Checkerboard calibration not detecting corners: what to check

A checkerboard that works with one camera may fail with another because the detector receives different pixels, not because the calibration mathematics changed. This troubleshooting sequence is for inspection teams facing missing corner overlays, inconsistent detections or doubled target images. Start with the board definition, then inspect saved frames from the actual calibration stream. Only consider replacement hardware after separating target problems from acquisition defects.

Identify the failure before changing settings

Calibration estimates camera geometry from detectable image features. It does not reconstruct sharp corners from ghosted images or correct a wrongly described target.

Use the following tests in order. Change one variable at a time and keep the original frames for comparison.

A successful detection on another camera confirms that the target can work under some conditions. It does not establish adequate sharpness, sampling or contrast in the failing stream.

Verify inner-corner counts, not just square counts

Most conventional checkerboard calibration tools request the number of internal intersections along each axis. Those are not the numbers of printed squares.

For a complete rectangular checkerboard:

For example, a target containing 10 squares across and 8 squares down has 9 × 7 inner corners. Enter those dimensions according to the software's documented ordering convention.

Check the actual target in the image, rather than relying on its original print file. A fold, obstruction or cropped edge may make the visible pattern different from the intended board.

Avoid folding a larger target into a smaller one. The fold can introduce curvature, shadows and an ambiguous outer pattern. Print a complete target with the required dimensions instead, and retain the surrounding margin recommended by the detector documentation.

Open one saved image at native resolution and check:

• Every required inner intersection is visible.
• The outermost rows and columns are distinguishable from the background.
• No label, crease or reflection interrupts the pattern.
• The configured detector expects an ordinary checkerboard rather than another target family.

Do not repeatedly adjust calibration coefficients while these checks fail. The detector must first identify the intended geometry.

Measure square pitch and keep the board flat

Square size and corner count solve different problems. Corner count tells the detector which pattern to find. Square pitch defines the physical spacing assigned to those detected points.

In many conventional workflows, an incorrect square-size value does not prevent initial corner detection. Instead, it gives the estimated translations the wrong physical scale. Other implementations may use size information during processing, so follow the software's documented requirements.

Measure the printed target rather than assuming the printer preserved its design dimensions. Measuring several square pitches reduces the relative effect of ruler placement error.

For example, if 10 consecutive pitches measure 200 mm:

Use the unit required by the calibration software. Do not enter a millimetre value into a field expecting metres.

Check horizontal and vertical pitch separately. Unequal print scaling produces rectangular cells, which violate a square-grid model. Also inspect the mounted target for curling, bubbles and creases. The calibration target flatness guide explains when a paper target is adequate and when a more stable backing is justified.

Flatness usually matters most after detection succeeds, because deformation can bias calibration without causing obvious detection failures.

Check corner detail in the actual streamed images

A sharp-looking preview can hide insufficient detail. Preview windows may resize images, while the detector receives a different resolution, crop or processed stream.

Save the exact frames supplied to calibration. Inspect them without display scaling, especially at the farthest and most tilted target positions you intend to use.

For a nearly front-facing target, projected square width can be estimated from the horizontal field of view:

Suppose a 1280-pixel-wide image covers 400 mm horizontally and the target has a 20 mm square pitch:

A target that is 200 mm wide would occupy approximately 640 pixels across under the same conditions. Tilting it reduces projected spacing along the foreshortened direction, so examine the smallest visible squares rather than relying only on the front-facing estimate.

There is no universal detector threshold in pixels per square. As a practical starting point, aim for several tens of pixels across each square, then verify detection using your actual images. If you choose 30 pixels per square for initial planning, the example requires:

Use the camera resolution selection guide to distinguish nominal pixel count from usable feature detail. The calculator can support this initial sampling estimate, but it cannot predict blur, compression or detector performance.

Moving the target closer can improve sampling, but may cut off its edges or exceed the available focus range. The target coverage and lens margin guide covers that tradeoff.

Separate doubled images from geometric distortion

Lens distortion changes where features appear. It does not normally create a second transparent copy of the entire checkerboard. Treat doubled edges as an image-formation or acquisition problem before attempting calibration.

Use a rigidly mounted camera and a stationary target under steady illumination. Then compare capture paths methodically:

1. Save several consecutive frames directly from the path used by the calibration application.
2. Save images through the manufacturer's supported viewer, if available.
3. Match resolution, crop, exposure and processing settings wherever possible.
4. Compare full-resolution files, not screenshots of differently scaled previews.
5. Introduce a small target movement and observe whether the duplicate follows an earlier position.
6. Repeat with moving background objects removed from the scene.

A clean still image and a poor stream do not automatically prove a driver defect. The still-image path may use different exposure, resolution, compression or processing. Record those differences before drawing conclusions.

In custom acquisition code, verify that the detector receives a complete frame with the documented pixel format, dimensions and row stride. If a capture buffer can be reused, retain or copy it according to the acquisition API before processing asynchronously.

For queue and timing problems, the frame-to-result timestamp guide helps separate stale frames from current detections.

Remove blur and contrast problems without hiding them

Once the target definition and capture path are correct, address ordinary image quality. Focus on the target plane using native-resolution crops, then inspect both the centre and image edges.

Keep the target stationary during each exposure. Exposure time is still relevant when the board is hand-held or the camera vibrates.

An approximate image-motion calculation is:

Using the earlier imaging geometry, the scale is 1280 / 400 = 3.2 pixels/mm. At a target speed of 80 mm/s and an exposure of 0.005 s:

Reducing exposure to 0.001 s lowers that estimate to about 0.26 pixels. These are motion estimates, not detector acceptance limits, and they exclude vibration and other blur sources.

Add controlled illumination rather than relying on aggressive gain or sharpening. The calibration target lighting guide explains how front lighting can support shorter exposures.

Check white cells for clipping and black cells for lost contrast. Glare can erase intersections even when average brightness looks acceptable. Once exposure is suitable, hold relevant settings constant while collecting the dataset. Use per-frame exposure and gain verification when requested settings and observed brightness disagree.

Verify detection before collecting calibration views

First establish reliable detection on a stationary, moderately sized, nearly front-facing board. This is a diagnostic baseline, not the complete calibration dataset.

Next, test deliberately different positions:

• Near the image centre and toward each edge.
• At several orientations with clearly resolved corners.
• At different distances within the intended operating range.
• With enough border visible to preserve the complete pattern.

Log which conditions fail. A failure confined to distant views suggests sampling or focus. Failure under one lighting angle suggests glare. Failure only near the boundary suggests target clipping, peripheral image quality or strong distortion.

Some wide-angle workflows require detectors and initialization methods designed for substantial distortion. If sharp, correctly described targets still fail, check the calibration tool's documented lens-model support rather than assuming every detector handles every field of view.

Collect calibration images at the final image geometry. Resolution changes, cropping and resampling can change the applicable camera matrix. If calibration output later has unexpected borders, investigate cropping after undistortion separately from initial checkerboard detection.

Decide whether hardware changes are justified

Do not replace a camera solely because another capture path gives a cleaner-looking preview. First demonstrate that the failing images cannot meet the required sampling, focus, motion or acquisition conditions.

When evaluating area scan cameras, compare the required stream configuration through the camera finder. Interface choice matters for system integration, but neither interface type nor nominal resolution guarantees checkerboard detection.

The LS-CMR-0053 is a GigE option to evaluate against your acquisition requirements.

The LS-CMR-0166 is a USB3 option for a different host connection.

Neither listing establishes compatibility with your calibration application. Confirm software support, required image modes, exposure controls and access to suitably detailed frames before ordering.

If the limitation is optical, review machine vision lenses through the lens finder. Check sensor coverage, mount compatibility, working distance and focus range. A higher-resolution camera cannot restore detail that the lens, target or acquisition pipeline has already removed.

Final checks before repeating the calibration

Use this release checklist before collecting a new dataset:

• Pattern: configured dimensions match the visible inner-corner grid.
• Scale: measured square pitch uses the software's required unit.
• Target: the complete pattern is flat, undamaged and clearly separated from its background.
• Sampling: the smallest projected squares retain distinct, sharp intersections.
• Capture: saved calibration-input frames are free from unexplained duplicates and tearing.
• Settings: exposure, gain, resolution and processing are recorded and controlled.
• Coverage: accepted views span useful positions and orientations without clipping.
• Validation: independent images show consistent geometry, not merely a low fitting error.

If those checks identify a genuine hardware constraint, start the camera configuration process. For an unresolved failure, send representative full-resolution frames, target dimensions and capture settings through technical support before selecting replacement parts.