Clarify output of loading dataset

IDEA-Research / OSX

[CVPR 2023] Official implementation of the paper "One-Stage 3D Whole-Body Mesh Recovery with Component Aware Transformer"

https://osx-ubody.github.io/

MIT License

655 stars 57 forks source link

Clarify output of loading dataset #96

Closed longphamkhac closed 10 months ago

longphamkhac commented 10 months ago

Hi, tks for your outstanding job!!! I'm new to 3d human recovery. I'm reading your loading MPII dataset function, can I ask what are the differences between smplx_joint_img and smplx_joint_cam ?

linjing7 commented 10 months ago

Hi, smplx_joint_img is 2D keypoints, smplx_joint_cam is 3D keypoints.

longphamkhac commented 10 months ago

Another question is that I see in the human_models.py, you changed the joints number of SMPL-X model to 137. However, in the original paper it has only 55 joints include the root. Can you explain why?

linjing7 commented 10 months ago

Hi, the 137 joints include face keypoints while the 55 joints do not. We add face keypoints for supervision to obtain a better face result.

longphamkhac commented 10 months ago

How do you get the face keypoints in MPII dataset when using SMPL-X model ?

linjing7 commented 10 months ago

The pseudo GT labels are from NeuralAnnot. You can have a detailed check on that project.