Age-Style and Alignment Augmentation for Facial Age Estimation

Abstract

Facial age is an important soft biometric trait for better identification of a human subject. The development of a facial age estimation system requires a large collection of age-labeled data. However, the imbalanced data distribution across age poses a major challenge to making a decent model to describe the variation of facial appearance caused by age. The cross-age data imbalance can be observed in common facial age datasets, for example, the MORPH [8], FG-NET [7] and the MIVIA dataset [3] considered in the GTA Contest. It can be often seen that insufficient data are provided for younger ages and senior ages, and the insufficiency becomes worsened as the age moves close to both ends. To deal with the data imbalance issues, many approaches implement various data augmentation schemes. In our approach, we propose a data augmentation scheme built upon the Age-Style GAN (ASGAN), which we propose for facial age regression and progression. In addition to the ASGAN-based data augmentation, we leverage the mean-variance loss to improve the age classification accuracy, and exploit face alignment as an auxiliary scheme to augment the whole dataset with an aligned subset. We conducted extensive experiments on the MIVIA dataset for verifying the performance of our approach.