Fully automated facial picture evaluation using high level attributes

People automatically and quickly judge a facial picture from its appearance. Thus, developing tools that can reproduce human judgments may help consumers in their picture selection process. Previous work mostly studied the position of facial keypoints to make predictions about specific traits: trustworthiness, likability, competence, etc. In this work, high level attributes (e.g. gender, age, smile) are automatically extracted using 3 different tools and are used to build models adapted to each trait. Models are validated on a set of synthetic images and it is shown that using attributes increases significantly the correlation between human and algorithmic evaluations. Then, a new dataset of 140 images is presented and used to demonstrate the relevance of high level attributes for evaluating faces with respect to likability and competence. A model combining both facial keypoints and attributes is finally proposed and applied to picture selection: which picture depicts the most likable face for a given person?