Large-Scale Supervised Hashing for Cross-Modal Retreival

Matrix factorization technique has become a promising approach for cross-modality hashing, where projections are learned to represent data from multiple interlinked sources in a single hamming space. In this paper we propose a novel method based on Collaborative Matrix Factorization Hashing (CMFH) that reduces the error found between multimodal data and their unified representation. The new method uses an iterative learning algorithm to initialize projection matrix. To achieve better search accuracy, we measure similarity using the cosine distance. Extensive experiments done on two different datasets prove that our method outperforms CMFH and several states of the art methods and achieves stable performance over code length variation.