Fisher׳s linear discriminant embedded metric learning

In this paper, we propose a novel distance metric learning model, which embeds Fisher׳s linear discriminant into the classical maximum margin criterion. Specifically, given pairs of similar samples and pairs of dissimilar samples, our metric learning model aims to maximize the margin between these two kinds of sample pairs while maintaining a large mean squared distance ratio. In this way, the learned model benefits from exploiting the distributions of sample pairs and thus becomes more reliable and effective. Furthermore, the optimization problem of our model can be solved efficiently by a proposed generic iterative approximate method. The effectiveness of our model is demonstrated on various datasets including a challenging face verification dataset called Labeled Faces in the Wild.