Item

Abstract

The ability of observers to recognize faces across changes in viewpoint has been found previously to decline as the angle between learn and test view grows. We replicate this result, but show that the recognizability of individual faces across changes in viewpoint cannot be characterized adequately in this way. An analysis of the recognizability of individual faces in different viewpoint transfer
conditions indicated that the representations of faces we make from different views may be only modestly related. We propose and implement a two-stage computational model of face recognition following Lando and Edelman (1995) that operates by: (a) transforming a previously unknown view of a face into other views, one of which may be known; and by (b) recognizing the transformed face views via an interpolation process that operates directly on the structure of the similarity relationships among faces. By combining model and human measures at the level of individual faces, using factor analysis, we isolate relatively consistent orthogonal loading patterns for three
learn view conditions (full, three-quarters, profile) with respect to the human and model measures. The axes stemming from the factor analysis are interpretable in terms of recognition transfer success based on: (a) large scale distinctive global features such as head shape, (b) smaller scale distinctive features such as large noses; and
(c) distinctive skin tone.