A single camera cannot measure depth: a small person nearby and a large person far away make the same picture. Single-camera 3D methods fill in depth from what bodies usually look like, so they give convincing estimates that can be wrong in size, distance, speed and joint angle. A second camera measures depth by triangulation. A third reduces blind spots and lets the software check its own accuracy by rebuilding each camera from the other two.
A photo flattens the world onto a sheet. Everything along one line of sight from the lens lands on the same pixel, whether it was one metre away or fifty. That isn’t a flaw in the camera, it’s what a picture is. And it is the reason one camera can only ever estimate motion in 3D.
01What one picture can’t know
Drag the slider in Fig. 1. At ×0.8 the runner is 20% closer to Cam 2 and 20% smaller, a 1.4 m athlete. At ×1.3 he is further away and nearly 2.3 m tall. Cam 2’s picture is identical in every case, pixel for pixel, because each joint only slid along its own line of sight. Every number you might want from the clip, his height, his distance, how fast he ran, scales with that slider, and one camera has no way to tell you where it should be.
It gets worse for limbs. An arm pointing toward the camera and one pointing away can make almost the same shape in the picture. That’s the classic trap for single-camera joint angles, and it’s worst for exactly the joints coaches care about: a knee coming forward, a hip in a lunge, a throwing arm.
02How single-camera 3D works anyway
Single-camera (monocular) 3D pose models learn what bodies usually look like: typical limb lengths, how joints bend, how people stand on the ground. Given one picture, they pick the most plausible 3D pose. On everyday movement the results often look right, and for a lot of uses, like a fitness app counting reps, that’s enough.
What they give you is an estimate. Size is assumed from average proportions. Distance from the camera follows from the assumed size. Speed follows from the distance. If the athlete is taller than average, or moving toward the lens, the errors stack up, and nothing in the output tells you they happened.
03What the second camera fixes
Switch on Add Cam 1. Cam 1 sees the runner from a different direction, so its lines of sight cross Cam 2’s. Slide the runner along Cam 2’s rays now and he drifts off Cam 1’s rays at once: the miss in Cam 1 jumps to tens or hundreds of pixels. Only one position fits both pictures, and that position is a measurement. That is triangulation, and it’s the same geometry surveyors have used for centuries.
How well two cameras measure depends on the angle between them. At 90° each covers the other’s blind direction. At 10° they are nearly the same view and the crossing is shallow, so a one-pixel error slides the answer a long way. Here it is with one phone, to start from.
A top view of phones around a running lane. With one phone, depth along its line of sight is unknown; adding and spreading phones shrinks the uncertainty to a small ellipse.
04What the third camera adds
Two cameras can measure anything both of them see. The catch is “both”. The body hides itself constantly: the far arm behind the torso, the back leg behind the front one. With two cameras, any joint one of them loses goes back to being a guess. With three spread around the athlete, nearly every joint is seen by at least two at any moment.
The other gift of a third camera is harder to see and more important: it lets the result check itself. Hide one camera, rebuild the joints from the other two, and project them into the hidden view. If they land where the hidden camera saw them, the reconstruction is consistent. If they don’t, something is off: the sync, the calibration, the person the detector followed. Fig. 3 runs that check on the real take.
Each camera in turn is hidden and its joints rebuilt from the other two; the rebuilt joints land about 6 to 9 pixels from what it saw, on a 3,840-pixel-wide frame.
On this take the misses are 6 to 9 pixels on frames 3,840 pixels wide. In our pipeline that same number sets the grade on every clip. With two cameras there is nothing left over to check against.
05Side by side
| 1 camera | 2 cameras | 3 cameras | |
|---|---|---|---|
| 3D body motion | Estimated, depth assumed | Measured | Measured, fewest blind spots |
| Speed and distance | Side-on straight runs only | Any direction | Any direction |
| Joint angles | Side facing the camera | Both sides, gaps when hidden | Both sides, few gaps |
| Replay you can orbit | No | Between the two views | Around the covered arc |
| Accuracy self-check | No | No spare view | Each camera rebuilt from the others |
| Setup effort | One phone | Two stands | Three stands, ten minutes |
06Which should you use?
If you want to see technique, one camera side-on is a fine start, and most coaching apps are built for it. If you want numbers you’d act on, like a top speed, a knee angle at landing or a left–right difference, use two cameras at least 60° apart. If the numbers matter enough to be checked, or the movement turns and twists, use three.
Questions
Can you get accurate 3D motion from one video?
You can get an estimate. Depth is inferred from what bodies usually look like, so size, distance, speed and some joint angles can be wrong without any sign in the output. Two or more cameras measure depth instead.
Why do two cameras measure depth when one cannot?
Each camera sees a joint along its own line of sight. With two cameras from different directions, the two lines cross at one point, which is the joint’s position in 3D. This is triangulation.
Is a third camera worth it?
Yes if accuracy matters. It reduces the time joints are hidden from all but one camera, and it lets the software check the result by rebuilding each camera’s view from the other two and measuring the difference.
What is monocular 3D pose estimation?
Estimating a 3D body pose from a single image or video. Models learn typical body shapes and movements to fill in the depth a single camera cannot see.