Reconstruction
Skeleton Repair & 2D→3D Lifting
Real tracking data is imperfect: joints drop out, occlusions cut sequences, 2D detections carry noise. The foundation model repairs all of it in one forward pass — denoised skeletons with the gaps filled, lifted from flat 2D video coordinates into consistent 3D motion.
Where this shows upPersonal PerformanceTeam IntelligenceLive preview
What it is
Real tracking data is damaged data: a leg occluded behind another player, a dropped frame at the worst moment, jitter on every fast turn. Skeleton repair treats that damage as a first-class input — give the platform imperfect skeleton sequences, get back complete, anatomically consistent ones. Where pose estimation reads bodies out of pixels, repair works on skeleton data itself, including sequences produced by other tracking systems.
The before and after is concrete. Before: flat 2D detections with holes — missing joints, missing frames, noisy coordinates. After: the same movement, complete — every joint present, bone lengths consistent, velocities smooth, lifted into 3D joint trajectories ready for biomechanical measurement.
How it works
Under the hood there is exactly one model. The foundation model treats an imperfect sequence — joints missing, frames dropped, coordinates noisy — as partial evidence of one complete movement, and reconstructs the version consistent with anatomy and with time: bone lengths stay fixed, velocities stay smooth. A 2D sequence is handled the same way, as a flat view of a 3D movement whose depth can be reconstructed. Denoising, occlusion fill, and 2D→3D lifting are therefore not three tools but one reconstruction — produced in a single forward pass, with no per-case tuning.
The output lands on the 21-joint production topology; how the model learned to complete movement is documented in the foundation-model docs.
What you get
Repair is why the rest of the platform stays dependable on real footage — signatures, search, and baselines all consume repaired sequences. And it stands on its own: tracking data you already have can be repaired and lifted without re-recording a single session.
| Parameter | Value | Notes |
|---|---|---|
| Input | skeleton sequences · 2D or 3D | From our own pipeline — or tracking data you already have. |
| Repairs | denoising · occlusion fill · gap fill | Missing joints and dropped frames restored as anatomically consistent motion. |
| Lifting | 2D → 3D | Flat video detections become 3D joint trajectories on the 21-joint topology. |
| Output | complete sequences · T × 21 × 3 | Every joint present, bone lengths consistent, velocities smooth. |
| Boundary | reconstruction ≠ recording | A filled joint is the most plausible motion, not an observation — long gaps stay uncertain. |
Status & roadmap
Prototype. Reconstruction is the validated core of the proof of concept: denoising, occlusion fill, and 2D→3D lifting are measured against independent ground-truth references — the same forward pass, no extra machinery. Packaged repair for customer data follows the ingestion roadmap; metrics, protocols, and caveats live in the evaluation docs.