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

  • 2021-09-30 13:07:04
  • Erroll Wood, Tadas BaltruĊĦaitis, Charlie Hewitt, Sebastian Dziadzio, Matthew Johnson, Virginia Estellers, Thomas J. Cashman, Jamie Shotton
  • 60

Abstract

We demonstrate that it is possible to perform face-related computer vision inthe wild using synthetic data alone. The community has long enjoyed thebenefits of synthesizing training data with graphics, but the domain gapbetween real and synthetic data has remained a problem, especially for humanfaces. Researchers have tried to bridge this gap with data mixing, domainadaptation, and domain-adversarial training, but we show that it is possible tosynthesize data with minimal domain gap, so that models trained on syntheticdata generalize to real in-the-wild datasets. We describe how to combine aprocedurally-generated parametric 3D face model with a comprehensive library ofhand-crafted assets to render training images with unprecedented realism anddiversity. We train machine learning systems for face-related tasks such aslandmark localization and face parsing, showing that synthetic data can bothmatch real data in accuracy as well as open up new approaches where manuallabelling would be impossible.

 

Quick Read (beta)

loading the full paper ...