Why this app talks about confidence so much
A phone video has real limits. Honest uncertainty beats a confident wrong answer.
Published 20 June 2026 · Updated 8 October 2026 · OpenBikeFit
Camera field guide
Three questions. No guarantee.
Visible point
Is the landmark tracked?
Usable clip
Are its cycles usable?
Independent agreement
Do separate takes agree?
Visibility concerns an estimated point; quality concerns a clip.
Independent agreement concerns the recorded condition.
None alone certifies physical accuracy, comfort or installation.
On this page
Understand before recording
Inspect quality in a fictional analysis
Use the existing free workflow and read its source and scope checks; opening a page does not accept consent or verify a rider.
Open toolConfidence is useful only when you know what it refers to. A pose model may be confident that a point is visible, while the bicycle is filmed obliquely. An angle series may be stable, while the same estimated joint centre is misplaced in every frame. Two takes may agree, while both share that camera error. None of these observations demonstrates that a riding position is ideal.
OpenBikeFit therefore keeps quality bands and reasons near the estimates and checks independent recordings before allowing its controlled advice workflow to continue. This guide explains those software signals without converting them into a guarantee. It changes no engine thresholds, reference bands or safety gates. A clean technical result still needs its recording context and limitations; a low-quality result should not be disguised by extra decimal places.
Signal one: landmark visibility
The pose engine provides estimated body points and visibility-related scores. Google's landmark definition describes visibility as a score concerning whether a landmark is visible or occluded, including being outside the scene. That is not a measurement of how far the estimated point lies from the true anatomical centre. It should not be read as an angle-accuracy percentage or a probability that a component needs moving.
In the app's side-view processing, the constituent landmarks contribute to per-frame quality. An obscured ankle can weaken a knee-angle observation even when the hip is clear. Tracking the near-side joints consistently is a practical preparation goal; increasing a display score does not itself prove that a camera was square, a bicycle was level or a physical reference was correctly located.
Signal two: measurement quality within a clip
The current measurement-quality calculation combines landmark coverage, the number of usable cycles and cycle-to-cycle variability. It presents coarse high, moderate, low or insufficient bands with reasons. These are app-specific technical screening categories. They are not a validated clinical scale, a confidence interval or a statement that an estimate is within a particular number of degrees of the true motion.
Read the reasons as part of the result. Inconsistent tracking, too few usable cycles and variable motion answer different questions and need different preparation checks. A clearer view may help tracking; a separately repeated settled recording may reveal unstable motion. Do not erase a reason by quoting only the most favorable joint or the overall band. A result is a collection of named metrics and conditions, not one universal score.
Signal three: agreement between independent takes
A second recording asks whether the first observation can be repeated under the same declared setup and protocol. It is not another sample obtained by replaying the first file. The recording-condition workflow checks each accepted take and multiple decision metrics. Agreement includes source eligibility and the app's observed-noise comparison; those checks should remain visible rather than being compressed into the word passed.
When an additional accepted take is inconsistent, the workflow does not hide it by choosing a matching pair. That restraint protects the interpretation of the software evidence. It does not prove the physical setup stayed identical, nor does an agreeing pair eliminate a shared systematic error. If the condition changes, keep that distinction instead of pretending a new recording belongs to the same unchanged baseline.
Repeatability and accuracy answer different questions
Repeatability asks whether comparable observations agree. Accuracy asks how close the measurement is to the relevant physical quantity or reference method. A camera can produce stable projected angles that differ from a calibrated measurement. A numerical example with a small observed spread describes its own readings, not every error that could exist in the measurement system.
Professional camera systems can use physical calibration to locate and orient the bicycle relative to their cameras. That is an additional measurement procedure, not something automatically supplied by a phone model emitting three-dimensional-looking coordinates. OpenBikeFit's sagittal joint angles use image-plane geometry. Do not assume that model output depth, sharper pixels or a high band reproduces another system's calibration or validation evidence.
Keep confidence attached to its quantity
A report can contain a knee-angle quality band, an overall analysis band and an independently confirmed condition. Keep their labels separate. Quality for one joint cannot make a different missing joint usable. A camera-derived degree value is not a physical saddle height in millimetres. Comparing a mechanical coordinate needs its own endpoint, datum, instrument and repeated readings, not a conversion based on a confidence label.
Likewise, reference bands describe the app's stated mechanical comparison context, not the certainty that a rider will be comfortable. A displayed change, an eligible before-and-after comparison and rider-reported experience are separate pieces of evidence. Do not call any of them a diagnosis or installation approval. The article supplies a way to read the evidence, not a reason to bypass a blocked workflow.
What to do with an uncertain result
First read the specific reason, then return to the recording conditions it concerns. Check whole-cycle framing, the visible side, lens, lighting, camera stability and the settled riding condition. Begin a genuinely independent take after correcting preparation rather than repeatedly replaying the same clip. If uncertainty remains, keep it unresolved; no advice is better than inventing a precise correction from poor evidence.
You can inspect the public fictional analysis without camera permission or a saved fit, and use the readiness tool when the device itself is uncertain. If the intended decision needs physical installation checks or knowledge outside the side-view protocol, use the exact manuals and appropriate human assessment instead of a quality badge. Reading these static guides does not grant permissions, accept the fit scope, create records or send a video. Explicit tool actions still have their own requirements.
Practical questions
Frequently asked questions
Does a high quality band mean the angle is accurate?
Not by itself. The app band screens technical quality using tracking, usable cycles and variability. It is not a calibrated degree-error bound, a fit probability or clinical validation.
Can several identical results still be wrong?
Yes. An unchanged projection or consistently misplaced landmark can repeat across every take. Agreement checks repeatability, not the absence of a shared systematic error.
Why are quality and repeatability separate?
Quality concerns whether a clip provides usable technical evidence. Independent agreement concerns whether eligible separate takes match under the declared protocol. Neither alone establishes an ideal physical setup.