What 2D video analysis can and cannot tell you

30 July 2026·6 min read

Any tool that measures your body from a phone video has a ceiling, and it is better to know where it is than to find out by trusting a number too far. This is the honest version: what a side-on video read is genuinely good at, what it is mediocre at, and what it simply cannot see.

What it does well

Finding a saddle that is clearly wrong. This is the strong case. Knee flexion at the bottom of the stroke is a large, well-defined angle formed by three joints that are all clearly visible from the side, measured at a repeatable moment in the pedal stroke, and averaged across dozens of revolutions. A saddle 15 mm too low is unmissable.

Telling you whether a change helped. Even where absolute accuracy is limited, relative accuracy is better. If you film two clips with the same camera setup and only change the saddle, the difference between the two readings is more trustworthy than either reading on its own, because the systematic errors are the same in both.

Catching locked-out arms. A near-0° elbow is a big, obvious signal.

What it is mediocre at

Reach. The shoulder angle is the weakest of the four readings. Shoulder landmarks are hard to place from a single view, and the number depends on exactly where your hands were during that clip. Useful as a hint, not as a basis for buying components.

Anything near a band edge. A phone video resolves joints to a few pixels, which is a few degrees of angle. A reading of 41° against a 40° limit is not meaningfully different from 39°. The analyzer allows about 2.5° of tolerance at each edge for this reason, and a difference of one or two degrees between two clips is noise, not progress.

Very aggressive or very upright positions. The target ranges come from research on fairly conventional road positions. A deep TT tuck or a very upright city bike sits at the edge of what those numbers were derived from.

What it cannot see at all

Cleat position and rotation. Invisible from the side, and it changes the effective hip-to-pedal distance — so it interacts with the one measurement the tool is best at.
Q-factor and stance width. Not measurable from a side view.
Saddle pressure and soft-tissue comfort. The single most common reason people are unhappy on a bike, and completely invisible to a pose model. Green angles and a saddle you hate is an entirely possible combination.
Leg-length differences. A real asymmetry needs measuring in person.
Foot and ankle mechanics. Arch support, varus/valgus wedging, how your ankle actually behaves under load.
Anything about the frame. The tool measures your body, not your bike's geometry. It cannot tell you your bike is the wrong size — only that your position on it is outside a range.

Why the front and rear views are still beta

Knee tracking and pelvic rock are reported as hints, not graded green/amber/red, and that is deliberate rather than unfinished. Two findings in the literature set the ceiling.

First, 2D frontal-plane measurement has a noise floor. Reported standard error of measurement is around 2.7–3.0°, and the smallest difference that can be reliably detected is roughly 7.5–8.9°. Any threshold tighter than about 9° is inside the method's own uncertainty, no matter how good the pose model gets.

Second, correctly-fitted riders move. Healthy adults with the saddle at 85.5% of inseam show about 7.1° of pelvic coronal range of motion, and knee coronal motion around 6.6°. An early version of this tool flagged pelvic rock above 6° — below the healthy mean — which would have told the majority of perfectly well-fitted riders their saddle might be too high. The threshold now sits at 12°, above the healthy range rather than in the middle of it.

The honest caveat on top of that: those frontal thresholds come from the literature, not from clips run through this tool. They need checking against real front and rear footage before the beta label comes off.

Where a professional fit is genuinely different

A good fitter is not just holding a more expensive protractor. They can watch you ride and see compensations a still frame cannot show. They can put you on an adjustable rig and change three things in ten minutes. They can measure pressure, check your cleats, assess your flexibility directly, and ask what actually hurts and when. And they carry the thing no tool has: a few hundred previous riders' worth of pattern recognition.

What this tool is for is the step before that. It finds the obvious, free-to-fix problems, and it gives you numbers to describe your position with. If you do end up booking a fit, arriving with "my knee reads 47° at the bottom and my back hurts after an hour" is a much better start than "something feels off".

How to use it well

Trust the knee angle most, the shoulder angle least.
Keep your camera setup identical between clips so comparisons stay meaningful.
Act on 10° outside a band; ignore 2°.
Treat a grey reading as "not measured", not "fine" — re-film it.
Let how you feel on the bike outrank the screen. If the numbers are green and you are miserable, the numbers are not measuring what is wrong.

Free, runs entirely on your own device, and it tells you when it is not confident.

Analyze your clip

Sources

Reliability of 2D video assessment of frontal-plane knee valgus — SEM 2.72–3.01°, smallest detectable difference 7.54–8.93°.
Cycling kinematics in healthy adults — pelvis coronal ROM 7.1 ± 2.5°, knee coronal ROM 6.6 ± 2.7°.
Knee alignment and frontal-plane knee biomechanics in cycling — how much normal variation exists between riders.
Dynamic vs static measurement validity — the ~8° offset that makes video ranges differ from goniometer ranges.