Full-space local topology extraction for cross-modal retrieval.

Abstract

With the ever increasing availability of various kinds of multimedia data, cross-modal retrieval, which enables information retrieval from various types of data given various types of query, has become a research hotspot. Hashing-based techniques have been developed to solve this problem, however, most previous works cannot capture the shared underlying structure of real-world multimodal data, which degrades their retrieval performances. In this paper, we propose a novel hashing method based on the extraction of the common manifold structure shared among different feature spaces. To faithfully represent the common structure, two kinds of local topology information are exploited in our method. Local angles are incorporated within the extraction of local topology of each feature space, which is then used to learn a common intermediate subspace. After heterogeneous features being embedded into this subspace, local similarities are exploited to extract the local topology between different feature spaces, and learn compact Hamming embeddings to facilitate cross-modal retrieval. The proposed method is referred to as full-space local topology extraction for hashing. Extensive comparisons with other state-of-the-art methods on three benchmark multimedia data sets demonstrate the superiority of our proposed method in terms of retrieval recall and search accuracy.