In this paper, we investigate the contribution of dynamic evolution of 3D faces to identity recognition. To this end, we adopt a subspace representation of the flow of curvature-maps computed on 3D facial frames of a sequence, after normalizing their pose. Such representation allows us to embody the shape as well as its temporal evolution within the same subspace representation. Dictionary learning and sparse coding over the space of fixed-dimensional subspaces, called Grassmann manifold, have been used to perform face recognition. We have conducted extensive experiments on the BU-4DFE dataset. The obtained results of the proposed approach provide promising results.