questions about the main paper

[maolingao]

Hi Garvita,

I have 2 questions when reading the paper, can you please help?

1. Does the learned pose manifold depend on the identity of a person, i.e. it needs to be retrained for a different person?
   Since the input of the network is SO(3)^k poses, I would say that poseNDF shouldn't depend on the person's identity. However, for person with large shape variation, e.g. tall/small, fat/slim, there may be differences in the respective SO(3)^k, even if these different persons are in the exact same pose.

2. The Eqn. 6 projects a random SO(3)^k pose to a valid SO(3)^k pose on the learned pose-manifold. Why a re-projection to the quaternion sphere is still required? Does this suggests that the manifold contains some invalid poses, or is it just minor numerics?
   (see the line below eqn. 6. https://virtualhumans.mpi-inf.mpg.de/papers/tiwari22posendf/posendf_paper.pdf)

thanks for your clarification and the nice paper, I enjoyed a lot reading it.

Best,
Maolin