problems in function inverse_kinematics_np

EricGuo5513 / HumanML3D

HumanML3D: A large and diverse 3d human motion-language dataset.

https://ericguo5513.github.io/text-to-motion/

MIT License

828 stars 81 forks source link

problems in function inverse_kinematics_np #131

Open azuki-miho opened 7 months ago

azuki-miho commented 7 months ago

In the 58th line of common/skeleton.py, the order of "l_hip, r_hip, sdr_r, sdr_l = face_joint_idx" is different from the comment "right hip, left hip, right shoulder, left shoulder", whether this will influence the processed data?

Meanwhile, in the 74th line of the common/skeleton.py, "root_quat = qbetween_np(forward, target)" is the rotation from forward to the [0,0,1], whether it should be the rotation from [0,0,1] to the forward?

EricGuo5513 commented 5 months ago

Hi, thanks for pointing out this. The first one seems a bug exactly. I guess it’s not major problem, but it may still cause few discontinuities for root angle. However since many experiments have been done on this version, I could not update this anymore. I will highlight this issue to note others. For the second one, as long as you keep the same rule when you extract features and recovering motions, it should be fine.