Six stages. Click any one to see what it does.
Each stage is independently deployable. We build them together, or wire the ones you need into your existing stack.
Pitch homography
Detect pitch lines and landmarks, solve the perspective transform between the camera view and the 2D pitch coordinate system. Robust to camera pans, zooms, and occluded landmarks.
What ends up in your hands.
No black boxes. Every output has an inspection format and an evaluation contract.
Annotated training datasets
COCO · YOLO · Pascal VOC formats. Stratified splits, class balance reports, edge-case packs.
Trained models
Detection, tracking, pose estimation, event classification. Quantized variants for edge deployment.
Homography & calibration
Court/pitch templates, automatic and manual calibration tools, exported transformation matrices.
Dot-map / tactical view
2D top-down projections with per-frame positions, velocities, and heatmap aggregation.
Event timelines
Passes, shots, possession changes, fouls — synchronized to broadcast timestamps.
Evaluation harness
Tracking accuracy (MOTA, IDF1), homography error, event detection P/R — reproducible on your data.
Tools & libraries we lean on.
Where this pipeline shows up.
Live tactical overlays
Real-time dot-map, passing networks, and possession overlays generated from broadcast feeds for studio commentary.
Match analysis for coaching
Post-match player workload, off-ball runs, defensive coverage maps. Per-player and per-phase breakdowns.
Talent ID at scale
Automated indexing across a season of footage — surface plays, players, and patterns that match a brief.
Interactive replay & stats
Click-and-track viewer experiences, AR-style overlays, and stat-on-demand product features for OTT platforms.
Have footage you'd like to see this run on?
Send us a clip. We'll run a stage of the pipeline on your data so you can evaluate it directly.