Abstract
Why does human face identification mainly rely on the processing of the horizontally oriented cues? Is it because the face image contains most of its contrast in this range, and more particularly at the level of the eyebrows, eyes and mouth, features known to be crucial for face recognition? Alternatively, could the horizontal range of face information provide cues to identity that are more resistant to the drastic variations in face appearance resulting from e.g. changes in viewpoint? To disentangle these accounts, twenty-two human observers performed an identity recognition task with face stimuli presented one by one under different viewpoints (from full-front to profile views), and filtered to preserve contrast in selective orientation ranges. We found that the gaussian-shaped function relating the sensitivity to face identity to orientation peaked around the horizontal angle irrespective of viewpoint. A model observer performed a similar task by comparing each stimulus (pixel-wise) to the set of face images in different orientation ranges, within each viewpoint. Its performance suggests that image identity cues shift away from horizontal as viewpoint aversion increases. A second model observer performed the same task by comparing face images in selective orientation ranges, but across viewpoints. The performance of this model observer kept sharply tuned to the horizontal range, no matter the viewpoint of the stimulus. This result supports the notion that the horizontal orientation range of the face image conveys cues to identity that are more stable across viewpoint variations. The comparison of model and human performance suggests that human face identification tunes to horizontal orientation because this range conveys the richest feature cues in full-front views while it provides the most stable identity cues when moving away from full-front. These results yield novel insight on the strategies developed by the human visual system to achieve invariant face recognition.