What can a phone video actually measure about your squat?

This has been measured directly. Ten physically active volunteers performed three squats to a maximum depth of 90 degrees while two iPhones recorded them from the side and a ten-camera marker-based system recorded the same reps as a reference. The phone footage was processed with MediaPipe, which is the same pose estimation model behind most free form-check tools.

Read that second row again. The 95 percent limits of agreement span nearly 69 degrees. The average miss alone, 17.49 degrees at the hip, is larger than the depth difference most people are trying to see. The authors found the overestimation was systematic rather than random, and concluded the method is usable in settings where laboratory capture is not available, provided the bias is taken into account.

A single camera cannot see through you. When one leg passes behind the other, the model is guessing. A separate study filmed walking with a single camera and compared both legs against marker-based capture, and the gap between the near leg and the far leg is large.

The hip spread roughly doubles once the joint is hidden, from about 18 degrees of range to about 37. Those figures come from walking rather than squatting, so treat them as the direction of the problem rather than a squat measurement. The practical instruction is the same either way: film the side you care about, and ignore what the far leg appears to be doing.

Knees travelling inward, a hip dropping on one side, uneven weight between feet: these happen across the body, not along it, and a camera at your side is looking straight down the axis they move in. Filming from the front instead swaps one problem for another. In the same walking study, frontal-plane limits of agreement equalled 35 to 46 percent of the total range of motion at the knee and hip, and the authors state plainly that the frontal-plane ankle should not currently be used in clinical or sporting applications.

Nothing in either study measures spine position, bar path, or how load is shared between joints. A 2D pose model returns the coordinates of points it thinks it can see. It does not know what you are holding, how heavy it is, or what your back is doing between the landmarks.

Three things, and they are worth more than they sound.

Counting. Rep detection only needs the angle to cross a threshold consistently, not to be correct. A model that reads every rep 17 degrees too deep still counts eight reps as eight.

Tempo. How long the descent takes is measured in frames, not degrees, so the angular bias does not touch it.

Your reps against your own reps. This is the one that matters. Because the error is systematic rather than random, it largely cancels when you compare two reps filmed in the same session from the same position. Rep one reaching 10 degrees deeper than rep eight is a real finding even when neither absolute number is trustworthy. Fatigue showing up as creeping shallowness is visible this way. An absolute depth target is not.

Several free tools will give you a number out of 100 and a list of faults. Work backwards through what that requires: an accurate joint angle, a threshold for correct that applies to your limb proportions, and evidence linking the deviation to an outcome. The first of those three already fails by 17 degrees at the hip. The published work supports a statement about consistency and none about correctness.

That is why there is no form grader on this site. Measurements that the method genuinely produces, such as a rep count and the spread between your deepest and shallowest rep, are defensible. A grade is a judgement dressed as a measurement.

If you want numbers from a lift that are exact rather than estimated, those exist. The one rep max calculator works from a weight and a rep count, the barbell plate calculator from what is on the bar, and the DOTS calculator scores a total against bodyweight using a published polynomial. None of them has to guess where your knee is.

If you are going to record yourself, the setup in the squat study is a reasonable template, because it is the setup the numbers above were measured with: camera 1.5 m to one side, at roughly hip height, square to the direction you are moving, with your whole body from head to feet in frame for the entire set.

Square to the movement matters more than it sounds. The study aligned the cameras perpendicular to the direction of movement specifically to reduce parallax error, and a camera angled up from the floor or in from a corner adds error on top of everything described above. Keep the phone in the same place between sessions, or you lose the one thing a phone video is reliably good for.