Camera intrinsics unchanged after calibration: what to check

A calibration tool reports a better result, but the intrinsic matrix returned by your application remains identical. That mismatch can look like a failed write, a stale SDK connection or an incomplete correction. It can also be expected behavior when the tool adjusts internal reconstruction geometry rather than the published image projection model.

For an existing 3D inspection station, the useful question is not simply whether a matrix changed. You need to establish what the correction updates, which output it affects, and whether independently measured depth errors improved.

Decide whether an unchanged matrix is actually a fault

Start with the manufacturer's documented calibration scope for your exact camera, software release and capture mode. A tool's name alone does not establish which mathematical parameters it changes.

A reconstruction pipeline can contain several calibration layers:

• Lens projection and distortion parameters for an imaging channel.
• Relative positions and orientations of internal cameras or a projector.
• Rectification maps that generate an output image on a defined pixel grid.
• Reconstruction corrections used when converting captured patterns into depth.
• Transforms between depth, texture and merged output frames.

An API may expose only some of those layers. Internal corrections can therefore improve reconstructed depth while preserving the same published intrinsic matrix. A rectified output can also retain a fixed projection model while the mapping from raw measurements changes internally.

These are possible architectures, not a description of every camera. If the documentation explicitly says the exposed matrix should change, an unchanged result requires further investigation. Otherwise, treat matrix equality as an observation, not a failure criterion.

For the distinction between projection models attached to different streams, see camera intrinsic parameter conversion for depth and texture images.

Separate projection parameters from reconstruction geometry

The usual pinhole model describes how a point already expressed in a camera coordinate frame projects onto an image. It does not describe every operation used to reconstruct that point from sensor measurements.

Here, fx and fy are focal lengths expressed in pixels. The principal-point coordinates are cx and cy. This simplified expression assumes the applicable distortion correction has already been handled and that the published model uses zero skew.

Do not copy an internal module correction into a camera-to-machine transform. They describe different relationships and may use different coordinate conventions.

Check scope and capture mode before changing anything

Work through these checks in order. They separate a misunderstood correction from a genuine persistence or integration fault without immediately disturbing a working station.

1. Identify the connected device. Record its identifier, firmware, SDK release and calibration-tool release. Confirm that the tool and application connect to the same physical camera.
2. Record the active capture mode. Include channel selection, merged-output selection, resolution, cropping, binning and depth-to-texture alignment where applicable.
3. Read the documented correction scope. Find which internal parameters change, where they are stored and whether they are exposed through the API.
4. Check save and activation requirements. Follow the documented apply, save, reconnect or restart sequence. A computer restart is not evidence that a correction was committed to the camera.
5. Query parameters through a fresh session. Use the documented API for the current output, rather than an application object created before the correction.
6. Capture new data. Confirm that the test uses newly acquired frames, not a recording, cached image or previously exported point cloud.

Print matrices at adequate numerical precision when comparing them. Rounded displays can conceal small changes, although additional decimal places do not establish that a change was required.

If binning or image size changed during the investigation, camera intrinsics after binning addresses a different source of apparent inconsistency.

Validate depth against an independent reference

A tool's improved calibration score shows that its own evaluation changed. It does not, by itself, establish the accuracy of your production measurement.

Choose a reference that challenges the actual inspection task. A known step height checks local depth differences. A reference with known lateral spacing checks scale across the image. A flat surface checks shape distortion, but fitting a plane to that same surface cannot establish its absolute distance from the camera.

For absolute range checks, the reference distance must be known relative to the documented output origin. Measuring from the front housing is not equivalent unless the manufacturer defines that relationship.

Keep the following conditions matched between the before and after datasets:

• Camera mounting, reference placement and measurement regions.
• Capture mode, exposure strategy and reconstruction settings.
• Warm-up state and relevant ambient lighting.
• Filters, invalid-point handling and measurement algorithm.
• Sample count and distribution across the working volume.

A practical initial investigation could use 10 captures at each of 3 positions across the field. That is a diagnostic starting point, not a universal qualification requirement. Add near and far positions if the station uses a substantial depth range.

Use 3D camera working distance and positioning guidance when deciding which positions represent the intended measurement volume.

Work through an example with measurable acceptance criteria

Suppose an independent reference has a step height of 20.00 mm. Your application measures the difference between two surface regions, using the same processing before and after correction.

For repeated captures, calculate signed error and mean error:

Assume 10 captures give these results:

The mean-error arithmetic is straightforward:

This indicates reduced bias for that step and position, even if the intrinsic matrix remains unchanged. The unchanged standard deviation suggests that repeatability did not improve in this example.

Now suppose the application's maximum permitted absolute error is 0.10 mm. The corrected dataset still exceeds it because 0.12 mm is greater than 0.10 mm. A better average does not make the station acceptable.

These figures are illustrative, not camera specifications. Include reference uncertainty, measurement repeatability and coverage of the working volume when setting your actual acceptance limits. A step test also cannot prove absolute range accuracy or texture alignment.

Verify the output coordinate frame separately

Depth accuracy and coordinate-frame correctness are separate checks. A geometrically improved point cloud can still produce incorrect machine coordinates if the application applies the wrong transform.

For a rigid transform, the basic relationship is:

The point must belong to the camera frame named by that transform. A transform calibrated for a texture channel cannot automatically be applied to a merged depth frame.

Check the documentation for axis directions, handedness, units, transform direction and matrix storage conventions. Also establish whether depth means axial distance along the camera's viewing axis or radial range along a viewing ray. Those quantities are not interchangeable away from the image center.

Check texture registration by projecting depth points onto visible reference features, not by inspecting depth accuracy alone. Check machine placement against independent locations outside the calibration fit.

For stations that combine several cameras, multi-camera extrinsic calibration in one shared coordinate frame explains the separate frame relationship. If downstream placement changed, use vision calibration recovery after camera or robot changes to plan revalidation rather than assuming every transform must be replaced.

Diagnose failures from the observed symptom

Once you have controlled measurements, the next action becomes more specific.

Avoid repeated recalibration while changing unrelated settings. Each additional variable makes it harder to identify which operation produced the improvement or regression.

For a support case, preserve before and after captures, parameter queries, device identifiers, settings and the tool's report. Obtain software and firmware only from the manufacturer or its authorized distributors, following their compatibility guidance.

Keep a calibration acceptance record

A reusable acceptance record is more valuable than a screenshot showing unchanged matrix values. Store the configuration and evidence together so a later restart or software update can be checked against the same baseline.

Use this checklist before returning the station to service:

• [ ] The correction scope matches the device and capture mode.
• [ ] The documented save and activation procedure is complete.
• [ ] Parameter queries identify the current output stream and image dimensions.
• [ ] Independent dimensions pass the application's stated acceptance criteria.
• [ ] Tests cover representative positions, depths and surface conditions.
• [ ] Invalid-depth coverage and texture registration remain acceptable.
• [ ] Camera-to-machine placement passes its separate verification.
• [ ] A fresh connection reproduces the accepted behavior.
• [ ] Original records and recovery instructions are retained.

The guide to saving camera calibration parameters for reliable reuse covers configuration records and reload checks. Preserve manufacturer-managed calibration through supported backup methods; an exported intrinsic matrix may not contain the full reconstruction calibration.

Ask about calibration access before selecting hardware

An unchanged matrix alone is not a reason to replace a camera. First establish whether the existing system meets the measurement requirement after correction.

When selecting a new system, use the 3D sensors category and camera finder to identify candidates, then request documentation for calibration scope, output frames and supported recovery procedures.

The MV-DLS300P-04 is listed as a laser galvanometer stereo camera.

The MV-DPS200P-02 is listed as a projected structured-light 3D camera.

Those catalog descriptions do not establish which internal correction parameters their APIs expose. Confirm that requirement explicitly rather than inferring it from the sensing method.

For help defining a suitable validation test or comparing documented calibration capabilities, contact the applications team with your capture mode, working volume, reference measurements and required tolerance.