Deep learning methods for 3d human pose estimation from RGB images require a huge amount of domain-specific labeled data for good in-the-wild performance. However, obtaining annotated 3D pose data requires a complex motion capture setup which is generally limited to controlled settings