What is the basis for good performance to symmetric views of faces?

Purpose. Recently, we investigated human performance to generalize to novel views of a learned face (Troje & Bülthoff, 1996, Vision Research, in press). Among other results, we made the observation that generalization to views that are symmetric with respect to the frontal view is much better than to otherwise different views. Here, we present new psychophysical experiments investigating the nature of this performance. In particular, our question is, whether this performance is based on the bilateral symmetry of the 3D-object or on the resulting mirror symmetry of the images. Methods. Two experiments were performed. Both used a SAME/DIFFERENT recognition paradigm in which two images of faces were shown in immediate succession. The subject then decided, whether or not the two images showed the same person. The images were made from 3D head models and showed the face without its natural texture, only applying a Lambertian shading model. This allows decoupling the symmetry of the view from the mirror symmetry of the image. Symmetric views yield mirror symmetric images only if the simulated light source in both images comes from the direction of the camera or if it is also symmetric with respect to this direction. However, if the light comes from the same side, the images taken from symmetric viewpoints are no longer mirror symmetric. In experiment 1 training and testing views where either identical or symmetric. The light direction was also either identical or symmetric, yielding four different conditions. In experiment 2 the light always came from the direction of the camera. Training and testing views were either identical, symmetric or otherwise different. In a fourth condition we used instead of the symmetric view the flipped, perfectly mirror symmetric image for testing. Results. Experiment 1. If both viewpoint and illumination were symmetric, performance was much better (p<0.005) than when only the view or the illumination changed. Experiment 2. Generalization to the perfectly mirror symmetric image was even better (p<0.05) than to the symmetric view. Conclusions. The better generalization to the symmetric view of a head model is not based on knowledge about the almost bilaterally symmetric three-dimensional structure of the head but rather on the simple image operation of identifying mirror symmetric images.