What a plumb line can and cannot tell you about your posture
We built an app that measures posture from four photos. The most useful thing it taught us is how unreliable a single photo is.
Another disclosure. Posture is an unreleased app of mine. It takes four photos of a person, front, back and both sides, finds the joints, drops a vertical line through them and reports how far each part of the body sits from it.
This piece is about what that kind of measurement is good for, and where it stops.
The plumb line
The idea is a century old and physiotherapists still use it. Hang a weighted string, stand beside it, and look at what lines up.
Ear over shoulder, shoulder over hip, hip over knee, knee over ankle. Deviations from the line get names: forward head, rounded shoulders, an anterior pelvic tilt.
A phone camera and a pose model can automate the string. The model we use, MediaPipe’s pose landmarker, finds thirty-three landmarks on the body.
The app measures the horizontal distance between each landmark and the line, scales it to the person’s height, and sorts each finding into mild, moderate or significant.
That is where it gets honest.
What we learned from forty-five people
During testing, forty-five people ran real assessments. Twenty-eight of them produced photos clear enough to re-analyse against a stronger reference model, and the comparison was humbling.
A bug in how we handled the photo’s aspect ratio had been inflating every deviation by a little over double. Three findings were reporting the wrong direction entirely, so a head tilted left was labelled right.
On the back view, the software was borrowing the nose position from the front photo because the back of a head has no nose.
After those fixes and a better detector, the median joint error fell by a third, and by more than half at the ankles.
When the same person was photographed twice, the app gave the same severity rating on the side view sixty-four per cent of the time. Before the fixes it was forty-seven.
Read that last number again. On a good day, with the bugs gone, a third of repeat measurements land in a different bucket.
What that means for you
It means a posture score from a photo, ours or anyone’s, is a rough instrument. It can tell you that your head sits noticeably forward of your shoulders. It cannot reliably tell you whether that changed since last month, unless the change is large.
It also means the confident diagrams in posture marketing, the ones with red angles drawn on a silhouette, are drawn on top of the same uncertainty. The angle is real. The precision is not.
There is a bigger caveat that has nothing to do with software. Whether a given posture is a problem is contested among the people who study it.
Plenty of pain-free people have a forward head. Plenty of people in pain have textbook alignment. A deviation is a description, not a diagnosis.
What a photo is good for
Two things. A starting picture, so you have something to compare against in six months rather than a memory. And a nudge toward the one thing that does hold up in the research, which is moving more and holding any single position less.
We keep building it, with a wellness disclaimer on every report and a score that says what it is. If a tool tells you your posture is a seven out of ten, ask it how it knows.