Learning Discriminative Representations for Semantic Cross Media Retrieval

Heterogeneous gap among different modalities emerges as one of the critical issues in modern AI problems. Unlike traditional uni-modal cases, where raw features are extracted and directly measured, the heterogeneous nature of cross modal tasks requires the intrinsic semantic representation to be compared in a unified framework. This paper studies the learning of different representations that can be retrieved across different modality contents. A novel approach for mining cross-modal representations is proposed by incorporating explicit linear semantic projecting in Hilbert space. The insight is that the discriminative structures of different modality data can be linearly represented in appropriate high dimension Hilbert spaces, where linear operations can be used to approximate nonlinear decisions in the original spaces. As a result, an efficient linear semantic down mapping is jointly learned for multimodal data, leading to a common space where they can be compared. The mechanism of "feature up-lifting and down-projecting" works seamlessly as a whole, which accomplishes crossmodal retrieval tasks very well. The proposed method, named as shared discriminative semantic representation learning (\textbf{SDSRL}), is tested on two public multimodal dataset for both within- and inter- modal retrieval. The experiments demonstrate that it outperforms several state-of-the-art methods in most scenarios.