A major obstacle in modeling the mental representation of human faces is the lack of an appropriate parameterization of face stimuli. Methods that use pixilated face images result in parameterizations that are overly sensitive to viewing conditions such as illumination, alignment, and rotation in depth, while methods that rely on the location of landmarks on face images result in incomplete parameterizations that cannot be used to reconstruct the original images. Here we present a method that avoids the encoding problems of pixel-based parameterizations yet retains the power to reconstruct the original stimuli. Two-toned silhouetted face profiles are obtained from gray-scale profile images by thresholding at a median gray level and cropping the images at the forehead and neck. The parameterization consists of locating a set of 23 landmark points along the image contour, normalizing their coordinates across all faces, and subjecting these values to a principal components analysis (PCA). This process results in a 42-dimensional vector description of each face. Face silhouettes are then accurately reconstructed from the vectors by using bicubic splines to interpolate between adjacent points along the face contour. In four studies we confirm the psychological validity of silhouetted face profiles as face stimuli by obtaining people’s ratings of gender (Study 1), age (Study 2), attractiveness (Study 3), and distinctiveness (Study 4). We conclude by describing how a detailed mathematical description of face space based on face shape allows for the creation of realistic artificial face stimuli suitable for experiments on face memory and representation.

Thomas Griffiths, Michael Ramscar, Joshua Tenenbaum, and Daniel Yarlett