Highlights
- Transfers motion capture priors into a monocular network, improving quad jump body alignment.
- Produces depth-aware joint trajectories compatible with inverse dynamics toolchains.
- Offers a lightweight inference mode for integration with temporal segmentation networks like PoseTSN.
Implementation Notes
The released checkpoints expect calibrated camera intrinsics; for broadcast feeds, approximate parameters using rink geometry. Pair with sensor data from Tanaka et al. (2023) to validate landing forces.