What 2D analysis misses when a limb leaves the camera plane

How to tell whether a 2D angle can be trusted

1. 01Check the phone is really side-on

   On a track, the lane lines should cross the picture horizontally and stay parallel to each other. If they clearly converge towards one side, the camera is filming at an angle, and every angle read on the image is distorted, slightly in the middle and more at the edges.

2. 02Read the angle while the athlete is mid-frame

   In the middle, the camera sees the athlete in profile. Towards the edges it sees them more and more from three quarters: with a phone's main lens, the line of sight is around thirty degrees off square at the sides of the picture. Motion that stays exactly in the plane does not suffer from it, but any movement in depth shows up more the further you are from the centre.

3. 03Look at the movement from the front at least once

   A side view cannot show whether the knee caves in or the foot crosses the running line. A short separate clip from the front or from behind tells you whether the limb stays in its plane. If it clearly leaves it, angles taken from the side are orders of magnitude, not measurements.

4. 04Be wary of the far leg

   In 2D, the leg furthest from the camera is often hidden behind the other one, and it is further away, so it looks smaller and is seen from a slightly different angle. Comparing left and right on a single side-on video gives a gap that partly comes from the camera, not from the athlete.

5. 05In 3D, turn the view around

   With a 3D skeleton, rotate the view to watch the run from the front or from above. You then see directly whether a segment leaves the running plane, and the angle shown no longer depends on where the phone stood.

Why an angle changes when it leaves the plane

A video is a projection: every point of the body is flattened onto the sensor. Lengths parallel to the sensor are kept, up to a scale. Lengths that run in depth, towards or away from the camera, are foreshortened, and vanish entirely when they point straight at the lens.

A joint angle is made of two segments. If both stay in a plane parallel to the sensor, the angle reads true. If that plane is turned, part of each segment runs in depth, the two are shortened unevenly, and the angle on the image is no longer the real one. Every app that measures on the image works this way, whether you draw the angle by hand as in Kinovea, Dartfish or OnForm, or a model places the points on the image as in DrivePhase, Track & Field AI, Astra or VueMotion.

Worked example: a 90 degree knee, turned 30 degrees

Take a knee bent at exactly 90 degrees, and assume the whole leg moves in a plane turned 30 degrees about the vertical from the image plane, because the pelvis has rotated or the phone is not square to the track. Ignore perspective, which is fair at around ten metres.

In that projection, everything vertical keeps its length and everything horizontal is multiplied by cos 30°, which is 0.866.

First case, front leg, knee leading: the thigh slopes forward and down at 45 degrees below horizontal, and the shank comes back at 45 degrees. Each segment has a horizontal and a vertical component of the same length, 0.707. After projection the horizontal drops to 0.707 × 0.866 = 0.612. Each segment then sits at arctan(0.707 / 0.612) = 49.1 degrees from horizontal instead of 45, and the knee angle on the image is 2 × 49.1 = 98.2 degrees.

Second case, back leg in recovery: the thigh 45 degrees behind vertical, the heel drawn up, the shank 45 degrees on the other side of vertical. This time each segment sits at arctan(0.612 / 0.707) = 40.9 degrees from vertical, and the same 90 degree knee reads 81.8 degrees.

So one real angle shows up as 98 or 82 degrees, and the error flips sign with posture. No fixed factor can correct it without knowing the rotation, which a single image does not give you. At 45 degrees of rotation you would read 109 or 70 degrees. On the other hand, a phone only 15 degrees off square gives 92 or 88 degrees: a slightly misaligned camera costs little, it is the limb's own rotation that makes the big errors.

Where it happens in a sprint

• Pelvis rotation: with every stride the pelvis turns about the vertical, and the thigh flexes in a plane that turns with it.
• A knee caving in at ground contact, or a foot crossing the running line: motion in the frontal plane, which a side view sees poorly or not at all.
• An arm crossing in front of the trunk: the elbow moves in depth, and the arm angle read from the side is foreshortened.
• A camera not quite square to the track, or an athlete filmed at the edge of the frame, seen from three quarters.
• The ultra-wide lens, which bends the edges of the picture on top of everything else.

What a 3D reconstruction from one phone does

TrackGenius estimates a 3D skeleton on every frame, then places it in space. The four ground markers, whose size you give, tie the image to the ground. Depth, which a single camera cannot see, is fixed by the lane you declare and the athlete's height from their profile: the athlete runs in that lane, in a straight line, at one height, without sinking into the track.

Angles are then computed on that skeleton, not on the image. Knee and elbow are the true angle between the two segments, whatever plane they are in. The hip and the upper arm to trunk angle are measured in the body's own sagittal plane, the one set by the pelvis or the shoulders, which follows the rotation of the trunk instead of staying locked to the camera. Every angle is given for the left and the right side separately, with a correction for the two legs being swapped.

What 3D from a phone cannot do

It is an estimate by a model, not lab motion capture with markers and several cameras. Depth is the direction the camera sees worst: the model infers it from what it learned about the shape of the human body, and the leg hidden behind the other one is estimated, not observed. The phone's focal length is not read but assumed, which can shift the scale a little.

Accuracy therefore depends on the setup: markers clicked precisely, the right lane and height, a side-on view, good light. TrackGenius has not been validated against a lab reference, and it does not measure knee valgus or any frontal plane angle.

And when the movement really stays in the camera plane, a good 2D measurement is right. 2D is not wrong by nature, it is wrong outside its plane.

Frequently asked questions

In short

• An angle read on the image is only right if both segments stay in a plane parallel to the camera. Pelvis rotation, a caving knee, an angled camera and the edges of the frame all take it out of that plane.
• A 90 degree knee in a plane turned 30 degrees reads 98 or 82 degrees depending on posture (horizontals multiplied by cos 30° = 0.866). A phone 15 degrees off square on its own only costs about 2 degrees.
• TrackGenius measures on a 3D skeleton placed in space by four ground markers, the lane and the athlete's height, in the body's own planes and for left and right separately. It is a model estimate, with no lab validation.