This work uses GANs to generate synthetic data to use for supervised training of facial recognition systems. More specifically, they use an image-to-image GAN to improve the quality of faces generated by a face simulator. The simulator is able to produce a wider range of face poses for a given face, and the GAN is able to refine the simulators output such that it is more closely aligned with the true distribution of faces (i.e. They show that by fine tuning a facial recognition system on this additional synthetic data they are able to improve performance and outperform previous state of the art methods. Pros: - This method is simple, apparently effective and is a nice use of GANs for a practical task.

Synthesizing realistic profile faces is promising for more efficiently training deep pose-invariant models for large-scale unconstrained face recognition, by populating samples with extreme poses and avoiding tedious annotations. However, learning from synthetic faces may not achieve the desired performance due to the discrepancy between distributions of the synthetic and real face images. To narrow this gap, we propose a Dual-Agent Generative Adversarial Network (DA-GAN) model, which can improve the realism of a face simulator's output using unlabeled real faces, while preserving the identity information during the realism refinement. The dual agents are specifically designed for distinguishing real v.s. In particular, we employ an off-the-shelf 3D face model as a simulator to generate profile face images with varying poses.