can smpl's beta directly be used in STAR?
wells-wei-wei opened this issue · comments
I have just used same beta parameters in SMPL and STAR, but they are totally different. whether I use it in a wrong way or not?
smpl_para=joblib.load("20200901_zhh_y_t.pkl")
device = torch.device('cuda')
for i in tqdm(range(len(smpl_para[1]['frame_ids']))):
pose=smpl_para[1]['pose'][i]
pose=pose[np.newaxis, :]
beta=smpl_para[1]['betas'][i]
beta=beta[np.newaxis, :]
poses = torch.cuda.FloatTensor(pose)
poses = Variable(poses,requires_grad=True)
betas = torch.cuda.FloatTensor(beta)
betas = Variable(betas,requires_grad=True)
trans = torch.cuda.FloatTensor(np.zeros((1,3)))
trans = Variable(trans,requires_grad=True)
d = star(poses, betas, trans)
f="%06d" % i
vertices=d.detach().cpu().numpy().squeeze()
faces=d.f
out_mesh = trimesh.Trimesh(vertices, faces)
out_mesh.export("star_obj/"+f+".obj")
Hi @wells-wei-wei , both STAR pose and shape space are different from SMPL, so we can't plug-in the SMPL parameters in STAR. I will commit a convertor script to convert from SMPL to STAR. Thanks,Ahmed
Hi @wells-wei-wei I just pushed a convertor script that given 'smpl meshes' returns and save the star parameters. It's under https://github.com/ahmedosman/STAR/blob/master/convertors/convert_smpl_to_star.py
Is there a way to regress the STAR parameters directly like HMR or VIBE? Thanks.
Hi @leeskyed, one would need to retrain the networks (HMR and VIBE) with STAR. At the moment, there isn't a direct regressor of parameters STAR from images/videos.