Hashing with Binary Matrix Pursuit

TL;DR

This paper introduces a theoretically grounded and empirically validated two-stage hashing method that leverages residual learning and high-capacity hash functions like CNNs to produce highly accurate binary codes for image retrieval.

Contribution

It provides a theoretical analysis of binary code quality, simplifies code inference with CNNs, and proposes a new hashing method that outperforms previous approaches.

Findings

Residual learning achieves arbitrary accuracy in fitting neighborhood structures.

High-capacity hash functions simplify binary code inference.

Proposed method outperforms previous hashing techniques on image retrieval benchmarks.

Abstract

We propose theoretical and empirical improvements for two-stage hashing methods. We first provide a theoretical analysis on the quality of the binary codes and show that, under mild assumptions, a residual learning scheme can construct binary codes that fit any neighborhood structure with arbitrary accuracy. Secondly, we show that with high-capacity hash functions such as CNNs, binary code inference can be greatly simplified for many standard neighborhood definitions, yielding smaller optimization problems and more robust codes. Incorporating our findings, we propose a novel two-stage hashing method that significantly outperforms previous hashing studies on widely used image retrieval benchmarks.