Hi, I have two questions about the rotation data in the genearted dataframe:
You only utilize the relative xyz (ric data) in the generation and motion recovery. Is the rotation just redundant data for network training assistance? Have you tried using rotation data for control guidance?
Did you evaluate the quality of the rot data in the generated results? The FID measures it, but not directly. Have you tried recovering motion from the rotation data and assessing control error and visual performance?
Yes the rotation part is just for network training. In the control guidance, we only use relative xyz. We didn't try to use rotation data. Probably it works as well.
No. When converting to human joints from the redundant representation (263), we always use relative xyz.
Hi, I have two questions about the rotation data in the genearted dataframe:
You only utilize the relative xyz (ric data) in the generation and motion recovery. Is the rotation just redundant data for network training assistance? Have you tried using rotation data for control guidance?
Did you evaluate the quality of the rot data in the generated results? The FID measures it, but not directly. Have you tried recovering motion from the rotation data and assessing control error and visual performance?