Why I got the wrong global orient results

[jhkim0759]

I rendered the obj with pytorch3d library and fit the SMPL skinned model using the parameters provided. but, I got the results as above. I think this is upside down result. Can you help me to get the correct results?

[ytrock]

Dear jhkim0759,

can you first try to generate the corresponding SMPL mesh and save it as .obj file, and then check the alignment between the two obj meshes using the Meshlab?

[jhkim0759]

Thank you for your answer

I saved the mesh as obj and check between the data what you provided. but, I got the exactly opposite results as the image shown. I think I have to aligned the mesh one by one. or is there an other way to obtain the Global Orient from camera parameters(Entrinsic, Intrinsic, Rotation, etc.)?

[ytrock]

That's weird, when generating smpl-x meshes, did you use the provided poses which already include the global RT? Or you can paste your code for generating the smpl-x meshes at here.

[jhkim0759]

smplr = SMPLR()

params_dir = "../dataset/THuman/THuman2.0_smplx/0525/smplx_param.pkl"

db = np.load(params_dir, allow_pickle=True)
dict_ = {}
for key in db:
    dict_[key] = torch.tensor(db[key])
    
global_ori = torch.from_numpy(dict_["global_orient"])[0]
pose = torch.cat([global_ori, torch.from_numpy(dict_["body_pose"][0])],0)
verts = smplr(pose.reshape(1,-1,3,3).float(),torch.from_numpy(beta).reshape(1,-1).float(), pose2rot=False)["verts"]

renderer = Renderer(resolution=(1024, 1024), orig_img=True, wireframe=False )
img = renderer.render(img, verts[0], [1,1,0,0])

This the code what is used. The SMPLR Model is from ROMP, and Renderer is from PARE.

Thank you

[jhkim0759]

I think the SMPL regressor is based on vibe form. That's why the verts show the opposite. Thank you

[ytrock]

You are welcome :) Thanks for your attention.