Fake It Till You Make It: Face analysis in the wild using synthetic data alone

We demonstrate that it is possible to perform face-related computer vision in\nthe wild using synthetic data alone. The community has long enjoyed the\nbenefits of synthesizing training data with graphics, but the domain gap\nbetween real and synthetic data has remained a problem, especially for human\nfaces. Researchers have tried to bridge this gap with data mixing, domain\nadaptation, and domain-adversarial training, but we show that it is possible to\nsynthesize data with minimal domain gap, so that models trained on synthetic\ndata generalize to real in-the-wild datasets. We describe how to combine a\nprocedurally-generated parametric 3D face model with a comprehensive library of\nhand-crafted assets to render training images with unprecedented realism and\ndiversity. We train machine learning systems for face-related tasks such as\nlandmark localization and face parsing, showing that synthetic data can both\nmatch real data in accuracy as well as open up new approaches where manual\nlabelling would be impossible.\n