WARank: Weighted Asymmetric Ranking for Approximate Nearest Neighbor Search

Binary hashing based methods have been widely used for large-scale approximate nearest neighbor search because of their two benefits: less memory usage and high search efficiency. In these methods, binary code ranking is usually implemented based on Hamming distance or asymmetric distance. Generally, asymmetric distance is more accurate than Hamming distance, thus recent work focuses on the asymmetric distance ranking. In existing asymmetric distance ranking, query-independent values are approximated by sample average values. However, when the distribution of data is not uniform, sample average values are not representative, leading to wrong ranking results. To address this problem, we propose Weighted Asymmetric Distance Ranking (WARank) algorithm which consists of two parts. First, we present an otsu threshold-based method to obtain more appropriate query-independent values in which the otsu threshold performs almost the same with the average value when the distribution of the data is uniform but much better when it is not uniform. Second, we present bit-level weight calculation method by which we can assign different weights to different bits in order to minimize the negative effect of any bit without uniform distribution. The experiments on public datasets show that the proposed WARank algorithm further increases the search accuracy compared to state-of-the-art methods.