search

jfzhang95 / PoseAug

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

I encountered some tricky problems #22

Closed psw13583112 closed 3 years ago

psw13583112 commented 3 years ago

Can this model run a single picture? When I ran a picture, there was some misalignment. The model used was MLP trained by PoseAug, and I found that the result was not very satisfactory. 20210908-104340 20210908-170632 20210908-170643

Garfield-kh commented 3 years ago

Hi, thank you for the interest! I think the problem might be 2D pose normalization. The 2D pose here has absolute value > 1 (head, feet), which is not usual in training data. Can you try scale down the 2D pose by 0.5?

psw13583112 commented 3 years ago

@Garfield-kh Thanks a lot for your prompt reply, I tried to reduce it by 0.5 times, the effect is indeed much better, but the results obtained by predicting the same picture every time are not the same.

Garfield-kh commented 3 years ago

Did you fix the random seed and also model.eval() for the dropout layer?

psw13583112 commented 3 years ago

@Garfield-kh Thank you, I will pay attention to these details