Abstract
Due to their superior performance on efficient search and discrete loss, quantization methods have attracted considerable attention for approximate nearest neighbor (ANN) search on large-scale multimedia data. In this paper, we aim to introduce quantization into cross-modal similarity search with a focus on learning discriminative binary codes. Different from existing cross-modal quantization algorithms that transform heterogeneous data to a common subspace by unsupervised approaches, we propose a novel cross-modal quantization method embedded in a supervised framework by exploring the discriminative property of label information. The proposed approach learns common semantic space from label information by linear classification, enabling the generated category-specific features to produce more discriminative quantization codes. Furthermore, the unified codebooks and quantization codes are adopted to preserve the correlation of similar inter-modal pairs in the learned semantic space. The overall optimization algorithm jointly learns the linear classifiers, category-specific features and the unified quantizer in an alternated strategy. Extensive comparative experiments on three benchmark datasets show the superiority of our approach over some state-of-the-art methods.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have