Kernel Grassmannian distances and discriminant analysis for face recognition from image sets

We address the problem of face recognition from image sets, where subject-specific subspaces instead of image vectors are compared. Previous methods based on Grassmannian subspace distances mainly take linear subspaces as input. The non-linearity exists when the input data contain complex structure such as pose changes. We generalize Grassmannian distances into high dimensional feature space with kernel trick to handle the underlying non-linearity in data. We show that kernel Grassmannian distances in feature space can be implicitly computed from the input data. Furthermore, we propose to use projection kernel in feature space for discriminant analysis. Comparisons with several state-of-the-art methods were performed on two databases, CMU PIE and YaleB. The proposed methods have demonstrated promising performance.