7. H. Jegou: Improving projection-based approximate nearest neighbors search

Omedrank is a similarity search approach based on random projections associated to median rank aggregation. It has shown to be an efficient solution to the general problem of finding the k-nearest neighbors of a query point in a high-dimensional space. In this poster, the
relevance of the search on synthetic data is significantly improved by implementing a proper selection of the projection directions and by modifying the voting mechanism, and this for a comparable computing cost. As a preliminary result, the method is adapted to SIFT descriptors in a context of image retrieval.