Whether support larger num_betas?

[Queequeg92]

In your code, num_betas is set to 10. But in some other papers, they have different num_betas.
If num_betas changes, should I retrain the models for whole body mesh recovery?
Thanks for your great work!

[moonbow721]

That's a great question, and thank you for your interest in our work!

The setting for num_betas is flexible in the whole-body mesh recovery task. We chose 10 simply because it is the most common value, but you are free to modify this parameter directly to fit your specific needs or dataset.

The reason you don't need to retrain is that our DPoser-X model only learns a prior for the pose parameters of the smplx body model. For the shape parameters (betas), we use a standard L2-loss as the prior and did not train a separate, learned model for it. Therefore, changing the number of shape parameters will not affect the performance of our pose prior.

I hope this helps!