In my experience pose-estimation is relatively solved problem. The biggest issue is estimating camera movement, especially across cuts. There are few commerical solutions for AI mocap (move.ai, deepmotion, wonderdynamics, cascadeur and others), and it's still far from being solved. For accurate motion estimation, lens correction is reuqired, and lens focal distance and sensor size information is required. I don't know if any of these solutions support zoom level change.

There are tons of practical applications for the AI mocap, especiall in sports. But requirement of calibrating users (standing in T pose) and using only calibrated and fixed camera limits the use of these AI mocap solutions to the lab settings only.

They use IMU sensors not cameras. The linked article is a bit bad in that it does not make it very clear and it has a quite misleading hero image (in the tradition of useless hero images). They try to estimate pose in detail through IMU signals. While pose estimation through IMU is not new, I am not sure if something like getting into that deep detail (out of a general classification) has really been done.

Lens adjustment, sensor adjustment, this is a problem ripe for the deep learning "bitter lesson"