Camera calibration: move the camera or the chessboard?

You want to calibrate a camera with a chessboard, and every capture guide shows the camera bolted down while someone waves the board around. Your rig is the opposite: the board is fixed and the camera can move. For a single camera the two procedures are equivalent, because calibration only ever sees the relative pose between camera and board in each frame. This explainer shows why, gives the pose geometry that drives accuracy, works the depth-of-field cost of tilted views with real numbers, and names the cases where the answer flips: stereo pairs and line scan.

Yes: single-camera calibration never asks which side moved

A monocular calibration estimates one camera matrix — focal length, principal point and skew — plus the lens distortion coefficients. Every photo of a flat target with known spacing contributes one plane-to-plane mapping between the board and the sensor, and the intrinsics are the only parameters shared across all of those mappings. The board's pose enters each image separately, and it is always the relative pose between board and camera. Nothing in the computation references a world frame or asks which object was in your hand.

Here r1, r2 and t are that image's board-to-camera rotation columns and translation. Whether you realize a pose by shifting the board 50 mm left or the camera 50 mm right, the image records the same relative geometry and the solver recovers the same numbers. The camera-fixed, board-moving procedure in most guides is a bench convention, not a requirement of the mathematics. The per-view poses that come out are written in each shot's own board frame and are discarded once the intrinsics converge.

That last point matters on machines: a fixed overhead camera guiding a robot gets its world-frame pose from a later extrinsic or hand-eye step, solved with the camera bolted where it will run. Camera intrinsics for hand-eye calibration: fix the pose first separates the two steps, hand-eye calibration target frame missing: what to check covers the target side, and the relative-pose logic holds for any planar target, chessboard or not — single camera calibration when your device has no stereo pair applies the same reasoning to a depth module.

The views decide the accuracy, not the mover

Collect 10 to 20 usable views. Tilt the target up to about 45° to the lens axis: tilt is what separates focal length from scene translation, and fronto-parallel views leave those two parameters fighting each other. Place the target in every quadrant of the frame, corners included, because distortion is only measurable away from the image center. Spread the working distances across roughly a 2:1 range; a set captured at a single distance conditions the focal length poorly.

Two failure modes dominate real captures: focus creep and missed corners. A focus ring that drifts between frames changes the effective focal length mid-run, so lock it before the first shot — calibrate a camera with autofocus: lock the focus first covers the autofocus case. Steep tilts foreshorten the pattern, and past roughly 45° corners begin dropping out of detection, while every extra frame adds another chance for a miss. Chessboard not detected in calibration: lighting and focus fixes lists the usual causes.

Worked example: paying for a 45° tilt with f-number and distance

The set-up: a 76 mm square target; a 1/1.7" sensor about 7.4 mm wide with 4000 pixels across it, so 7.4 mm ÷ 4000 px = 1.85 µm per pixel; a 12 mm lens; and a corner-blur allowance of 2 pixels, giving a circle of confusion of 2 × 1.85 µm ≈ 0.004 mm. Sensor width and pixel pitch are the standard figures for that format; board, lens and tilt are this example's choices. Three lines of thin-lens arithmetic do the rest.

1. Close views first: let the board fill 70% of the frame width. FOV = 76 mm ÷ 0.70 = 109 mm, so call it 110 mm.
2. Magnification m = 7.4 mm ÷ 110 mm = 0.067.
3. Working distance WD = 12 mm × 1.067 ÷ 0.067 = 191 mm. Set the stand near 190 mm.
4. Tilt the board 45°. The near edge comes 38 mm × sin 45° = 27 mm closer and the far edge moves 27 mm away: a 54 mm depth span, 163 mm to 217 mm.
5. Depth of field at f/8: DOF ≈ 2 × 8 × 0.004 mm × (190 mm)² ÷ (12 mm)² = 16 mm. The span needs 54 mm, so every tilted frame would carry soft corners.
6. Stop down to f/16 and pull back to 250 mm: DOF ≈ 2 × 16 × 0.004 mm × (250 mm)² ÷ (12 mm)² = 56 mm, which covers the 54 mm span.
7. The new view: m = 12 ÷ (250 − 12) = 0.0504, FOV = 7.4 mm ÷ 0.0504 = 147 mm, and the board fills 76 ÷ 147 = 52% of the width.
8. Light: f/8 to f/16 is two stops, one quarter of the photons. If 10 ms was right at f/8, budget 40 ms at f/16 — free on a stand with a still board.
9. Scale check: 147 mm ÷ 4000 px = 37 µm per pixel on the object, so a 0.5 mm dot spans 14 px, far above the 2 px blur allowance.

The rule those numbers leave you with: the tilt span — half the board times the sine of the tilt — must fit inside the depth of field. If the lens stops at f/8, cap the tilt at 25°, because 38 mm × sin 25° = 16 mm, matching the f/8 depth of field at 190 mm. If the iris reaches f/16, buy distance instead: at 250 mm the full 45° tilt fits with margin. The DOF line is the thin-lens estimate; at this magnification, m ≈ 0.05, it lands within a few percent of the exact hyperfocal form.

Stereo flips the answer: lock both cameras, move only the board

A stereo calibration estimates intrinsics for each camera plus one rigid transform between them: the rotation and translation of the second camera in the first camera's frame. Every image pair is assumed to share that single transform. Move either camera between captures and the assumption breaks; the solver then averages incompatible geometry into one wrong transform and the depth output is quietly wrong. With a pair, both cameras bolt to a rigid bar and only the target moves. Stereo depth camera calibration: target capture and stream setup walks the capture, and stereo calibration with angled cameras: what still works covers converging rigs.

Line scan: motion is part of the camera

A line scan sensor images one line per encoder tick; without relative motion there is no image at all. Fix the camera, sweep the target past it at constant speed, and calibrate scale against encoder counts. Line scan camera calibration with a moving checkerboard target details the moving-target procedure.

Keeping a moving camera sharp

Handheld camera motion is workable at short exposures: stay at or below about 10 ms per frame and hold still until the frame completes. A stand is the better tool, because poses repeat, the lighting on the target never changes between frames, and the longer exposures that f/16 tilt views need stop being a constraint. The LS-MNT-0040 camera stand is a professional slide-table type with a height of 700 mm on a 400 x 450 mm base.

For the camera itself, the interface matters more than the shutter. Two 12 MP cameras share the same 1/1.7" rolling-shutter sensor, one USB3 at 30 fps and one GigE at 9 fps; every calibration frame is a still, so rolling shutter adds no skew, and either body calibrates as cleanly as a global shutter sensor of the same format.

All 2 share: Sensor size 1/1.7". Values as printed in each product's data sheet; open a part for its drawing and full specifications.

Moving the camera or the board: which to choose

Decide by rig, not by convention. A small loose board on a bench: moving the board is simpler. A large or fixtured target: move the camera, on a stand, with the focus ring taped. A camera that will live in a cell: calibrate intrinsics wherever it is convenient, then solve extrinsics after mounting — static camera calibration accuracy: why moved parts drift off target explains why a later bump, not the calibration method, is what usually costs accuracy. A stereo pair or a line scan camera: the target moves, never the camera.

Checklist before you accept the result

• Focus and iris locked — a creeping focus ring changes the effective focal length mid-run, and the iris position belongs in the calibration you will actually run.
• 10 to 20 usable views — tilts to about 45°, positions in every quadrant, distances spread about 2:1.
• Every corner found — re-shoot frames where detection failed; a silently dropped view can remove exactly the tilt diversity the set needed.
• Grid pitch entered correctly — a wrong square size scales the whole calibration; checkerboard camera calibration error: fix the grid pitch shows the symptom.
• Residuals reviewed — read the average and worst reprojection errors against sensible limits in camera calibration residual values: average and max targets.
• Rig untouched for pairs — both stereo cameras fixed between every frame, or the transform is void.

With the method settled, the system configurator walks the camera, lens and stand in order; if your target has to stay in a fixture, send its size and working distance through the contact page and we will run the geometry from the worked example for you.