What gets measured, exactly

The measurement

Two open models run on your device: MediaPipe's face landmarker (478 facial points) and its hair segmenter (a per-pixel hair-probability map). From the landmarks we anchor vertical scan columns to your brows — one at the centre, one at each brow peak, and a 5-column sweep per side between the brow peak and its outer end that keeps the deepest bare notch it finds, which is where the frontotemporal angle actually lives. Each column walks upward from just above the brow until it meets a solid run of hair; that meeting point's height, in brow-to-chin units, is your hairline profile.

The staging signal is recession: average corner-notch height minus centre height. Below 0.08 face-units reads as stage 1, up to 0.18 as stage 2, up to 0.35 as stage 3, beyond that as deep frontal loss. A second signal floors the verdict: a centre hairline above 0.52 face-units reads as at least stage 2, above 0.66 as at least stage 3 — straight-back recession raises the whole line without deepening the M, and corner depth alone misses it. Within 0.03 face-units of any cut-off the verdict is presented as borderline, with the neighbouring stage named.

These thresholds were calibrated in August 2026 against real front-facing portraits of men spanning the full scale, staged by eye against the Norwood criteria and run through the exact pipeline that runs in your browser. That pass is also where the corner sweep comes from: a single column above the brow's outer end reads temple-point hair below the notch and calls receded men full-haired. The corpus keeps growing; when a threshold changes, this page changes with it.

A front-facing photo contains no crown. Rather than pretend otherwise, the tool stages the front from measurement and asks you two questions about the rest: the state of your crown, and — when it separates stages — whether a band of hair still divides front from crown. Those answers route the verdict through the same table a clinician's intake questions would.

• No face found — nothing is staged from a photo without one.
• Head turned or tilted too far — beyond ±14° of turn or ±15° of tilt the temple columns drift off the frontotemporal angles, so the tool asks for a straight-on shot instead of mismeasuring.
• No readable frontal hair — very light or shaved hair and an already-bare top look identical to a segmenter (under 10% hair coverage in the zone where frontal hair sits). A second, multiclass segmenter then checks what actually fills that zone: bare skin gets a precise notice, but either way the tool switches to questions rather than guess — reading a hat as a Norwood 6 is this tool's worst possible failure, so it is the one most explicitly designed out.
• A hat, or hair over the forehead — a "hairline" measured below 0.34 face units sits lower than any real anatomy (calibration put cap brims at 0.24–0.31 and the lowest real hairline at 0.39). The multiclass segmenter decides which it saw, and the tool asks for a retake — take the hat off, or push the fringe back — instead of measuring headwear.

The Norwood scale itself is a shorthand — real heads sit between stages and off the chart's idealised patterns, which is why borderline verdicts name both stages. Segmentation quality drops on very light hair against light skin and in harsh backlight; the refusal rules catch most of this, not all. And a stage is a position, not a trajectory — the tool cannot tell a stable hairline from a moving one, which is why the maturing-or-receding guide leans on year-apart photos. None of this is medical diagnosis, and sudden or patchy loss belongs with a doctor, not a scale.