Large Scale Metric Learning for Matching of Heterogeneous Multimedia Data

Heterogeneous multimedia data are widely encountered in many applications, such as photo-sketch face recognition, still image to video face recognition, cross-modality image synthesis, cross media retrieval, etc. With the ubiquitous use of digital imaging devices, mobile terminals and social networks, there are lots of heterogeneous and homogeneous data from multiple sources, e.g., news media websites, microblog, mobile phone, social networking, etc. Matching of heterogeneous multimedia data becomes increasingly important to achieve cross modal and cross media information retrieval. One popular approach to the matching of heterogeneous data is metric learning, which learns a positive semi-definite matrix to measure the similarity of heterogeneous data. However, there are several challenging issues that the current metric learning methods cannot adequately address. First, the current metric learning methods are limited in dealing with the highly diverse and complex data types in real-current methods have poor scalability, which is a critical issue in handling the tremendous amount of multimedia data. Third, the labels of most data are unavailable, making them difficult to be used by current metric learning methods. Fourth, the data in the same modality may have different representations, and thus a multiple feature