search

wangsen1312 / unsupervised3dhuman

58 stars 5 forks source link

Depth demo #9

Closed DavidBoja closed 2 years ago

DavidBoja commented 2 years ago

Hi,

I ran the demo you provdied for fitting the SMPL model onto a partial scan:

python generate_depth.py --filename ./demo/demo_depth/shortshort_flying_eagle.000075_depth.ply --gender male

but the surface_EM_depth minimization outpus NaN s for the demo example you provided. Does the code work for you?

Thanks in advance, David

DavidBoja commented 2 years ago

Even though there is a known issue for LBFGS giving NaNs (https://github.com/pytorch/pytorch/issues/5953), I think the issue here results from the smpl model which cannot handle the "crazy" inputs given by LBFGS. That is at least where the first NaNs start appearing when running the surface_EM_depth optimization step.

Did you have the same issues when evaluating on the CAPE / CMU / FAUST datasets?

wangsen1312 commented 2 years ago

@DavidBoja Actually I met the same problem before. From my understanding, two things may trigger this issue.

  1. the first step's result is so bad, the optimize function can not work very well.
  2. LBFGS lr can trigger this, when you change it smaller, the NaNs things will disappear. However, the small lr may not get very good results sometimes.

In my evaluation, if I met the problem, I will log it and use other set of optimize parameters to run it again.