Supervised Incremental Hashing

TL;DR

This paper introduces Supervised Incremental Hashing (SIH), a scalable method for updating hash functions efficiently in large-scale image search, using a two-stage classification framework with kernel SVMs.

Contribution

It presents a novel incremental learning approach for supervised hashing that efficiently updates hash functions with new data or classes.

Findings

Outperforms state-of-the-art supervised hashing methods on large datasets

Efficient incremental update procedure for new classes or images

Effective binary code optimization via cyclic coordinate descent

Abstract

We propose an incremental strategy for learning hash functions with kernels for large-scale image search. Our method is based on a two-stage classification framework that treats binary codes as intermediate variables between the feature space and the semantic space. In the first stage of classification, binary codes are considered as class labels by a set of binary SVMs; each corresponds to one bit. In the second stage, binary codes become the input space of a multi-class SVM. Hash functions are learned by an efficient algorithm where the NP-hard problem of finding optimal binary codes is solved via cyclic coordinate descent and SVMs are trained in a parallelized incremental manner. For modifications like adding images from a previously unseen class, we describe an incremental procedure for effective and efficient updates to the previous hash functions. Experiments on three large-scale image datasets demonstrate the effectiveness of the proposed hashing method, Supervised Incremental Hashing (SIH), over the state-of-the-art supervised hashing methods.