Most researchers agree that faces are represented in a multidimensional face space with the “mean face” at the origin. It is often assumed that the component dimensions of this space involve appearance based features, but there is relatively little empirical data to support that view. Therefore, the dimensions of the face space are still unclear. A preferable approach is to systematically probe the face space to determine which types of information are utilized by the human recognition system. In particular, due to the parsimonious nature of human perceptual processes, sources of information that are invariant across a variety of conditions are good candidates for dimensions of the face space. The research described here was designed to address this issue in two experiments.
Experiment 1 included four image manipulations. The first condition was comprised of line drawings of faces from the AR face database. The second condition included images of faces manipulated with the threshold function in Adobe® Photoshop, leaving only black and white in the image. The third condition was similar to the second except that the faces were filled in with a mid-gray. In the final condition, the Fourier amplitude information was randomized, while the phase information was left unaltered. The second experiment systematically manipulated images of human faces using a variety of homeomorphic transformations, including a cardioidal strain, affine shear, a sinusoidal variation, and a local perturbation. Each type of transformation was applied with several different magnitudes.
In both experiments, the differences in image structure they produced were measured using several possible metrics involving either pixel intensities or wavelet outputs. Observers performed a match-to-sample task in which they viewed an image of a standard face, followed by transformed versions of the same face and a different face, and they were asked report which of the alternatives depicted the same individual as the sample image.
The results of the first experiment revealed that as information was removed, performance suffered, particularly with the removal of contrast polarity information (i.e. the difference between the threshold condition and the line drawing condition). Experiment 2 revealed that performance deteriorates with the magnitude of image change for each possible transformation type. Additionally, there were also clear differences among the different types of transformations that cannot be explained by simple differences in low level image structure. These findings suggest that observers' judgments may have been based on configural relations among facial features that can remain relatively invariant over some types of transformations, but not others. Future directions for more sophisticated computational modeling and important methodological considerations for face recognition research are discussed.