davrempe / humor

Code for ICCV 2021 paper "HuMoR: 3D Human Motion Model for Robust Pose Estimation"
MIT License
518 stars 72 forks source link

why the prior loss can be negative #25

Closed AIML closed 2 years ago

AIML commented 2 years ago

Hello, thank you for this great work! In stage3, I found the motion prior and init motion prior loss can be negative. As they are log-likelihood, is this Normal ?

AIML commented 2 years ago

I also find that humor rollout sequence may be oversmooth. But as it's trainned from mocap dataset, the high dynamic action should be support?

AIML commented 2 years ago

Sorry, In the firt question, I have confuse with probability and probability density for continuous random variables. In the second question, as the optimizer just make a trade-off between different loss, when we estimate the high dynamic action, the loss of humor will be bigger than data term loss, which cause the over smooth. I have refine the 2D keypoint confidence estimate by openpose, in which, let confidence be 1.0 when it's bigger than 0.3.