Positions/rotations on input and whether whole animation orientation matters

[dj-kefir-siorbacz]

Hi!

I was wondering, whether the orientation of animation matters for training performance - i.e. if it matters whether the head/chest/hips joints are oriented towards the global "forward" axis (for example Z axis). I was thinking of rotating all my animations to face this common global "forward", but it's tricky and in the end I don't know if it'll even make a difference.
Let me know!

Also, what's not clear for me: what kind of representation is being used as input to the network - joint positions or rotations? In the paper it's not clear. If I can use rotations only, then my abovementioned issue with global "forward" is not a problem.

[GuyTevet]

We use the SMPL (thetas only) or the HumanML3D representations, depending on the dataset. Training should work without the front-facing trick. You should try and see.