HumanML3D
HumanML3D copied to clipboard
Motions in HumanAct12 are not facing z+?
Hi Chuan,
I have been working with the HumanML3D dataset and recently visualized some motion sequences from the HumanAct12 subset. After preprocessing, I noticed that all the motions in this subset are facing the z- direction. Additionally, when I directly visualized the provided pose data in https://github.com/EricGuo5513/HumanML3D/blob/main/pose_data/humanact12.zip, I observed that the left and right parts of the poses appear swapped, which is consistent with other subsets' motions in './pose_data/' folder.
I noticed that other subsets' motions are processed with two steps in the raw_pose_processing.ipynb file: (1) Swapping the x-axis and y-axis (means a mirror operation). (2) Applying data[..., 0] *= -1 for another mirror operation. These steps ensure the poses are correct after mirroring twice.
However, I am unsure about the exact processing steps applied to the HumanAct12 subset. Did you also first swap the x-axis and y-axis? If so, should the HumanAct12 subset also undergo the data[..., 0] *= -1?
I would appreciate your clarification on this matter.
Best regards, Zan Wang