Multi-view dimensionality reduction via subspace structure agreement

Abstract

Many state-of-the-art biometric systems make use of high-dimensional features to represent data samples, which often contain substantial amount of intra-class variation. In this paper, we propose an unsupervised multi-view dimensionality reduction approach to extract discriminative low-dimensional features and apply it to multi-modal biometric retrieval problems. The proposed approach is based on a novel concept referred to as multi-view subspace structure agreement, which aims to learn a subspace projection for each view, such that the k-nearest-neighbour similarity graphs built in subspaces of different views are maximumly compatible. The proposed method is unsupervised in nature, but exhibits high discriminative power and is thus well suited to applications where class labels are generally unavailable such as retrieval and clustering. We evaluate the performance of the proposed algorithm under an audio-visual speaker retrieval experiment, as well as a multi-feature face retrieval experiment. Experimental results show that the retrieval performance of the proposed approach out-performs other competing methods with a significant margin.