About prior token

[ltphuongunited]

Thank you for the great work.

I wanted to ask if it seems that 'Prior tokens' are not initialized from SMPL_MEAN_PARAMZ but are initialized based on a normal distribution. This may result in non-physical poses and not leverage the SMPL template mesh. I am a bit curious about the use of the term 'prior' here. Looking forward to hearing from you soon.

[yangsenius]

Thanks for your attention. These "prior tokens" are initialized with a normal distribution but also constantly updated during training. Finally, these tokens learn some kind of prior knowledge about plausible human poses as an initial state, so we call them 'prior tokens'. Yes, it's challenging to estimate a reasonable human pose when not leveraging the SMPL template. Effective training is necessary to ensure it.