densebody_pytorch
densebody_pytorch copied to clipboard
Lotayou
about the last fitting stage
Can you provide the code about how to get SMPL parameter from the resampled mesh?
@ZephirGe You can use Bodynet's code to do this. https://github.com/gulvarol/bodynet/blob/master/fitting/fit_surreal.py
@ypflll The Bodynet's code is about fitting SMPL to the volumetric results. As for fitting SMPL to UV maps, we should first transform UV map to SMPL-type vertices, right? Do you know how to do this transformation? Thanks!
@gaizixuan0128 You can resample smpl vertices from UV map by interpolation. Or you can directly use all valid vertices on the UV map(about 50k) as the fitting input, and this will make the fitting slower.
You can resample smpl vertices from UV map by interpolation. Or you can directly use all valid vertices on the UV map(about 50k) as the fitting input, and this will make the fitting slower.
thanks！I‘ll try it！
@gaizixuan0128 Have you tried the BodyNet fitting code? I've resampled 6890 smpl vertices, the fitting code also needs the 3D joints as input. Right?
@willie1997 I think it's right. The fitting code in the bodynet takes the 3D joints as input which could be regressed from the resampled smpl vertices.
I have another question. I find that the fitting stage introduced in the paper needs scale s, rotation R, translation t. So do we still need to optimize theta,beta,s,R,t jointly if follows the densebody paper?
@Rhyssiyan I feel confused too. The objective function written in the paper uses the smpl model after s, R, t. The resampled result is the transformed vertices without the parameter s,R,T. But the output doesn't include the s,r,t. As far as I know, smpl model has a parameter for global rotation and trans and the 1st PCA of shape is related to the scale change. Maybe these are inlucded in the theta,beta.
@ypflll How can we get the s,r,t used in the fitting stage? Or just add them to the param list when optimize.
@ypflll which means if the input is a new image taken by myself without the cameara paras, we cannot use the fitting stage?
@willie1997 Absolutely you can optimize by adding extra camera paras, but I didn't tried this. Maybe you can take a look at smpl-x or smplify-x.
Hi,@ypflll what do you mean by camera paras? Is it camera intrinsics? if it's camera intrinsics, could you explain how to infer s,R,t from camera intrinsics? In my understanding s,R,t are obtained by aligning the predicted joints with the ground truth 2d,3d joints.