A Prior-Free Weighting Scheme for Binary Code Ranking

Abstract

Fast similarity search has been a research focus in recent years. Binary hashing, which embeds high-dimensional data points into Hamming space, is a promising way to accelerate similarity search, since its search process can be performed in real-time by using Hamming distance as similarity metric. However, as Hamming distance is discrete and bounded by code length, its resolution is limited. In practice, there are often many results sharing the same Hamming distance to a query, which poses a critical issue for problems where ranking is important. This paper proposes a weighted Hamming distance ranking algorithm (WhRank) to give a better ranking of results with equal Hamming distances to a query. By assigning different bit-level weights to different bits, WhRank is able to distinguish between the relative importance of different bits, and to rank the results at a finer-grained hash code level rather than the original integer Hamming distance level. We show that an effective weight is not only data-adaptive but also query-sensitive, and give a simple yet effective prior-free weight learning algorithm. Evaluations on three large-scale image datasets containing up to one million points demonstrate the efficacy of the proposed algorithm.