search

davrempe / humor

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

Wrong coordinate system in motion_optimizer.py #52

Open linjiangya opened 2 months ago

linjiangya commented 2 months ago

Hi,

I am trying the experiments on AMASS dataset and found that the code here might be wrong:

https://github.com/davrempe/humor/blob/fc6ef84f0baa153be15427402e0147ed1a63a11a/humor/models/humor_model.py#L843

the past_t is the input to the sample_steps(). It is said to be in the local coordinate system as mentioned here:

https://github.com/davrempe/humor/blob/fc6ef84f0baa153be15427402e0147ed1a63a11a/humor/models/humor_model.py#L795-L796

However, the initial_input_dict is still in world coordinate system as it is not yet fed into apply_world2local_trans() Did I miss something?

THanks