search

IDEA-Research / Motion-X

[NeurIPS 2023] Official implementation of the paper "Motion-X: A Large-scale 3D Expressive Whole-body Human Motion Dataset"
https://motion-x-dataset.github.io
Other
566 stars 15 forks source link

Difference between motionx2humanml and the original HumanML3D data #73

Open weihaosky opened 7 months ago

weihaosky commented 7 months ago

Hi, I convert the MotionX data to humanml format according to motionx2humanml, but the converted data is slightly different from the original HumanML3D data, as shown in the following:

https://github.com/IDEA-Research/Motion-X/assets/23240241/0a350415-c08d-4c8a-9ece-0c0a632fd501

output

The first one is the original HumanML3D data '000000.npz'. The second one is converted from motionx2humanml. The third one is MotionX. Their framerates are different. Then I adjust the framerate to make them in the same framerate, as shown in the following:

https://github.com/IDEA-Research/Motion-X/assets/23240241/0bbb59ab-244e-44fa-ae8a-56e1e744e63b

output

However, they are still slightly different if you watch carefully. Also, the scale of the trajectory on the ground is different.

May I ask what is the cause of the difference and how to solve it? Many thanks!

XingliangJin commented 7 months ago

I found the same issue. It's more evident in the clip "000021.npy". Here, the first one is the original HumanML3D data. The second one is converted from motionx2humanml. The third one is MotionX. It can be seen that the trajectory difference between them is obvious. The motion trajectory of MotionX is elongated, and there is a serious foot sliding problem.

shunlinlu commented 7 months ago

Hi,

Thanks for your information. It may be because we use different subsets of AMASS. Original humanml uses smplh, while we use smplx. We checked that there are differences between the same case of smplh and smplx. We can not solve it now, since it's more like the AMASS problems. If you want to keep the same with the original Humanml, you may use smplh subset.

Shunlin

XingliangJin commented 7 months ago

Thank you for your checking! Since it's an AMASS problem, I will use smplh subset for the Humanml set.