pose prior loss with SMPL-X model

[yyvettey]

Hi, I have some questions regarding the pose prior loss line385

        # Calculate the loss from the Pose prior
        if use_vposer:
            pprior_loss = (pose_embedding.pow(2).sum() *
                           self.body_pose_weight ** 2)
        else:
            pprior_loss = torch.sum(self.body_pose_prior(
                body_model_output.body_pose,
                body_model_output.betas)) * self.body_pose_weight ** 2

The body_pose_prior takes input of both body_pose and betas if use_vposer==False. However, the pose prior is trained on the AMASS dataset where the SMPL model is used. Since SMPL and SMPL-X do not share shape parameters, will it cause problems when I use pose prior with the SMPL-X model?

[geopavlakos]

In this case, the shape parameters betas are given as an input argument, but if you dig deeper into the code, you will see that they are not used in the loss computation. You could potentially define a body pose prior that takes the shape parameters into consideration too, but the priors used in the code (i.e., VPoser and GMM) depend only on the body pose parameters.

[yyvettey]

Thank you for your reply!!

May I know where I can find the GMM pkl files? Seems that they are not included in this repo.

[dimtzionas]

The GMM is a contribution of SPMLify [1] and not SMPLify-X [2]. This repo offers using the PKL as a non-default option for the sake of completeness and comparisons.

For downloading the PKL file that you are interested in:

• Please go to http://smplify.is.tue.mpg.de
• Create an account and agree with the license
• Download SMPLIFY_CODE_V2.ZIP (first button in the Downloads section)
• Unzip and find the file smplify_public/code/models/gmm_08.pkl
• If you find the PKL files useful for your work, please cite [1]

[1] Bogo et al. "Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image". ECCV 2016
[2] Pavlakos et al. "Expressive Body Capture: 3D Hands, Face, and Body from a Single Image". CVPR 2019