Analysis¶
A useful calibration must work for your application, not just fit the calibration images. Use the steps below to review a result and decide what to improve.
1. Check that calibration completed¶
Read the optimization outcome. Reaching a time or iteration limit may mean the result needs more work.
Confirm that the intended cameras and observations were included. If you changed exclusions after the last run, optimize again before assessing the result.
2. Inspect detections and outliers¶
Use RPE Bars to find images with unusually large errors, then inspect the original image and its detected features.
Look for:
- Blur, glare, saturation, or occlusion.
- Incorrect target dimensions, spacing, or IDs.
- Mistaken or poorly localized features.
- Movement or synchronization problems in multi-camera capture.
Exclude observations for a clear reason. Removing every difficult image may lower the error while leaving a weaker calibration.
3. Check acquisition geometry¶
Coverage¶
Features should cover the usable sensor, including edges and corners. Check both Coverage and the actual detections: the coverage map shows target footprints, not feature density.
Regions with no observations are less well supported. A low overall error does not tell you how well those regions are calibrated.
Tilt, distance, and size¶
Use varied target tilts in both directions, different distances, and different image positions. A set of almost identical, front-facing views may leave important parameters poorly determined.
The target should be large enough in the image for accurate localization without making all views identical. Very oblique views can also be difficult to detect reliably.
Keep focus and aperture appropriate to the final application. Use stable mounting, sharp images, diffuse light, and a rigid, dimensionally stable target.
To learn more about capture planning, read Calibration Best Practices.
4. Look for residual patterns¶
Read RPE Magnitudes and Directions together. Random-looking small errors are generally more reassuring than smooth, repeated patterns.
A pattern may come from the camera model, target, detection, or capture conditions. Investigate before adding model parameters. Use tooltips and statistics to compare numerical values; colours adapt to each camera.
Statistical tests can highlight behaviour worth inspecting, but a p-value is not a calibration acceptance criterion.
5. Compare model candidates using held-out performance¶
Enable Compute Test RPE when comparing models or parameter choices. It evaluates observations held out from fitting.
| Result | What to consider |
|---|---|
| Low fitted error, much higher Test RPE | Check model flexibility, pose diversity, and difficult images. |
| Similar fitted and Test RPE | Encouraging, but still compare the error with your application needs. |
| Incomplete or unavailable validation | Resolve the limitation before relying on the test result. |
Check per-camera results and validation coverage, not just the overall number. There is no universal acceptable pixel error or percentage difference.
If you repeatedly use Test RPE to select a model, keep a separate acquisition for a final check.
Uncertainty versus validation¶
These diagnostics answer different questions:
| Diagnostic | Main question |
|---|---|
| Reprojection error | How well does the calibration fit these observations? |
| Test RPE | How well does it predict held-out observations? |
| Parameter uncertainty and correlations | How well do the data constrain the estimated parameters? |
| Projection or triangulation uncertainty | What might that uncertainty mean for image or 3D measurements? |
| Independent reference measurements | How accurate is the result for the actual task? |
Large parameter uncertainty may call for better poses or a simpler model. Small uncertainty does not rule out a wrong target scale, model mismatch, or other systematic error. Fixed values are assumptions, not perfectly measured quantities.
For triangulation, evaluate realistic working distances and feature noise. Use independent reference measurements to establish task accuracy.
If you optimize target geometry¶
Inspect Target Geometry and check whether the fitted changes are plausible. Allowing target points to move can reduce image errors without improving the true camera calibration.
For dimensional work, verify the target independently rather than treating its fitted shape as a measurement certificate.
Review¶
Pro feature
Offline Review requires Calib GEOM Pro and a supported installed model, and works on air-gapped devices. Cloud Review for Web is planned but is not currently available.
Review helps you interpret calibration results and identify useful next steps. It summarizes the available diagnostics and highlights up to three priorities for review.
It is an advisory tool: check its findings against the plots, your images, and your application's requirements.
Running a review¶
- Optimize a calibration, or open a project with saved results.
- Wait for any diagnostic calculations to finish.
- Open the Review results tab.
- Choose an available model and select Analyze Calibration.
- Review the summary, evidence, and suggested actions.
Use Cancel to stop a review, or Analyze Again to request another one. Changing the model or calibration clears the displayed assessment.
If a model is disabled in the list, it is not installed.
Understanding the findings¶
| Assessment | What to do |
|---|---|
| No obvious concerns | Continue your normal quality checks; this is not an accuracy certificate. |
| Concerns found | Inspect the cited evidence and investigate the suggested priorities. |
| Insufficient evidence | Calculate missing diagnostics or collect more information before drawing a conclusion. |
Findings are marked Info, Warning, or Critical to indicate the model's suggested priority. These labels are not independently verified diagnoses. Different models, or repeated reviews, can reach different conclusions.
The review uses numerical summaries such as errors, coverage, validation, and uncertainty—not the original camera images. It cannot see blur, glare, target defects, or other capture problems directly. Missing information remains a limitation of the review.
Privacy and saving a review¶
Offline Review runs on your device; calibration diagnostics are not uploaded and it can be used on air-gapped devices. Cloud Review for Web is planned but is not currently available. When available, it will share calibration data with Calib.io's servers, where the data will be treated as confidential.
Use Copy Review to keep the assessment. Reviews are not saved in the project. Copy Prompt copies the information used for the review so you can inspect it or use it elsewhere. If you paste it into an external service, you are sharing that information with the service; follow your organization's data policy.
Keep a record¶
Record the intended application, capture conditions, model choice, exclusions, and any remaining concerns. Define acceptance limits around the measurement task, not a universal pixel threshold.
Keep the source data, export the parameters, and create a calibration report. In Pro, also save the project. Offline Review, where available, can provide a second reading of the diagnostics, but the final decision is yours.
To learn more about interpreting these diagnostics, read Understanding Reprojection Error and Understanding Parameter Uncertainty.
Glossary¶
- Pose: one target placement in the image sequence.
- Feature: an identified point on the target.
- Intrinsics: parameters describing a camera's projection and distortion.
- Extrinsics: the position and orientation of one coordinate frame relative to another.
- Residual: the difference between a predicted and detected feature position.
- Covariance: a description of estimated uncertainty and how parameters vary together.