search

jfzhang95 / PoseAug

[CVPR 2021] PoseAug: A Differentiable Pose Augmentation Framework for 3D Human Pose Estimation, (Oral, Best Paper Award Finalist)
MIT License
368 stars 57 forks source link

Hyper-parameters for reproducing PoseAug on H3.6M #18

Closed yangchris11 closed 3 years ago

yangchris11 commented 3 years ago

Thank you for your interesting and exciting work!

After going over the code, I wonder if you can also provide the hyper-parameters (total epoch, learning rate ...) to reproduce the result in Table 4, namely the PoseAug using different keypoints source on H3.6M?

Thank you!

Garfield-kh commented 3 years ago

Hi, thank you for the interest!

There is some modification in the released code in the BA operation (fixed backbone) and the feedback loss (allowed the hardness within a proper range), which helps achieve better performance for the highest one in 3DHP (vpose-gt: 71) and H36M (stgcn-gt:36).

To reproduce the result in Table 4, you may try with the original Hyper-parameters setting and code here.

The folder structure is the same, just put the data folder from the released one to the original one and run the script in script_final.sh.

yangchris11 commented 3 years ago

Thank you so much!