Checkerboard calibration targets: image quality requirements

A checkerboard detector is a machine for finding one thing: crisp boundaries between dark and light squares. Feed it a blurry, glare-washed or noisy frame and it finds nothing, and the error will not tell you whether the board, the light or the focus is at fault. This guide turns the way these detectors work into capture-side numbers: pixels per square, aperture, exposure, lighting and framing. A worked example then runs the arithmetic from board size to focal length, so you can set the shot before you press the trigger and know the corners will come back.

How a checkerboard detector reads your image

Every checkerboard detector in common use runs the same three stages. It thresholds the frame, usually with a local adaptive threshold, into a black-and-white map. It finds the dark regions shaped like squares of similar size and checks that they form a consistent quadrilateral grid: neighbours share corners, rows and columns alternate, spacing is regular. It then takes each rough corner and refines it to subpixel accuracy by fitting the two intensity edges that meet there.

Assisted tools run the same logic. The automatic checkerboard corner detection tool published by NIST starts from four corners you click by hand and leaves the grid walk and the refinement to the software, so accuracy still rides on the image the clicks seed (NIST checkerboard corner detection tool).

Each stage sets a requirement you can control. The threshold stage needs real black-white contrast, which is lighting and exposure. The grid stage needs every square complete and inside the frame, which is framing and glare control. The refinement stage needs edges that span a few clean pixels, which is focus, aperture and noise. The sections below put numbers on each.

Pixels per square: the number that decides detection

The detector consumes squares as pixel regions, so the printed size of the squares matters less than how many pixels each side spans. The working rule: aim for at least 20 pixels per square side, treat 8 pixels as the floor below which squares tear or merge during thresholding, and expect little further gain past about 25, where blur and lighting set the accuracy instead. Tilt compresses one axis by the cosine of the tilt angle, so keep the floor in mind for the compressed axis too.

Two things follow. Square size in millimetres cancels out: a 10 mm square at 20 px/mm and a 20 mm square at 10 px/mm both give 200 pixels per side. The number you control at the bench is the board's share of the frame, so zoom or step closer until the squares are big enough. How many squares to print is the companion question — checkerboard calibration target size: how many squares and how big works it with the same arithmetic.

Sharp edges: focus, aperture and motion

Corner refinement fits the intensity transition at each edge, so the transition must stay narrow. Diffraction sets the hard floor:

On 3.4-3.45 µm pixels those disks span 1.1, 2.2, 3.1 and 6.2 pixels. Keep the blur disk under about 2 pixels: on this pixel pitch that means F5.6 or brighter. A flat, rigid board needs almost no depth of field, so there is nothing to trade against; stop down only when the board cannot sit at one distance. The diffraction arithmetic continues in airy disk size and the diffraction limit: match pixel to feature, and the depth side in aperture f-number and depth of field in machine vision cameras.

Motion blurs corners the same way. Keep exposure time × board speed below one pixel's footprint on the board. One pixel spans 0.086 mm on the board at the worked example's scale below, so a board moving 100 mm/s forces an exposure under 0.9 ms, while a static rig has no such limit. Keep gain near zero and take the extra light instead: noise widens every transition and scatters refined corners between repeated shots of the same pose, which surfaces later as calibration scatter you cannot trace.

Even light and a matte board

The threshold stage needs black squares that stay black and white squares that stay white across the whole frame. Light the board diffusely from the front — an LED panel, or two bar lights at 30-45° — so no square carries a shadow or a highlight. Placement rules for the lights are in lighting geometry for image contrast: camera, light and part positions.

Set exposure from the histogram: white squares near 200-230 of 255, black squares near 30-60, transitions in the middle of the range, nothing clipped at either end. A clipped white reads as a bigger blob to the threshold stage; a black square buried in noise tears into fragments. Both break the grid walk before the refinement stage ever sees a corner.

Gloss fails even under good light. The lamp reflects as a saturating highlight that splits squares, and the detector reports a partial grid. Print on matte paper or spray a laminated board matte; if the board must stay shiny, spread the reflection with a diffuse dome. For frames that already came back empty, the checks in checkerboard corners not detected: fix gaps in your target and checkerboard calibration not detecting corners: what to check cover the usual gaps.

Save lossless files — PNG or the camera's raw format. Heavy JPEG compression rings the edges and shifts refined corners by fractions of a pixel; across a calibration set that becomes a systematic error.

Framing, tilt and full grids

Keep the whole grid inside the frame with margin. The grid walk needs each square's neighbours present, so one border square clipped by the frame edge can end detection for the whole board even though every visible corner is perfect. Leave at least one square width of margin on all sides; at the 20-pixel target that margin is nearly free.

Tilt is wanted, not avoided — a calibration set needs the board at many angles. The per-frame limit is arithmetic: foreshortening scales one axis by cos(θ), so a square that spans 174 pixels face-on spans 87 at 60°. Detection survives; corner precision along the compressed axis drops with it. For metrology-grade results keep most poses within about 45° of face-on and let a few go steeper to condition the solve.

Two checks before you shoot. The grid needs odd-by-even inner corners, which why a calibration checkerboard needs odd by even squares explains. At long standoffs the pixel budget, not the detector, becomes the constraint: checkerboard calibration fails at distance: depth camera target size works that case.

Worked example: a 9 x 6 board, 15 mm squares and a 5 MP camera at 400 mm

Inputs, each one a decision you make at the bench:

• Board: 9 x 6 inner corners, which is 10 x 7 squares at 15 mm, so 150 x 105 mm — a common printed target.
• Camera: 2448 x 2048 pixels at 3.45 µm pitch, the standard 5 MP 2/3" class; sensor width = 2448 × 3.45 µm = 8.45 mm.
• Working distance: 400 mm, set by the clearance around the rig.
• Fill: the board spans about 70% of the frame width.

The arithmetic:

1. FOV = 150 mm / 0.70 = 214 mm.
2. f = 400 mm × 8.45 mm / 214 mm = 15.8 mm, so the nearest standard focal length is 16 mm.
3. FOV at 16 mm = 8.45 mm × 400 mm / 16 mm = 211 mm.
4. Scale = 2448 px / 211 mm = 11.6 px/mm.
5. Pixels per square = 11.6 px/mm × 15 mm = 174 px. Cross-check: the board fills 150 / 211 = 71% of the frame, 0.71 × 2448 = 1738 px, over 10 squares = 174 px.

Decision: 174 pixels per square is nine times the 20-pixel target, so pixel count is nowhere near the limit in this set-up. The board could shrink to about 8% of the frame width — 0.08 × 2448 / 10 ≈ 20 px per side — before touching the target, and to about 3% before falling under the 8-pixel floor. The limits that bind are the ones above: F5.6 or brighter, even light, gain low, board rigid. Depth of field is not one of them: at F5.6 with magnification m = 8.45 / 211 = 0.04, DoF ≈ 2 × 5.6 × 0.00345 mm / 0.04² ≈ 24 mm total — the depth of field calculator does this conversion — far more than a flat board needs.

Run your own board, distance and sensor through the focal length calculator:

Common failures and the fix for each

On preprocessing: resist it. Sharpening, contrast stretching and denoising each move edge positions before the detector runs, so you end up calibrating the filter chain along with the camera. Fix exposure, aperture and light at the source instead.

The decisions, and the parts behind them

• Pixels: 8 per square side is the floor, 20 the target; frame the board over at least a tenth of the frame width.
• Aperture: blur disk under 2 pixels, which is F5.6 or brighter on 3.4 µm pixels; stop down only for real depth ranges.
• Exposure: board travel during exposure under one pixel footprint; static rigs expose freely with gain at zero.
• Light: diffuse from the front, matte board, whites 200-230, blacks 30-60, nothing clipped.
• Geometry: full grid in frame with a square of margin; tilt and spread the poses for the solve.

A 16 mm focal length lands almost exactly on the worked example. The LS-FFL-0567 is a 16 mm, 1.1", 20 MP C-mount lens with F2.8-22 and <0.3% distortion. The 16-22 mm image circle leaves margin around an 11 mm diagonal sensor. Tighter fields on a 2/3" sensor come from the 35 mm LS-FFL-0679, a 6 MP-class C-mount lens. Where the camera rides a moving axis, the anti-vibration LS-FFL-0475 holds its 36.61 mm focal length under shake.

None of this survives a wobbling mount, so hold the geometry with a stand: the LS-MNT-0040 slide-table stand gives a 700 mm column and a 400 x 450 mm base.

Pick the lens from the computed focal length and sensor, not the other way round. The machine vision lenses page filters by focal length, sensor format and distortion class, the lens finder narrows by mount and resolution, and the lens configurator walks the chain from board to part number. The light finder filters diffuse panels for the even front light, and camera stands hold the geometry. The lenses topic page collects the rest of the optics and calibration guides.