Why not train a non-linear landmark estimator?

[hallwaypzh]

[universome]

We wanted to show that for our model, the geometry information is stored in a very "disentangled" manner in the latent code. If your keypoints estimator is very powerful (in the extreme case it can be equivalent to [generator + off-the-shelf keypoints detector on top of it]) then it will also be able to recover the geometry information, but it will not say anything about how difficult it is to manipulate this information by editing a latent code (since it might be very entangled)