Triple-GAN - Progressive Face Aging with Triple Translation Loss.

Face aging is a challenging task which aims at rendering face for input with aging effects and preserving identity information. However, existing methods have split the long term into several independent groups and ignore the correlations of age growth. To better learn the progressive translation of age patterns, we propose a novel Triple Generative Adversarial Networks (Triple-GAN) to simulate face aging. Instead of formulating ages as independent groups, Triple-GAN adopts triple translation loss to model the strong interrelationship of age patterns among different age groups. And to further learn the target aging effect, multiple training pairs are offered to learn the convincing mappings between labels and patterns. The quantitative and qualitative experimental results on CACD, MORPH and CALFW show the superiority of Triple-GAN in identity preservation and age classification.