Open davidpagnon opened 3 months ago
Might be related to this issue (not entirely sure about it): https://github.com/MarilynKeller/SKEL/issues/7
Hi David,
Definitely the second option is the most straightforward. From the bioamass dataset , we indeed trained a regressor that given the posed smpl vertices, predicts the opensim skeleton joints location.
I have been meaning to add a demo of this but did not find the time so far.
Here is how it works: https://github.com/MarilynKeller/SKEL/blob/af881e69f6d818f5e077c301777d0db73bb6fbdf/skel/skel_model.py#L301
skel_model = SKEL(gender)
J = torch.einsum('bik,ji->bjk', [v_shaped, skel_model.J_regressor_osim])
Where v_shaped are the posed SMPL mesh vertices sequence (Fx6080) and J (Fx24x3) are the regressed 24 anatomical joint location that correspond to the OpenSim model. F is the number of frames of the sequence.
Thanks, I can't right now but I'll check this in about 2 weeks!
Hi Marilyn,
I wonder if you have any perspective which of the following two approaches would be better for obtaining joint centers from an SMPL mesh:
skel_data = pkl.load(open(f'skel_{gender}.pkl', 'rb')
).If you think the second one is a good solution, do you know how I can use it to obtain joint centers from SMPL vertices?
Thanks!