How to achieve the score of MPII test and val which were displayed in your paper?

zhengli97 / OKDHP

[ICCV 2021] Official PyTorch Code for "Online Knowledge Distillation for Efficient Pose Estimation"

https://zhengli97.github.io/OKDHP/

43 stars 4 forks source link

How to achieve the score of MPII test and val which were displayed in your paper? #2

Closed tt1234562 closed 2 years ago

tt1234562 commented 2 years ago

Dear researchers, I would like to follow your work, but I can not get the score as described in the table1 and table2 of your paper. Could you teach me the details of training？Thanks :)

zhengli97 commented 2 years ago

What is your score? Here are the settings I used to achieve 90.034 on the MPII validation set.

Head | Sho. | Elb. | Wri. | Hip | Knee | Ank. | Mean 97.003 | 96.111 | 90.864 | 85.953 | 89.527 | 85.412 | 81.696 | 90.034

For a 4-Stack OKDHP with three branches, we set alpha=1 and beta=2 in Eqn.8. Each branch has two shared and two individual HG modules. LR schedule: [0, 90, 120, 150], LR: 2.5e-4. Different experimental environments may lead to different results. Try to run more times.