Data-oriented multi-index Hashing

The multi-index hashing (MIH) is the state-of-the-art method for indexing binary codes. However, it is based on the dataset codes uniform distribution assumption, and will lower efficiency in dealing with non-uniformly distributed codes. In this paper, we propose a data-oriented multi-index hashing method. We first compute the correlations between bits and learn adaptive projection vector for each binary substring. Then, instead of using substrings as direct indices into hash tables, we project them with corresponding projection vectors to generate new indices. With adaptive projection, the indices in each hash table are nearly uniformly distributed. Besides, we put forward an entropy based measurement to evaluate the distribution of data items in each hash table. Experiments conducted on reference large scale datasets show that compared to the MIH the time performance of our method can be 36.9％～87.4％ better .