Face-Front Occlusion Mask
You should use some deep learning to create a matte which essentially generates a "Face"/"Isn't Face" mask for rendering geometry. Using 3D Geometry works fine for general side-of-head occlusions but once the face turns more than 20-30 degrees generally it's the cheeks and nose which need to occlude, and ideally occlude to a pretty pixel accurate degree. We don't need a live 3D geometry generation that's accurate, but I bet a Face/NotFace mask would work fine just to separate the face profile from background geometry.
Hi Gavin! That's a cool suggestion! I will add it to our feature request list.
With the release of TensorFlow Mobile this should be fairly simple. The only question then is the complexity of its integration in your overall application and how would it affect the performance on users' devices.
Cheers,
Luka