search

caizhongang / SMPLer-X

Official Code for "SMPLer-X: Scaling Up Expressive Human Pose and Shape Estimation"
https://caizhongang.github.io/projects/SMPLer-X/
Other
1.01k stars 73 forks source link

Very big MAE SMPL-X Translation between GT and Predicted #54

Open GhostLate opened 8 months ago

GhostLate commented 8 months ago

Thank you for your research.

I trained model with custom dataset, it has low PA-MPVPE and PA-MPJPE errors, but absolute mean translation error is about [0.5, 0.5, 40] XYZ... Why does predicted SMPL-X translation have such a big difference compared to the translation of gt smplx translation on test dataset run?

Because of it I can't correctly build a body model in camera local coordinates (3D).

I think, these issues are related: #45 #41

Wei-Chen-hub commented 7 months ago

Hi, thanks for your interest in our project.

There's not enough information to identify the exact issue. In my opinion, the problem may be from the different camera scale, or camera space. I have a few insights to share.

  1. We assume the camera focal is 5000 (in bounding box space) during inference. Thus the transl might not the same scale with original camera parameters.
  2. During training, the data (smplx translation) is transformed into camera space.

I would also assign this issue to my colleague who was in charge of the training.