search

statho / ScoreHMR

ScoreHMR: Score-Guided Diffusion for 3D Human Recovery (CVPR 2024)
MIT License
367 stars 24 forks source link

Betas normalization #5

Closed smandava98 closed 3 months ago

smandava98 commented 3 months ago

Hey! Great work!

I just had one question:

Where did you get the min and max values to normalize the betas from? I see the file name is betas_stats_eft_fits.npz. Where is this initially derived from? Did you calculate this or?

smandava98 commented 3 months ago

@statho

statho commented 3 months ago

The min and max values for the SMPL betas were calculated from the pseudo ground-truth betas of the training datasets. The idea behind this normalization is to put the SMPL betas in [-1, 1] range, so that we can use common variance schedules for the diffusion process. The SMPL thetas are naturally within that range.

We found that using unnormalized SMPL betas works well too!