Multiple Query-Independent Values Based Asymmetric Ranking for Approximate Nearest Neighbor Search

Abstract

With the advantage of computational and storage efficiency, binary encoding and Hamming distance are widely used in approximate nearest neighbor search. However, a large number of candidate points share the same Hamming distance with the query when the database contains too many points, leading to the issue of confusing ranking. Therefore, asymmetric ranking is proposed to address this issue, in which they only encode the candidate points into binary codes while computing two query-independent values that represent 0 and 1 respectively for each bit, thereby computing asymmetric distance for ranking. The asymmetric distance between the query point and one candidate point equals to the 1-norm distance between the query and a vector of query-independent values corresponding to its binary code. It is more accurate than Hamming distance because of taking advantage of the full information of the query, but the information of candidate points is lost when only computing two query-independent for each bit of encoded candidate points. Therefore, we propose the Multiple query-independent values based Asymmetric Ranking(MARank) in which more than two query-independent values are computed for each bit in order to divide the distance space more densely. The MARank is applicable to many kinds of binary encoding methods such as LSH, SH, ITQ, PCAH and so on. We conduct experiments on five datasets: SIFT-10K, CIFAR-10, Caltech-256, MNIST and NUS-WIDE to compare MARank with the asymmetric ranking in terms of time and accuracy. The results show that the proposed MARank algorithms can achieve up to 22% performance gains over Hamming distance and 13% over the asymmetric distance.