Tri-space and ranking based heterogeneous similarity measure for cross-media retrieval

We study the problem of cross-media retrieval, where the query and the returned results are of different modalities. A novel method is proposed to measure the similarity between heterogeneous media objects for cross-media retrieval. While existing methods only focus on the original low level feature spaces or the third common space, our proposed tri-space explores both of the two kinds of spaces. On one hand, the low level feature spaces can reflect the original accurate information of each modality and the third common space can effectively explore the useful information hidden across modalities. On the other hand, combination of multiple spaces can lead to good results since we can fully use the rich information of tri-space. Moreover, we propose to use ranking orders to represent media objects. Ranking based similarity makes our proposed method less sensitive to actual distance values and thus more stable. Experiments on the Wikipedia dataset demonstrate the effectiveness of our approach.