Discrepancy in skating ratio: HumanML3D dataset vs. OmniControl results

[QingYuan-L]

Dear Authors,

While working with OmniControl, I've noticed some issues regarding the skating ratio that I'd like to bring to your attention:

1.Skating ratio in the HumanML3D dataset:

• The paper seems to assume a skating ratio of 0 for the HumanML3D dataset.
• However, when testing with the code in the repository, I'm getting a result of about 0.075.
• Does this discrepancy indicate a potential issue with the code? Or does the original dataset indeed have some degree of skating?

2.Skating ratio in OmniControl generated results:

• I've observed that the motions generated by OmniControl actually have a lower skating ratio(0.0547) than the ground truth motions.
• Is this an expected outcome? If so, could you explain the reasoning behind this?

3.Skating ratio calculation method:

• Does the current calculation method account for potential issues like clipping or floating?
• Is there a more precise method to calculate the skating ratio, especially when dealing with potential baseline shifts?
  Thank you for your time and consideration. Your insights will greatly help in better understanding and utilizing OmniControl.

[ymingxie]

Hi Qingyuan,

Thanks for asking. The Ground Truth motion is not perfect so it's reasonable to have foot skating. A better way might be also show the GT skating ratio in the table as a reference.

It's also possible that the predicted motion has a lower skating ratio or even lower FID than Ground Truth as the GT motion sequences are captured using some mocap methods that aren't flawless and may contain some errors.

Our calculation method exactly follows GMD code (https://github.com/korrawe/guided-motion-diffusion/blob/main/data_loaders/humanml/utils/metrics.py#L204). Not sure if it includes issues like clipping. As far as I know, there are no other public code scripts that can calculate the skating ratio.

Thanks