Definition of "Real Motion"

[bring-nirachornkul]

Dear team,

Thank you for your excellent work on "Generating Diverse and Natural 3D Human Motions from Text." I have a question regarding the term "real motion", as mentioned in Section 5.1.2 Quantitative Evaluation:

"The high R precision of real motions evidences the reliability of the proposed R-precision metric, which sets an upper performance limit for all methods."

You describe real motion as setting an upper performance limit for the R-precision metric but do not explicitly define it in the paper. Could you clarify:

1. Does "real motion" refer to the ground-truth 3D motion data from the HumanML3D and KIT-ML datasets?
2. How exactly is this real motion data used in the evaluation, particularly in the R-precision setup?
3. If real motion represents the ground truth and is expected to serve as the benchmark for comparison, how does it achieve a score of 0.511 (Top-1) in the HumanML3D evaluation table? Shouldn't the ground truth ideally achieve a perfect R-precision score since it directly corresponds to the text descriptions?

Thank you for your time and for addressing these questions. I look forward to your clarification!