NRL AI Coach — play tracking & try analysis

Stage 1 — Detection

Every player, the ball, the referee

A YOLOv8m detector (nrl_v1, trained at 1280px and fine-tuned on hand-labelled NRL footage) finds players, the ball and match officials on live broadcast frames. Below is the current model running on a calibrated NRL segment.

player referee ball

Detection in motion — nrl_v1 (1280px, in-domain NRL fine-tune) on a calibrated segment. Boxes are live model output, not hand-drawn.

wide detection — **Wide shot — 29 players + referee.** Strong recall on small, distant players; the orange box is the official (a class the baseline can’t produce).

head to head — **Head-to-head.** nrl_v1 (left) vs the in-domain baseline (right) — neck-and-neck on players, nrl_v1 also catches the ball here.

Stage 2 — Calibration

From broadcast camera to a flat field

Field-line geometry is used to solve a homography that maps the broadcast image onto a true-scale field model (~0.12 m accuracy on this segment). That’s what turns pixels into metres.

line overlay — **Field model projected onto the broadcast.** The lines land on the real paint — proof the homography is correct.

birds eye view — **Bird’s-eye view.** The same frame warped to a top-down field — the canvas for tracking and try location.

Tracking & analytics

From pixels to play data

With detection and calibration in place, the engine projects every player and the ball onto a true-scale field — the foundation for player tracking, movement heatmaps and try-location analytics, all recovered from a single broadcast feed.

Performance

What’s live

Running on the current production model and a calibrated broadcast segment.

+31%

player recall vs base model (in-domain fine-tune)

~0.12 m

field calibration accuracy (Seg B)

classes: player · ball · referee

Component	Status
Player / ball / referee detector (nrl_v1, 1280px)	● Live
Field calibration & bird’s-eye view	● Live
In-domain NRL fine-tune (1280px)	● Live — +31% player recall
Player & ball positions projected to a true-scale field	● Live

Detection shown uses nrl_v1 (1280px), fine-tuned on hand-labelled in-domain NRL frames — the current production model. This is real model output, not hand-drawn boxes.