Semi-Supervised Face Frontalization in the Wild

Synthesizing a frontal view face from single nonfrontal image, i.e. frontalization, is task of practical importance in wide range facial image analysis applications. However, to train the frontalization model supervised manner, most existing methods rely on availability nonfrontal-frontal pairs (typically Multi-PIE dataset) captured constrained environment. Such approaches, return, limit generalizability their application unconstrained scenarios. Unfortunately, although large amount in-the-wild datasets are available, they cannot easily be utilized for training since and images not paired. To network which generalizes well both environments, we propose semi-supervised learning framework effectively uses (labeled) indoor (unlabeled) outdoor faces. Specifically, achieve this goal, article presents Cycle-Consistent Face Frontalization Generative Adversarial Network (CCFF-GAN) consists (1) (2) unsupervised components. For (1), use paired data learn roughly accurate may generalize (in-the-wild) (2), cope with generalization issue, part unpaired under perceptual cycle consistency constraint semantic feature space controlled (indoor) uncontrolled (outdoor) Extensive experiments demonstrate effectiveness proposed method comparison state-of-the-art methods, especially