GrowBit: Incremental Hashing for Cross-Modal Retrieval

Abstract

Cross-modal retrieval using hashing techniques is gaining increasing importance due to its efficient storage, scalability and fast query processing speeds. In this work, we address a related and relatively unexplored problem: given a set of cross-modal data with their already learned hash codes, can we increase the number of bits to better represent the data without relearning everything? This scenario is especially important when the number of tags describing the data increases, necessitating longer hash codes for better representation. To tackle this problem, we propose a novel approach called GrowBit, which incrementally learns the bits in the hash code and thus utilizes all the bits learned so far. We develop a two-stage approach for learning the hash codes and hash functions separately, utilizing a recent formulation which decouples over the bits so that it can incorporate the incremental approach. Experiments on MirFlickr, IAPR-TC-12 and NUS-WIDE datasets show the usefulness of the proposed approach.