search

EricGuo5513 / HumanML3D

HumanML3D: A large and diverse 3d human motion-language dataset.
MIT License
805 stars 78 forks source link

6D rotation in new_joint_vecs #117

Open yw0208 opened 9 months ago

yw0208 commented 9 months ago

Hi, thanks for your great work. I am trying to recover the SMPL mesh from new_joints_vecs, but I find the result is weird. I covert the 6D rotation to axis angle and get mesh as shown in the figure. Do the 6D rotations in vectors fit SMPL pose parameters? If not，how should I get the SMPL pose parameters? image

MingCongSu commented 9 months ago

Hello @yw0208, I am also curious about this work. May I ask how do you get 6D rotation from new_joints_vecs, i.e. what's the specific dimension in (seq_len, 263)? And have you figured out how to recover the SMPL mesh correctly?

yw0208 commented 9 months ago

Hello @yw0208, I am also curious about this work. May I ask how do you get 6D rotation from new_joints_vecs, i.e. what's the specific dimension in (seq_len, 263)? And have you figured out how to recover the SMPL mesh correctly?

As for the specific dimension，you can refer to https://github.com/EricGuo5513/HumanML3D/issues/83#issue-1858841579 or the original paper. I didn't figure it out and have given up.

Diouo commented 8 months ago

Hello @yw0208, I am also curious about this work. May I ask how do you get 6D rotation from new_joints_vecs, i.e. what's the specific dimension in (seq_len, 263)? And have you figured out how to recover the SMPL mesh correctly?

Hi @MingCongSu , I use the 67:193 in 263 dimension and met the same problem. Have you found some methods to solve it?

MingCongSu commented 8 months ago

Hello @yw0208, I am also curious about this work. May I ask how do you get 6D rotation from new_joints_vecs, i.e. what's the specific dimension in (seq_len, 263)? And have you figured out how to recover the SMPL mesh correctly?

Hi @MingCongSu , I use the 67:193 in 263 dimension and met the same problem. Have you found some methods to solve it?

Hi @Diouo. I also gave up on recovering SMPL Mesh from 6D rotation in new_joint_vec. Instead, I followed MotionGPT's visualization process to render SMPL Mesh, which uses new_joints to fit SMPL.

If you are interested in it, you can take a look at MotionGPT's instruction about visualizing HumnaML3D data here

Diouo commented 8 months ago

Hello @yw0208, I am also curious about this work. May I ask how do you get 6D rotation from new_joints_vecs, i.e. what's the specific dimension in (seq_len, 263)? And have you figured out how to recover the SMPL mesh correctly?

Hi @MingCongSu , I use the 67:193 in 263 dimension and met the same problem. Have you found some methods to solve it?

Hi @Diouo. I also gave up on recovering SMPL Mesh from 6D rotation in new_joint_vec. Instead, I followed MotionGPT's visualization process to render SMPL Mesh, which uses new_joints to fit SMPL.

If you are interested in it, you can take a look at MotionGPT's instruction about visualizing HumnaML3D data here

Hi @MingCongSu, I have focused on this issue for several days and found this may help you too. https://github.com/GuyTevet/motion-diffusion-model/issues/114