Recognition of unfamiliar faces can be severely impaired when two pictures of a face are taken from different camera distances. This effect of perspective transformation may be predicted either by a model-based theory, for which the impairment shows a difficulty in constructing the 3D surface geometry, or by an image-based theory, for which it results from dissimilarity between 2D images. To test these hypotheses, we tested recognition performance in which face images were viewed either at their center of projection (the camera position) or at other distances. Based on past findings [11, 17], the same position should help correct marginal distortions of shapes due to large perspective convergence and hence facilitate reconstruction of 3D shape from perspective cues. However, the results showed little support for this prediction. The lack of 3D shape reconstruction and the effects of image similarity provided favorable evidence for the image-based theory of face recognition.