Rank beauty

Abstract

It is useful to automatically select the most attractive face images from large photo collections. Previous works in this area have little attention on facial attractiveness for one subject, but different objects. In this paper, we have a collection of subjects’ faces including a range of expression, postures, makeup, lighting and resolutions from Bing Search. Given training data of faces scored based on the majority of subjects’ tastes, we train a model to learn how to rank novel faces and show how it can be used to automatically mine attractive photos from personal photo collections. Our system achieves an average accuracy of 73 % on pairwise comparisons of novel faces.