Re-ranking with two-level hashing

As the exponential growth of the Internet database, searching nearest neighbors in a high dimensional space using linear time algorithms has become impractical. In recent years, hashing based approximated nearest neighbor search has been studied intensively and successfully applied to many fields. The major challenge of hashing researches is to trade-off the similarity preservation with the code length for storage efficiency. A long binary code is required to preserve a good similarity among images, but recall rate and efficiently decrease. A short binary code yields a high recall rate but a low precision rate. Almost all of the existing methods do not explicitly address this dilemma. In practice, most of them need to find the appropriate binary code length to balance precision, recall and computation efficiency. In this paper, we propose a re-ranking scheme that can be applied to most of existing hashing methods. We first generate a low dimensional binary code as the search index that roughly approximates the similarity. Then, a high dimensional binary code is generated to re-rank those roughly similar images. In this way, a higher precision rate is achieved while keeping the computation cost low.