CMML: a New Metric Learning Approach for Cross Modal Matching
Alexis Mignon, Frédéric Jurie · 2012
Abstract. This paper proposes a new approach for Cross Modal Matching, i.e. the matching of patterns represented in different modalities, when pairs of same/different data are available for training (e.g. faces of same/different persons). In this situation, standard approaches such as Partial Least Squares (PLS) or Canonical Correlation Analysis (CCA), map the data into a common latent space that maximizes the covariance, using the information brought by positive pairs only. Our contribution is a new metric learning algorithm, which alleviates this limitation by considering both positive and negative constraints and use them efficiently to learn a discriminative latent space. The contribution is validated on several datasets for which the proposed approach consistently outperforms PLS/CCA as well as more recent discriminative approaches. 1