search

mkocabas / VIBE

Official implementation of CVPR2020 paper "VIBE: Video Inference for Human Body Pose and Shape Estimation"
https://arxiv.org/abs/1912.05656
Other
2.86k stars 550 forks source link

Motion discriminator loss #204

Open GuyTevet opened 3 years ago

GuyTevet commented 3 years ago

Thanks for your great work! It seems that the l2 disc loss is not bounded (e.g. by a sigmoid function), although in the paper it is stated that the output is in [0, 1]. Do I miss something?

https://github.com/mkocabas/VIBE/blob/c0c3f77d587351c806e901221a9dc05d1ffade4b/lib/models/motion_discriminator.py#L77

https://github.com/mkocabas/VIBE/blob/c0c3f77d587351c806e901221a9dc05d1ffade4b/lib/core/loss.py#L208

AIML commented 2 years ago

you can use sigmoid instead linear, but as sigmoid suffers saturation problem, it may not be btter.