Abstract
That human face processing rests mainly on configural cues is the contemporary zeitgeist in the face recognition literature (e.g., Maurer, Le Grand, & Mondloch, 2002). Surprisingly, this hypothesis is supported by rather weak empirical evidence. We will not critically review the evidence here. Instead, we shall focus on two crucial but neglected points: If face processing is mainly configural then surely (1) real-world faces must vary configurally, and (2) human observers must be sensitive to these configural variations for faces over a range of realistic sizes. To address (1), four participants (to estimate between-subject error) each annotated 515 Caucasian portraits on 20 internal facial landmarks. All participants re-annotated 10% of these portraits, to estimate intra-subject error. One participant also annotated 309 Asian portraits in the same fashion. We translated, rotated, and scaled all face annotations to minimize the mean square of the distance between these annotations and their average. The differences between post-alignement face annotations reflect the configural differences between these faces. Ideal observers performed only modestly in gender, race, expression, and identity discrimination tasks on the basis of aligned face annotations. Next we turned to (2): Are human observers sensitive to this modest configural information available in real-world faces over a range of realistic sizes? Ten participants were submitted to a matching-to-sample (ABX) task. Each pair of stimuli were produced by applying two randomly chosen aligned Caucasian face annotations to a randomly chosen, aligned Caucasian portrait, preserving intact its parts but altering its configural cues. Participants completed the task at multiple viewing distances. Preliminary results indicate that human observers are capable of using real-world configural cues only at close range. We will apply multiple linear regression to the face annotations and response accuracy to quantify the importance of individual facial landmarks.