Checkerboard camera calibration error: fix the grid pitch

Your checkerboard calibration converges, the solver reports a mean error of 0.08 to 0.21 mm, and the guided system still lands on a different position every cycle. The camera is usually innocent. In almost every case like this, the grid pitch you typed does not match the physical target: printers and plotters scale the drawing when they print it, and on a non-English Windows install a decimal comma can truncate the entry without an error message. This guide ranks the causes, works the arithmetic of a 20 mm square that printed at 19 mm, and gives a gauge check that separates a wrong scale from a wrong calibration in five minutes.

The symptom: a fit that passes, positions that fail

A mean error in the calibration report tells you how tightly the detected corners fit one internally consistent grid. It says nothing about whether that grid's real-world size matches the pitch you entered. Type a pitch 5% larger than the printed square and the solver simply fits a 5% larger world: every corner lands within a fraction of a pixel of the stretched model, so the mean error stays low while every coordinate the system reports comes out 5% long.

That is why the failure looks random. A scale error is deterministic — the offset grows with distance from the calibration origin — but parts arrive at different places in the field of view, so each pick misses by a different amount. Random-looking misses across the tray, with a healthy mean error, are the signature of a wrong pitch, not of a failing solver.

First check: the printed target is not the drawing

Printers scale documents. A fit-to-page setting, a pattern drawn for A3 printed on A4, or an A4 file sent to a letter-size printer commonly moves the scale by 2 to 6% — 20 to 60 mm across a metre of field of view. The calibration screen still accepts the target, because the checkerboard is still a checkerboard; only its size changed.

The test takes a minute: measure outer edge to outer edge across ten squares with calipers and divide by ten. If the result differs from the value you entered by more than 0.2%, you have found the fault.

The fix: print at 100% scale (choose actual size, never fit), let the toner cool, mount the paper flat on aluminium composite or glass, and enter the measured pitch — not the value in the drawing file. Measure every copy: two prints of the same file can differ when one goes through a different tray or driver. The same trap exists outside vision — a motion axis calibrated without its actuator pitch carries the identical assumed-versus-real scale error into every move.

Second check: a decimal comma that truncates the pitch

Calibration software built for English locales parses numbers with C-style routines: reading stops at the first character that cannot belong to a number. Type 19,5 on an Italian, German or French keyboard and the field stores 19 — no warning, no error, half a millimetre gone. The field looks full and correct on screen.

The trap has a benign case: 19,0 truncated to 19 changes nothing, because the fraction was zero. It is fatal whenever the measured pitch carries a fraction, which it usually does once you measure a real print instead of trusting the drawing.

The test: re-enter the pitch with a point, recalibrate, and compare two reported positions of the same stationary part. Identical results mean your comma was harmless; a shift means it was not. If the field refuses a point, change the system region or use the software's numeric control rather than typing the separator.

Third check: the image itself

If both entries check out, look at the images before doubting the algorithm. Blur, a target that is not flat, and low contrast all raise residuals — but they raise the mean error too, which is how you tell them apart from a pure scale error.

• Motion blur: allow no more than half a pixel of smear during the exposure. At 0.11 mm per pixel and a tray moving 50 mm/s, that tolerates 0.5 × 0.11 = 0.055 mm of travel, so exposure must stay at or below 0.055 / 50 = 1.1 ms.
• Flatness: paper glued only at its corners bows a millimetre or more at the centre, and the bowl makes the scale drift across the field. Mount the target on a stiff plate.
• Missing corners: if the detector skips squares, the fit degrades quietly — what to check when checkerboard calibration is not detecting corners covers focus, contrast and square count.
• Acquisition: when the raw image shows vertical banding before any calibration runs, the fault sits in the camera or the frame grabbers, not in your numbers. Fix the image path first.

Worked example: a 20 mm target that printed at 19 mm

The numbers for a typical overhead pick camera, matching the field case:

• Sensor: 1-inch format, 20 MP, 5472 × 3648 pixels, 2.4 µm pixel pitch
• Lens: 16 mm focal length, focused and locked
• Field of view: 600 mm × 400 mm over the tray
• Target: squares drawn at 20.0 mm, printed and then measured at 19.0 mm; the operator entered 20.0 mm from the drawing

Work the magnification first: m = 13.1 / 600 = 0.022, so the lens sits at WD = 16 × 1.022 / 0.022 = 743 mm — call it a 750 mm stand-off. One pixel then covers 600 / 5472 = 0.11 mm on the tray.

The scale error: 20.0 / 19.0 − 1 = 0.0526, so every distance is over-reported by 5.3%. A part sitting 400 mm from the calibration origin is reported at 400 × 1.0526 = 421 mm — a 21 mm pick error, about 190 pixels. The mean error, meanwhile, reads 0.08 to 0.21 mm, which is 0.7 to 1.9 pixels: a pass.

The decision: any scale error above 0.2% — 1 mm on a 500 mm span — means the pitch entry is wrong; re-measure and re-enter. This one is 26 times over the limit.

These numbers suit most current area scan cameras; one match is the MV-CS200-10UM, a 20 MP USB3.0 mono camera with a 1-inch sensor and a C-mount:

To size the camera the other way round — from the smallest feature you must measure to the pixel count you need — run the resolution calculator before you order:

What the mean error cannot tell you

Mean error is a residual: the average distance between each detected corner and the model's prediction of it. A uniform pitch error stretches the model and the corners together, so the residual barely moves — that is how the report passed while picks missed by 21 mm. What the mean error does catch is local damage: blur, a bowed target, a lens model that does not fit, too few poses. As a working rule, 1 to 3 pixels of mean error is normal at this scale; above 3 pixels, fix focus, flatness, lighting and poses before suspecting anything else.

The report also says nothing about where the camera sits in the robot's frame. That is the extrinsic side of the problem, and camera extrinsic calibration accuracy beyond reprojection error covers what a reprojection number can and cannot certify there.

Verify the scale with a 300 mm gauge

The gauge check takes five minutes and needs no robot motion:

1. Machine or buy a bar with two fiducials exactly 300.0 mm apart, certified to ±0.1 mm or better — a CNC-drilled plate or a gauge-maker dot pattern both work.
2. Place it mid-field, roughly along the field diagonal.
3. Read the distance the vision system reports between the two fiducials.
4. Accept only 300.0 ± 0.5 mm. Rotate the bar 90° and repeat: a scale error that hides along one axis shows up across the other.

A 5.3% pitch error reports 300 × 1.0526 = 315.8 mm and fails by 31 times the tolerance band. A correct calibration with sub-pixel fiducial centres sits comfortably inside 0.5 mm.

On the bench, where you prove a camera and its software before hanging them over a line, a manual XY-plus-rotation stage such as the LS-MNT-0220, with a 60 mm base, presents the board flat and at controlled angles:

When the scale checks out but positions still drift

If the gauge passes and picks still miss, the pixel-to-millimetre scale is not your problem:

• The camera moved. A knock or a loosened column lock shifts the robot-to-camera transform; re-run that transform, not the grid calibration, and re-check the hardware on the camera stands column.
• Focus drifted with temperature. A lens that refocuses as the cell warms changes its focal length slightly, and the scale with it — camera focus drift with temperature gives the diagnosis.
• The pose set was poor. If the robot carries the camera, eye in hand calibration pose selection lists the robot positions that make the transform observable.
• The error has one fixed direction. That points at frames and calibration constants rather than scale, covered by fiducial pose at the wrong distance.

Checklist before you recalibrate

Work top to bottom; stop at the first failure and fix it.

Target

• Ten-square span measured with calipers; the measured pitch, not the drawing's, entered
• Printed at 100% (actual size, fit-to-page off), mounted flat on a stiff plate

Entry

• Decimal point in the pitch field, re-typed by hand
• Screenshot of the calibration screen filed with the target ID

Image

• Corners detected on every pose; mean error at or under 3 pixels
• Exposure at or below 1.1 ms whenever the part can move at 50 mm/s

Verification

• 300 mm gauge reads 300.0 ± 0.5 mm in both orientations
• Pitch, target ID, date and gauge reading logged

Next steps

Grid pitch errors are quick to fix and expensive to find after a week of wrong picks; the gauge makes them visible in minutes. If the camera is part of the same decision, start at the camera finder or the system configurator, which walks camera, lens, light and controller together. For the target itself, calibrating with a single square and circular grids on monochrome cameras trade convenience against accuracy, and more systems pieces sit under the systems topic page. Bring the gauge reading and a photo of the target to our applications team and we will work the numbers with you.