search

geopavlakos / hamer

HaMeR: Reconstructing Hands in 3D with Transformers
https://geopavlakos.github.io/hamer/
MIT License
326 stars 28 forks source link

About the relative coordinates of 3D hand key points #36

Closed Derek0907 closed 3 months ago

Derek0907 commented 4 months ago

Hi, It's a wonderful job! I've got a question, when dealing with my study, I only focus on the pose itself (Don't focus on the location in the picture/ frame view), which is reflected in metric PA-MPJPE. I'm confused with the reference of the predicted 3D hand key points directly running after the demo.py, because I didn't find a point where the coordinate is zero. Do I need to transfer these coordinates to the same reference( eg. set the root point coordinate to zero) when calculating the loss between the ground truth (predicted 3D hand points coordinates) and the predicted hand coordinates?

Thanks in advance!

fnuabhimanyu8713 commented 4 months ago

This may help you https://github.com/geopavlakos/hamer/issues/30#issuecomment-1961760722

geopavlakos commented 4 months ago

If you evaluate with PA-MPJPE, then the Procrustes Alignment will automatically translate, rotate and scale the prediction to optimally match your GT, so you don't need to do any updates to the predicted 3D hand coordinates.

Derek0907 commented 4 months ago

This may help you #30 (comment)

Thanks a lot!

Derek0907 commented 4 months ago

If you evaluate with PA-MPJPE, then the Procrustes Alignment will automatically translate, rotate and scale the prediction to optimally match your GT, so you don't need to do any updates to the predicted 3D hand coordinates.

Thank you!