Semi-supervised Distance Consistent Cross-modal Retrieval.

Most of existing cross-modal retrieval approaches only exploit labeled data to train coupled projection matrices for supporting retrieval tasks across heterogeneous modalities. However, the valuable information involved in unlabeled data is unfortunately ignored. In this paper, we propose a novel Semi-Supervised Distance Consistent method (SSDC) to solve the problem. Our approach firstly models the initial correlation between different modalities by constructing the pseudo label and corresponding data of unlabeled query. Then our method learns projection matrices by adaptively optimizing the pseudo label of unlabeled data. In this way, SSDC could learn discriminative projection matrices. Experimental results on two publicly available datasets demonstrate the superior performance of the proposed approach.