities and to define some research directions that can benefit both communities. Second, we have designed and studied an extremely efficient disk-based indexing scheme, called the NV-tree, for very large collections of advanced image feature vectors.