5±2 eigenimages suffice: an empirical investigation of low-dimensional lighting models

Abstract

Recently, P.W. Hallinan (1994) proposed a low dimensional lighting model for describing the variations in face images due to altering the lighting conditions. It was found that five eigenimages were sufficient to model these variations. This report shows that this model can be extended to other objects, in particular to those with diffuse specularities and shadows. We find that sharp specularities and shadows cannot be well represented by a low dimensional model. However, both effects can be adequately described as residuals to such a model. We can deal with occluders in a similar way. We conclude that low dimensional models, using 5±2 eigenimages, can be usefully extended to represent arbitrary lighting for many different objects. We discuss applications of these results to object recognition