search

ahmedosman / STAR

ECCV2020 - Official code repository for the paper : STAR - A Sparse Trained Articulated Human Body Regressor
https://star.is.tue.mpg.de
Other
654 stars 95 forks source link

Are arms unnaturaly long in STAR (SIMPL)? #37

Open dyollb opened 6 months ago

dyollb commented 6 months ago

If I remember correctly in the original CEASAR data hands are scanned as fists. I guess the registration process ignored the fit of hands to create a statistical shape model including hands with fingers spread apart.

I am trying to register the STAR model to a surface from the Virtual Population dataset, segmented from magnetic resonance images. I get reasonable fits, except for the hands. The arms seem to remain too long.

I measured the length in the mean model (female) and found the arms are 10% longer than those of Ella.

Has anybody else observed this?

dyollb commented 6 months ago

It seems the shape variation accounts for the arm length in higher modes. After increasing num_betas to > 60 I started to get the arms/hands to match. Sorry, should have tested that before. I will visualize the modes (shape_dirs) to understand which one(s) accounts for this difference.