The generated npy files are comparable to the HumanML3D dataset?

[AnnaRosse]

Hi, I find very interesting your work. I used this repo to convert my bvh files to smpl, changing the hierarchy in retargeting.py since mine is different from yours. I need to convert these files because I have to fine-tune the MotionDiffuse model that is trained using HumanML3D and so I have to convert my data in the format used by HumanML3D. Do you know if the retargeted files generated from your script can be used as the "joint" files used in HumanML3D? I ask because maybe you did something similar to me with the generated files. Thanks

[EmptyBlueBox]

Hi! I'm sorry, but I don’t have any experience with that, so I’m afraid I can’t really help. I hope you find the right person / coding agent to assist you though!

[AnnaRosse]

Thank you for the answer. Just a question, in the retargeting phase, if I use this cose ''' useful for optimization '''
# bvh.batch_forward_kinematics()
# bvh.print_dims()
# retargeted_joint_trans=bvh.joint_translation[:,idx_mapping]
# print("retargeted_joint_trans.shape: \n", retargeted_joint_trans.shape)
# np.save(retargeted_data_path+bvh_files[i].split(".")[0]+"-retargeted.npy",retargeted_joint_trans) instead of the other, I obtain the global position of each joint instead of the euler angles right? And in this case there are other elements that I have to change? Because I obtain a particular animation (for example the arms that should be in front of the mesh but instead are up and one is all the way back I don't know why). Thanks!

[EmptyBlueBox]

I’m afraid I can’t solve this with the limited information provided. However, please try using Claude Code or Codex before asking questions — they can usually handle all the questions!