Locally Optimized Hashing for Nearest Neighbor Search

Abstract

Fast nearest neighbor search (NNS) is becoming important to utilize massive data. Recent work shows that hash learning is effective for NNS in terms of computational time and space. Existing hash learning methods try to convert neighboring samples to similar binary codes, and their hash functions are globally optimized on the data manifold. However, such hash functions often have low resolution of binary codes; each bucket, a set of samples with same binary code, may contain a large number of samples in these methods, which makes it infeasible to obtain the nearest neighbors of given query with high precision. As a result, existing methods require long binary codes for precise NNS. In this paper, we propose Locally Optimized Hashing to overcome this drawback, which explicitly partitions each bucket by solving optimization problem based on that of Spectral Hashing with stronger constraints. Our method outperforms existing methods in image and document datasets in terms of quality of both the hash table and query, especially when the code length is short.