Question on normal sampling

[mks0601]

Hi,

Thanks for sharing your nice work!
I have a question about the normal sampling of the SMPL shape paramter.
In here, you add delta_betas with mean_shape, but the delta_betas is already summed with mean_shape at the above line.
Is this a mistake or the detla_betas is calculated by adding mean_shape?

Also, could you let me know how did you get the mean_shape and std_vector?

[akashsengupta1997]

Hi,

Ah yes that is a bug that I forgot to fix in the code release, thanks for reminding me!

mean_shape is the same mean shape vector used by SPIN (https://github.com/nkolot/SPIN).

std_vector is basically a hyperparameter that can be set as desired, to control the variation in body shapes seen during training. I use std_vector = [1.5, 1.5, ...., 1.5] for most experiments. I also messed around with trying to get SMPL to generate and predict really muscular people, which was quite challenging and required std_vector = [1.5, 1.5, ...., 10, 10] - I guess since muscular people are very out-of-distribution compared to the SMPL training set! Anyway, the point is that std_vector can be set as desired depending on what you want to model.

[mks0601]

Awesome. Thanks!

[melih-unsal]

Hi @akashsengupta1997, is there any way you had followed to find the std_vector hyperparameter when you need to apply your model to the people in a specific distribution or a country, or you had only tried till you found it?