smpl fit with MPII Human Shape model

[LSQsjtu]

Thank you for your valuable efforts in developing this work. I would like to inquire if there is a straightforward method to transferthe MPII Human Shape model to the SMPL model?

Looking forward to your reply.

[DavidBoja]

You could try to manually mark the landmarks on the MPII body (at least approximately) and then fit the SMPL to it, in order to find correspondences. You could play around with the loss weights and slowly diminish the landmark weight since they are approximate.

[LSQsjtu]

Thanks for your reply. I’m wondering if there are any easy tools for annotating landmarks. Specifically, are there any tools to get the landmark index of the MPII Human Shape to fit all instances of the MPII Human Shape into smpl?

[DavidBoja]

I’m wondering if there are any easy tools for annotating landmarks

There probably are, but I'm not really familiar with them. I used plotly to visualize SMPL and manually mark the CAESAR landmarks. I did it with a similar function as the one in visualize_smpl_landmarks in visualization.py. When you hover over a point, it should tell you the SMPL index for it. You could reuse the code and visualize the MPII Human Shape template to mark the landamrks. Once I had the initial SMPL landmarks, I fitted the SMPL to the whole CAESAR dataset scans. Then, I analyzed if the landmarks I annotated manually actually get fitted onto the ground truth landmarks of the CAESAR scans - if not, I improved the manual marking to make the landmark closer to where it was fitted on the scan.

fit all instances of the MPII Human Shape into smpl

I guess if you had the landmarks on the MPII template, you could use this codebase. Your scans, in that case, would be the MPII templates onto which you would fit the SMPL.