[ur-whale]

Just be patient, another decade and the tech will be at a point where his moves can be reverse-engineered back to 3D directly from his films.

[luma]

Pose Estimation from a 2D source is one of the default examples provided by TensorFlow: https://www.tensorflow.org/lite/examples/pose_estimation/ove...

Not only are we already there, it's now a trivial exercise used to instruct new users on the technology.

[mlyle]

The 6th grade robotic team I coach built from that example this year.

Note that this just comes up with coordinates in the 2D image for body parts. And that there's a lot of error and noise, so it's a long way from those list of points to a 3D kinematics model. But a tracking filter and linear algebra would get you a fair bit of the way there.

[ur-whale]

> Not only are we already there

We're not.

There are basic experiment that demonstrate it's feasible, but there isn't such a system in production today.

That being said, you are probably correct in that it's likely around the corner and my estimate was overly generous.

[iurysza]

We're already there.