Purchase this article with an account.
Samuel Anthony, Walter Scheirer, Ken Nakayama; Judgments of Personality Traits from Real World Face Images. Journal of Vision 2014;14(10):1280. doi: 10.1167/14.10.1280.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Oosterhof & Todorov (2008) have shown the surprising reliability of human judgments of personality from faces (traits like trustworthiness and dominance). However, the bulk of this work has relied on in-lab ratings of well-controlled stimuli with a high degree of similarity (including all frontal poses). We used a large web sample (using TestMyBrain.org) for judgments on a large, heterogeneous set of faces that allows us to explore the perception of trustworthiness and dominance in real world contexts, as well as developing a computational framework based on the details of human performance (Anthony et al., 2013). Subjects (N=9899) made pairwise comparisons of faces along three dimensions: dominance, trustworthiness, and age. 30% of the images in a block were sampled from a set of 66 neutral frontal face images previously rated for dominance and trustworthiness by Todorov. 70% were sampled stochastically from 13,971 face images collected from Flickr. These uncontrolled images are maximally variable with respect to pose, expression, lighting, occlusion and makeup. We examined reliability of the ratings via two methods. The ratings of the Todorov faces correlated essentially perfectly with Todorov's prior ratings. On the Flickr set we found (by splitting the dataset and looking at correlations between the two halves) 4422 images (those with more than thirty comparisons per dimension) had highly reliable ratings (average correlations of 0.9 and above). These images comprise our set for analysis and algorithm training. This work confirms that reliable human judgment of personality traits is maintained across a set of maximally variable images, opening the door to explore other determinants of these traits as well as providing data for machine vision applications. For example, we show that trustworthiness and dominance show characteristic relations to perceived age. Also, these images provide a sufficient training/test sample for algorithms to mimic human judgments of these traits.
Meeting abstract presented at VSS 2014
This PDF is available to Subscribers Only