Deep Learning
Movement Signatures
Learn compact movement embeddings for off-ball behavior — compare roles, styles, and tendencies across leagues and seasons.
What it is
A movement signature is a compact numerical description of how a body moves — not where it went, but how it got there. The foundation model compresses each four-second window of skeleton motion into a 384-dimensional embedding that keeps what matters: action type, speed, coordination, asymmetry, movement quality.
Two windows land near each other in this space when the movement is similar — regardless of camera angle, pitch position, body size, or which tracking system produced the data. That invariance is what makes signatures comparable across players, sessions, and data sources.
How it works
Signatures are learned self-supervised. The model reconstructs deliberately masked regions of the skeleton across millions of frames and aligns augmented views of the same movement — so it never needs a hand-labeled event. Physics-informed constraints keep the space anatomically honest: bone lengths stay consistent, velocities stay smooth. The full training recipe, masking strategy, and data-scale story are documented.
What you get
Concept probes confirm the space is readable: directions for speed, torso lean, and knee asymmetry can be recovered from embeddings alone — results, protocols, and caveats live in the evaluation docs. In the product, signatures are the unit of comparison everything else builds on.
| Parameter | Value | Notes |
|---|---|---|
| Invariant to | camera, mirroring, body size | By training design — verified with invariance probes. |
| Sensitive to | action, speed, asymmetry, quality | The signal signatures are built to carry. |
Status & roadmap
Prototype. Signatures are trained and probed on professional match data at proof-of-concept scale — a deliberately small corpus. Comparing roles and styles across leagues and seasons is the design goal the data roadmap works toward, not a shipped claim.