How should we represent faces for automatic recognition?

We describe results obtained from a testbed used to investigate different codings for automatic face recognition. An eigenface coding of shape-free faces using manually located landmarks was more effective than the corresponding coding of correctly shaped faces. Configuration also proved an effective method of recognition, with rankings given to incorrect matches relatively uncorrelated with those from shape-free faces. Both sets of information combine to improve significantly the performance of either system. The addition of a system, which directly correlated the intensity values of shape-free images, also significantly increased recognition, suggesting extra information was still available. The recognition advantage for shape-free faces reflected and depended upon high-quality representation of the natural facial variation via a disjoint ensemble of shape-free faces; if the ensemble comprised nonfaces, a shape-free disadvantage was induced. Manipulation within the shape-free coding to emphasize distinctive features of the faces, by caricaturing, allowed further increases in performance; this effect was only noticeable when the independent shape-free and configuration coding was used. Taken together, these results strongly support the suggestion that faces should be considered as lying in a high-dimensional manifold, which is locally linearly approximated by these shapes and textures, possibly with a separate system for local features. Principal components analysis is then seen as a convenient tool in this local approximation