Generalization to Novel Views of Faces: Psychophysics and Models concerning the Role of Bilateral Symmetry

Human faces are approximately bilaterally symmetric. We study the ability to generalize to novel views of human faces focusing on the role of that symmetry. Our hypothesis is that the ability to identify mirror symmetric images is used for viewpoint generalization by approximating the symmetric view of a learned view by its mirror symmetric image. Two psychophysical experiments are performed using a same/different paradigm. Experiment 1 shows that generalization to the symmetric view is better than generalization to otherwise different views. If the symmetric view is replaced by the mirror reversed learning view, performance further increases. Experiment 2 shows that the match between the learned view and the testing image is performed directly on the level of the images. Performance drops significantly if the symmetry between the intensity patterns of learning and testing view is disturbed by an asymmetric illumination, although the symmetry between the spatial arrangement of high-level features is retained. We show that a simple image-based model can explain important aspects of the data and we show how this model can be extended towards a general algorithm for image comparison. Experimental results are discussed in terms of their relation to existing approaches to object recognition.