@@realcrane I wonder how well it'd be able to adapt to that constraint. If the target pose has no velocity, but is far away with little time to reach it, the character would have to speed up more to give itself time to deccelerate. It'd be very interesting to see!
@@realcrane By dynamics, are you referring to physical simulation? I didn't mean that 😅 I just mean, basically your work synthesizes "artificial mocap data" to interpolate frames (if I understand it correctly) instead of a muscle control policy for physically simulated characters. That being said, both have a form of acceleration - the 2nd derivative of position. I just mean in the scenario I described earlier, it'd have to synthesize in-between frames with extreme agility - drawing from what it learned from dataset clips of athletes rapidly speeding up and slowing down.
Dear Mr. Wang, I am a graduate student in China, and I am very interested in the research direction of your laboratory. I hope to follow you to conduct research during my Ph.D., but my research direction during my postgraduate period is not computer vision. Will this affect PhD applications?
good question. Normally it is a tool to generate motions for the animators to further refine. So synthesizing high-quality, natural and controllable initial motions is important and is the point of the paper.
I think if you encoded the position AND the velocity in the target frame's pose, it'd help with the velocity discontinuities after the transition.
Interesting point. Something we can try later.
@@realcrane I wonder how well it'd be able to adapt to that constraint. If the target pose has no velocity, but is far away with little time to reach it, the character would have to speed up more to give itself time to deccelerate. It'd be very interesting to see!
@@WhiteDragon103 dynamics is harder to control. It's easier for optimization but harder for learning.
@@realcrane By dynamics, are you referring to physical simulation? I didn't mean that 😅
I just mean, basically your work synthesizes "artificial mocap data" to interpolate frames (if I understand it correctly) instead of a muscle control policy for physically simulated characters. That being said, both have a form of acceleration - the 2nd derivative of position.
I just mean in the scenario I described earlier, it'd have to synthesize in-between frames with extreme agility - drawing from what it learned from dataset clips of athletes rapidly speeding up and slowing down.
@@WhiteDragon103 no I meant motion dynamics, which doesn't necessarily mean dynamics in physics.
Cool Tech!!
Thanks!
Nice work🤘
Amazing work! Is there any released code and/or datasets for this?
Thanks for the support! The dataset is a public dataset. The code is currently proprietary but we are looking into other possibilities.
Dear Mr. Wang, I am a graduate student in China, and I am very interested in the research direction of your laboratory. I hope to follow you to conduct research during my Ph.D., but my research direction during my postgraduate period is not computer vision. Will this affect PhD applications?
have a look at my webpage: drhewang.com/. We do more than computer vision.
@@easy-sj8ns Probably you can send an email? TH-cam might not be the best place to discuss PhD applications....
@@realcrane thanks for your advice
what is it even supposed to be
good question. Normally it is a tool to generate motions for the animators to further refine. So synthesizing high-quality, natural and controllable initial motions is important and is the point of the paper.