Why a rep counter counts push-ups you didn't do

Prop your phone against something, stand in front of it, and do nothing at all. On a lot of apps, the rep counter starts climbing.

This is not a broken camera or a bad phone. It is what pose detection does by design when it cannot see a body properly — it guesses, and it tells you it is guessing, and an app that does not listen to that reads the guesses as movement. The interesting part is where the guessing comes from, because it explains why standing in front of the phone fools it more reliably than an empty room does.

What the phone is actually doing

An on-device pose detector — Google's ML Kit is the common one on Android — takes each camera frame and returns a skeleton: 33 points for shoulders, elbows, wrists, hips, knees and so on. A push-up counter then does something simple with it. Take the shoulder, the elbow and the wrist, measure the angle at the elbow, and watch it: arms straight is the top of the rep, arms bent is the bottom, and one full trip from top to bottom and back is one push-up.

That works. When the detector can see you, the angles are good and the count is right.

The catch is in what comes back when it cannot see you. It does not return nothing. Every one of those 33 points arrives with a confidence score, which Google's documentation describes as:

For each landmark, a measure that indicates the probability that the landmark is within the image frame.

A number between 0 and 1. And when the detector is unsure, it still hands you a complete skeleton — just one whose points carry middling scores and sit wherever the model's best guess put them.

Those guesses are not stable. Frame to frame they wander by a few pixels, because each frame is a fresh guess at an ambiguous picture. An elbow angle computed from wandering points wanders too. And a counter watching for that angle to cross a threshold will, sooner or later, watch it cross.

Why standing there is worse than leaving the room

This is the part that surprised me, and it is in the documentation too:

The user's face must be present in order to detect a pose.

So a camera pointed at an empty wall usually returns nothing at all, and nothing is easy to handle. But stand in front of it and you have given the detector the one thing it requires. It now has to produce a body — and you are standing, which is not the position it is being asked about.

What comes back is a skeleton with a face it is sure about and limbs it is not. Middling confidence, points bunched near the middle of the frame, drifting each frame. Exactly the input that produces phantom reps.

Which is why "I just stood there and it counted ten" is a more common report than "I pointed it at a wall and it counted ten". The second one often does nothing. The first one is the failure case.

The three checks that actually stop it

I spent yesterday fixing this in our own app, so the specifics below are what the fix looked like rather than a general recommendation. Each of the three is defeatable alone.

Check	What it catches	What gets through without it
Confidence floor	Guessed landmarks	A skeleton the model is unsure about, treated as fact
Posture gate	Wrong body position	Somebody standing and bending an elbow
Travel requirement	Small jitter	An angle that crossed a line without going anywhere

A confidence floor. Refuse any frame where the joints being measured score below a threshold. Ours was 0.50 and that was the root of it — 0.50 is squarely inside the band those guesses occupy. It is 0.70 now for the joints a push-up is measured from.

A posture gate. Check that the body is in the shape the exercise needs before counting anything. A push-up is a plank: the torso should read as roughly horizontal in frame, not vertical like somebody standing.

This one has a trap in it, and we fell in. Measuring "how horizontal is the torso" works when the phone is beside you and fails completely when it is in front of you looking down your body — head-on, the shoulders and hips land almost on top of each other and the torso barely projects at all. So we added an exemption for that case, keyed on the torso being short relative to the shoulders.

The exemption then became the hole. A guessed skeleton has all its points bunched together, so its torso is the shortest of all — meaning the poses that most needed the posture check were precisely the ones that skipped it. Fixing that took a floor as well as a ceiling: a genuine head-on push-up still has some torso, so anything that has collapsed to nearly nothing is a bad reading rather than foreshortening.

A travel requirement. Even with the first two, an angle can technically cross a threshold without the arm really moving, because the normalised signal a counter watches compresses the real range. So the last check works on the raw measurement: the elbow has to have actually swung through 40 degrees for a rep to count. A real push-up is far more than that. Jitter is far less.

The two-minute test for any app that counts you

You do not need to read anyone's code to find out whether their counter checks its inputs:

1. Prop the phone where you would normally put it.
2. Start the exercise mission or workout.
3. Stand in the frame, in view of the camera, and do absolutely nothing.
4. Wait one minute.

A counter that climbs is not checking confidence. A counter that sits at zero and tells you it cannot see you properly is.

Try the empty-room version too, for contrast — most apps do nothing there, which is why the failure survives testing. Somebody checking their own app points it at a wall, sees zero, and concludes it is fine.

Where this leaves you

If you use any app that counts reps from a camera, the honest summary is:

• It is probably accurate when it can see you. Good light, whole body in frame, camera to your side — the angles are reliable and the count is real.
• It may be nonsense at the edges, and the edges are where an alarm lives: a dark bedroom, a phone on the floor, a body half out of frame.
• The failure is silent, and it fails in the direction that flatters the app. Nobody reports an alarm that was too easy to turn off.

And if the reason you chose an exercise mission was that you do not trust yourself at 6am, that is exactly the reason to spend two minutes checking whether the counter can be fooled by standing still. It is the same instinct, applied one level down — the same one that makes five alarms worse than one snooze, and the reason an alarm you can dismiss in your sleep is not really an alarm. If you are picking between apps on exactly this, the missions each one offers — and which of them can be finished lying down — are set side by side here.

Sources

• Pose detection — Google's own ML Kit documentation. The definition of landmark likelihood and the requirement that a face be present are both quoted from this page.
• Pose detection concepts — the 33-landmark model and what is returned for a partially visible body.

The thresholds named above (0.70 confidence, 40 degrees of travel) are ours and are not a standard. They are the numbers that worked against real push-ups on real phones; anybody solving this in their own app should expect to find their own.