Computer Vision
Pose Estimation
Broadcast-ready multi-player pose inference at 25 fps — 17 keypoints per player, stabilized for real match footage and built to feed every downstream model.
What it is
Everything WithoutBall does begins with the skeleton. Pose estimation turns ordinary video — a broadcast feed, a tactical camera, a phone on a tripod — into structured skeletal motion: joint positions per player, per frame, ready for quantitative analysis. No wearables, no markers, no hardware on the athlete.
The output is not a highlight overlay; it is a measurement layer. Every capability on this site — signatures, search, clustering, risk signals — consumes these skeleton sequences as its raw material.
How it works
The planned production pipeline is deliberately unspectacular: detect players, track them across frames, estimate 2D keypoints — then let the foundation model do the hard part. Denoising, filling occlusions, and lifting 2D detections into consistent 3D skeletons happen in a single forward pass, because the model treats a 2D skeleton as a lossy projection of a 3D one. Stabilization is reconstruction, not post-processing.
The product truth is a 21-joint, 20-bone topology; standard 17-keypoint detector outputs map onto it. That mapping is documented joint by joint — not hand-waved.
What you get
One input, two products: clean 3D skeleton sequences for visual and biomechanical review, and the structured substrate every downstream model consumes. The table below is the honest datasheet — targets are labeled as targets.
| Parameter | Value | Notes |
|---|---|---|
| Input | single-camera video | Broadcast, tactical, or phone footage — no on-body sensors. |
| Output | 3D skeletons · T × 21 × 3 | Denoised and occlusion-filled, per player and per window. |
| Topology | 21-joint ↔ COCO-17 | Our 21-joint production topology — the joint table and mapping live in the skeleton docs. |
| Rate target | 25 fps | An engineering target for near-real-time operation — not a shipped benchmark. |
Status & roadmap
Prototype. The pose-to-representation stack is exercised on professional match footage with optical-tracking ground truth; packaged video ingestion for customer footage is roadmap. Metrics, protocols, and caveats live in the evaluation docs.