search

wkom / VirtualPose

[ECCV 2022] Offical Pytorch implementation of "VirtualPose: Learning Generalizable 3D Human Pose Models from Virtual Data"
MIT License
66 stars 6 forks source link

About Ground Center in MuPoTS-3D evaluation #4

Open shawnpark07 opened 1 year ago

shawnpark07 commented 1 year ago

Hi, thanks for your great work

I have a question on your code (Link) which I'm a bit confused. Could you explain the meaning of "ground center" argument here, used for MuPoTS-3D evaluation?

I'm not sure whether I understood it right, for me the purpose of it seems like aligning the center of persons between prediction and GT, which should not be done for a fair evaluation of PCK_abs.

Will be grateful if you correct me right if I misunderstood it.

shawnpark07 commented 1 year ago

And also want to ask you about the way you set the space center value for root estimation

wkom commented 1 year ago

Hi, "ground center" means we set the space center of root estimation as the center of all people in the scene. For convenience, we use this center as the new origin of the world coordinate system. Then we modify the camera parameters and 3D poses accordingly.