Video-based face recognition using Earth Mover's Distance

In this paper, we present a novel approach of using Earth Mover's Distance for video-based face recognition. General methods can be classified into sequential approach and batch approach. Batch approach is to compute a similarity function between two videos. There are two classical batch methods. The one is to compute the angle between subspaces, and the other is to find K-L divergence between probabilistic models. This paper considers a most straightforward method of using distance for matching. We propose a metric based on an average Euclidean distance between two videos as the classifier. This metric makes use of Earth Mover's Distance (EMD) as the underlying similarity measurement between two distributions of face images. To make the algorithm more effective, dimensionality reduction is needed. Fisher's Linear Discriminant analysis (FLDA) is used for linear transformation and making each class more separable. The set of features is then compressed with a signature, which is composed of numbers of points and their corresponding weights. During matching, the distance between two signatures is computed by EMD. Experimental results demonstrate the efficiency of EMD for video-based face recognition.