mks0601 / Hand4Whole_RELEASE

Official PyTorch implementation of "Accurate 3D Hand Pose Estimation for Whole-Body 3D Human Mesh Estimation", CVPRW 2022 (Oral.)
MIT License
315 stars 31 forks source link

About hand pose #62

Closed henrycjh closed 1 year ago

henrycjh commented 1 year ago

I got the pose of hands from the output of the demo code. When I apply the pose to an avatar, its hands have weird pose. I guess it is related to the parameter "flat_hand_mean" of the SMPLX model. So I am wondering if the Hand4Whole model is trained with flat_hand_mean=False? If I would like a initial "flat hand", should I train the model from scratch?

mks0601 commented 1 year ago

Yes it is traiend with flat_hand_mean=False, and you can check here to get that https://github.com/vchoutas/smplx/blob/main/smplx/body_models.py

henrycjh commented 1 year ago

Thanks for your reply. I would check it now.

sarsigmadelta commented 1 year ago

I got the pose of hands from the output of the demo code. When I apply the pose to an avatar, its hands have weird pose. I guess it is related to the parameter "flat_hand_mean" of the SMPLX model. So I am wondering if the Hand4Whole model is trained with flat_hand_mean=False? If I would like a initial "flat hand", should I train the model from scratch?

God, I Got the Same isssue, i think we need finetuing model with flat_hand_mean=True Flag

mks0601 commented 1 year ago

you can simply add subtract the mean pose from the current estimated hand pose if you want some hand pose with flat_hand_mean=True

sarsigmadelta commented 1 year ago

mean

thank you so much for reply, you meaning finetuing model is not actually necessary ? we just need to subtract the mean pose ,then we can get the right pose ?

mks0601 commented 1 year ago

Yeap. you can see what they do internally https://github.com/vchoutas/smplx/blob/1265df7ba545e8b00f72e7c557c766e15c71632f/smplx/body_models.py#L630 https://github.com/vchoutas/smplx/blob/1265df7ba545e8b00f72e7c557c766e15c71632f/smplx/body_models.py#L737C15-L737C15

sarsigmadelta commented 1 year ago

Yeap. you can see what they do internally https://github.com/vchoutas/smplx/blob/1265df7ba545e8b00f72e7c557c766e15c71632f/smplx/body_models.py#L630 https://github.com/vchoutas/smplx/blob/1265df7ba545e8b00f72e7c557c766e15c71632f/smplx/body_models.py#L737C15-L737C15

Appreciate it!