Item

Abstract

Humans are experts in processing faces, which are encountered in everyday life in many dierent poses and views. In the present study we investigated to what extend the processing
of such highly overlearned stimuli is dependent on viewpoint. The experiments
were modeled after the inter-extra-ortho experiments by Bultho Edelman, 1992,
PNAS, 60-64, who used novel objects as stimuli (wire- and amoebae-like 3D objects).
In Experiment 1, ten face-stimuli were presented in frontal view and 60 side view.
At test they had to be recognized among ten distracter faces at dierent viewpoints.
We found systematic eects of viewpoint (recognition performance: inter > extra >
ortho), which were consistent with recognition schemes based on local view interpolation.
These eects were replicated in Experiment 2, in which frontal and 45 side views
were learned, although recognition performance was less viewpoint-dependent (inter =
extra > ortho). This result is again consistent with recognition schemes based on local
view interpolation. In Experiment 3, the learning condition entailed frontal views and
45 upward views. The motivation for this experiment comes from the fact that in
everyday life, faces are encountered more often in dierent side views than in views
from above and that faces are only vertically symmetric. Indeed, systematic dierences
of viewpoint-dependent performance were found in Experiment 3 when compared to
Experiment 1 and 2 (inter > ortho > extra). The results are discussed within the
framework of a new computational model based on key-frames, which entails local view
interpolation and has been shown to be well-suited to model human face recognition
performance (Wallraven et al., 2002, LNCS vol. 2525, 651-660).