search

akanazawa / hmr

Project page for End-to-end Recovery of Human Shape and Pose
Other
1.54k stars 395 forks source link

Training the discriminator #154

Closed zhangy76 closed 3 years ago

zhangy76 commented 3 years ago

Hi,

Thanks for proposing this classic work! When training the model only using 2D keypoints and pose prior, I wonder how many epochs do we need so that the model can output realistic body reconstruction? Also, why don't we need an Sigmoid function on the output of discriminator?

Thanks, Yufei

akanazawa commented 3 years ago

Thanks Yufei. I forget how long really,, but in my recollection realistic poses start happening very quickly. From then on, is a matter or training the regressor so that reprojection loss is minimized. We don't need this bc we use Least-Squared GAN objective for the discriminator.

Best,

A

amituofo1996 commented 2 years ago

Hi, I am confused about which dataset is used in training discriminators, is h3.6m moshed data, or mpi-inf-3dhp dataset ? Thanks!

akanazawa commented 2 years ago

It is the moshed datasets, bc the data has to be in SMPL format. moshed h36m is not available anymore, but the other datasets still are (through link on readme) and you can also use AMASS https://amass.is.tue.mpg.de/

Best,

A

On Mar 4, 2022, at 12:45 AM, amituofo1996 @.***> wrote:

 Hi, I am confused about which dataset is used in training discriminators, is h3.6m moshed data, or mpi-inf-3dhp dataset ? Thanks!

— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you modified the open/close state.