Sharon Duvdevani-Bar, Shimon Edelman, A. Jonathan Howell, Hilary Buxton
Human observers are capable of recognizing a face seen only once before when confronted with it subsequently under different viewing conditions. We constructed a working computational model of such generalization from a single view, and tested it on a homogeneous database of face images obtained under tightly controlled viewing conditions. The model effectively constructs a view space for novel faces by interpolating view spaces of familiar ones. Its performance - 30% error rate in one out of 18 recognition, and 8% in one out of three discrimination - is encouraging, given that it reflects generalization from a single view/expression to a range of +/- 34 degree rotation in depth and to two additional expressions. For comparison, human subjects in the one out of three task involving only viewpoint changes exhibit a 3% error rate.
Download compressed postscript file