Unsupervised multi-view dimensionality reduction with application to audio-visual speaker retrieval

This paper presents a novel approach to unsupervised multi-view dimensionality reduction and reports its application to multi-modal biometrics retrieval, specifically audio-visual speaker retrieval. We propose a new concept referred to as multi-view subspace agreement, which aims to learn a subspace for each view which respects the similarity relationships between data points in the other view. The proposed algorithm is unsupervised but exhibits discriminative characteristics and is thus well suited to applications such as retrieval and clustering where class labels are generally unavailable. The effectiveness of the proposed algorithm is evaluated under an audio-visual speaker retrieval experiment with the MOBIO database. The retrieval performance of the proposed approach out-performs other single-view or multi-view dimensionality reduction methods with a significant margin.