A Computational Model of the Trait Impressions of the Face for Agent Perception and Face Synthesis

This paper reports a first attempt at developing a computational model of the trait impressions of the face for embodied agents that accommodates the social perception and social construction of faces. Holistic face classifiers, based on principle component analysis (PCA), were trained to match the human classification of faces along the bipolar rating extremes of the following trait dimensions: adjustment, dominance, warmth, sociality, and trustworthiness. Although results were marginally better than chance in matching the perception of dominance (64%), classification rates were significantly better than chance for adjustment (71%), sociality (70%), trustworthiness (81%) and warmth (89%). A second exploratory study demonstrates how PCA models of trait classes could be used by agents to generate faces. Novel faces were synthesized by probing specific PCA trait attribution spaces. Human subjects were then asked to rate the synthesized faces along a number of trait dimensions, and it was found that the synthesized faces succeeded in eliciting predicted trait evaluations.