It is important to look at the combined effects of lighting and view direction on face recognition, as both depend on the three dimensional shape of the face and are in some ways analogous. For example, both pitch rotations and the change between top and bottom lighting involve rotations about the left-right axis, just as yaw and left/right lighting involved rotations about the vertical axis. We investigated identity matching performance across 45° rotations of the light source or the face about pitch or yaw with all images shown either upright or inverted. Upright images were better matched than inverted images, showing that image differences are not sufficient to explain performance. This is consistent with the idea that the use of class-based knowledge of upright faces is important for this task. The effects of face and light rotation interacted and depended on the axis of rotation. Overall, rotations about pitch impaired performance more than rotations about yaw, and changes in view impaired performance more than changes in light. The interaction between face and light rotation was such that if both rotated in the same direction, equivalent to a rotation of the camera in the opposite direction, performance was as good or better than if the face alone rotated. This was particularly pronounced for pitch rotations, with a clear pitch up advantage but only when the light rotated in the same direction. The results are interpreted in terms of image, surface and three-dimensional models based accounts of face matching.
Available at: http://works.bepress.com/hhill/2/