When to recalibrate camera lens distortion coefficients

A calibration file sits in every vision station: k1, k2, p1, p2, the distortion coefficients measured once and trusted ever after. Then a lens cracks, a focus ring nudges, a filter gets swapped, and the corrected image quietly stops being metric. The distortion profile is not a property of the lens model or the camera model. It belongs to one physical camera and one physical lens, mounted and focused one way. This guide gives you the rule for when to recalibrate, the shop-floor events that invalidate a profile, the changes that do not, and the tolerance arithmetic that decides whether a stale profile can still pass your measurement.

What k1, k2, p1 and p2 correct

Machine-vision calibration libraries describe distortion with the Brown-Conrady equations: radial terms that pull image points toward or away from the centre, and tangential terms that absorb the skew introduced by elements sitting slightly off the common axis. k1 dominates near the centre, k2 refines the mid-field, k3 earns its keep only on very wide lenses, and p1, p2 stay close to zero on a decently mounted FA lens.

x and y are measured from the principal point, the spot where the lens axis actually lands on the sensor. Most libraries normalize them so the coefficients come out small and dimensionless; some toolchains store the same model in raw pixel coordinates, which changes every number. Open-source calibration projects have long issue threads about exactly that unit mismatch. Know which convention your file uses before you rescale anything, and if the signs and magnitudes look strange, the guide to reading distortion coefficients on an unrectified image walks through them. The machine vision glossary for lens choice keeps the surrounding terms straight.

Why the profile belongs to one camera and one lens

Distortion coefficients are intrinsic parameters: they describe the optics, not the scene, so they do not depend on what the camera is looking at. That is their strength and their trap. Element decentering and holder alignment sit inside manufacturing tolerances, so two lenses of the same model produce different p1 and p2, and their radial curves differ by hundredths of a percent. The principal point lands where that barrel's axis meets that sensor, so the camera contributes its own spread. A profile measured on station A describes station A alone.

Events that invalidate a calibration

Recalibration is event-driven, not calendar-driven. The trigger list:

• Another lens, even the same model. New glass, new profile; a different focal length needs one for the same reason.
• A moved focus or iris ring. The profile was measured at one focus setting and aperture, and turning either moves the elements. Refocus for a new working distance and the old k1 no longer describes the optics.
• A filter, protective window, spacer or adapter added or removed. Glass in the path and a changed flange distance both move the geometry. See what a protective filter window changes, and the C-mount vs CS-mount guide for the adapter side.
• A drop, a shock, or a mount that has walked. Vibration moves a lens in its holder over months. Cells that shake around the clock keep the profile honest by keeping the optics still:

The LS-FFL-0475 is a 36.61 mm, 2/3" anti-vibration C-mount lens, the class of part for exactly those cells.

• A camera or sensor swap. New sensor, new principal point and pixel grid; start from scratch.
• A large temperature swing. Barrel and mount materials expand at different rates and move the elements; the infrared athermalization guide shows how lens designers counter exactly this drift.

Changes that leave the profile alone

• Moving the camera-lens rig as one unit. Only the pose changes; the intrinsics survive. Repositioning the station or shifting the field down the conveyor needs no recalibration, as long as nobody touches focus.
• Exposure, gain, lighting, trigger timing. Photometric changes leave the geometry untouched.
• A different product or scene. The coefficients are scene-independent, by definition.
• Resolution, ROI and binning changes. With normalized coefficients, rescale fx, fy and the principal point and keep k and p. With pixel-unit coefficient sets, rescale or re-derive everything together; the guide on rescaling or recalibrating after a resolution mismatch works that decision through for stereo, and the same arithmetic carries over.

Worked example: 0.3% distortion against a ±0.1 mm tolerance on a 100 mm field

Take a dimension-check cell that must place a corner feature within ±0.1 mm on a part filling a 100 mm field.

Inputs, with their sources:

• Lens: 16 mm focal length, distortion below 0.3% (catalog, the LS-FFL-0567 below)
• Camera: 20 MP, 1" format (catalog, the MV-CS200-10UM below)
• 1" format geometry: 13.2 mm wide, 15.9 mm diagonal (sensor-format standard)
• Pixel pitch: 13,200 µm across 5,472 columns = 2.4 µm (sensor published geometry)
• Field: 100 mm across the sensor width (application)
• Tolerance: ±0.1 mm at the field corner (application)

The magnification is m = 13.2 mm / 100 mm = 0.132. A 16 mm lens reaches it at a working distance of WD = f (1 + 1/m) = 16 mm × (1 + 1/0.132) = 137 mm.

The corner sits at half the diagonal, r = 15.9 mm / 2 = 7.95 mm. Uncorrected, distortion moves it by 0.003 × 7.95 mm = 23.9 µm, which is 23.9 µm / 2.4 µm ≈ 10 px on this sensor. Push it through the magnification: 23.9 µm / 0.132 = 181 µm ≈ 0.18 mm in the plane of the part.

Now the decision. Without undistortion the corner lies 0.18 mm from where the software thinks it is, nearly twice the ±0.1 mm tolerance; this cell cannot run uncalibrated. With a competent calibration the fit residual lands at 0.1-0.3 px, so the surviving object-space error is 0.3 px × 2.4 µm / 0.132 ≈ 5.5 µm ≈ 0.0055 mm, an 18× margin under the tolerance. Calibrated, the cell passes; stale, it scraps parts.

Rerun the geometry for your own sensor and field:

How the lens choice sets the recalibration burden

The wider the field, the more the corner coefficients matter. An 8 mm lens such as the LS-FFL-1582 covers a 57.6° field of view; its corner radius loads the r⁴ term hardest, so k1 and k2 run at their largest and a stale profile costs the most. A 16 mm FA lens with a printed distortion limit hands you a budget you can run through the arithmetic above. A lens sold as low distortion with no number gives you nothing to compute; measure it once on a checkerboard and treat that measurement as the budget. Telecentric lenses sit at the other end: their distortion is usually specified an order of magnitude below a standard FA lens, and with magnification fixed, the thing that drifts after a knock is scale, not shape.

The three families, side by side:

All 3 share: Mount C-mount. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

The LS-TCL-0574 works at fixed 0.4x magnification with a 227 mm working distance on a 1.2" sensor. A station built around a telecentric recalibrates scale after a mechanical event and treats k1 as a rounding error. For the surrounding lens-choice material, the lenses topic page collects the guides.

The recalibration run and its acceptance check

1. Put a flat, rigid target in front of the lens. Paper on glass passes at short working distances; the target flatness guide draws that line. The target must reach into the frame corners, because the corner coefficients feed on those poses, and the coverage margins guide sizes it.
2. Grab 15-25 poses across the frame and tilt the target 20-45° in at least a third of them. If corners refuse to detect, the checkerboard troubleshooting guide lists the usual causes.
3. Lock the focus and iris before the first pose; the focus position guide shows a repeatable way to set and mark them.
4. Solve, then read the residual. A machine-vision lens on a rigid mount should land at 0.1-0.3 px RMS. Above about 0.5 px, fix the cause — warped target, blur, a moving lens — instead of shipping the fit.
5. Store the paper trail with the numbers: camera serial, lens serial, focus and iris positions, filter list, and the coefficient convention.

Decision rules: the budget against the tolerance

Reduce the whole question to one number, the uncorrected object-space budget:

d is the lens's printed maximum distortion, D the sensor diagonal, m the magnification, and B comes out in object units. In the example, B = 0.003 × (15.9 mm / 2) / 0.132 = 0.18 mm against a tolerance T = 0.1 mm.

• B ≥ T: a stale profile can fail parts on its own. Recalibrate after every trigger event, before the next production run.
• T / 3 ≤ B < T: recalibrate on events and re-check the residual monthly.
• B < T / 10: distortion barely reaches the tolerance. One calibration at set-up and an annual check is plenty, and you can question whether the undistortion warp belongs in the real-time path at all.

Set the re-check rhythm from the corrected margin, T divided by the surviving residual error: the example's 18× supports an annual look, margins between 5× and 15× deserve a quarterly one, and anything under 5× gets a monthly check. A guidance cell measuring to ±1 mm over a 30 mm field lands at B ≈ 0.05 mm, deep in the third row, and can go long stretches without recalibrating. When the measurement reads lengths from image markers, the length measurement calibration guide extends the same arithmetic, and the object-space pixel size guide converts magnification into millimetres per pixel.

When the usual answer is wrong

• A coefficient set in pixel units meets a resolution or binning change. Recalibrating wastes an hour of checkerboard poses; rescaling fx, fy, the principal point and the k, p terms is plain arithmetic and the correct move.
• Telecentric stations. k1 is close to nothing; the failure mode after a knock is magnification and telecentricity drift, so calibrate scale, not shape.
• Stereo and multi-camera rigs. A bumped camera breaks its own intrinsics and the pair's extrinsics at once. Re-run the paired-checkerboard stereo calibration rather than patching one camera and hoping.
• Lenses that refocus across a wide range. One profile per focus stop, or per production distance band, when you measure at more than one distance.
• No printed distortion limit. Then B is unknown. Measure the lens once on the checkerboard and treat the measured corner displacement as the budget.

Calibration hygiene checklist

• Record camera serial, lens serial, focus and iris ring positions, and the filter list with every coefficient set.
• Note the coefficient convention, normalized radius or pixel units, next to the numbers themselves.
• Photograph the ring markings after calibration; a photo costs seconds and settles arguments months later.
• After any event on the trigger list, run a residual check and recalibrate when it moves past about 0.5 px.
• Book the annual residual check when the corrected margin is 15× or better; go monthly under 5×.
• Keep a golden pose set on file: re-solve against it to separate optics drift from target wear.

When you specify the next station, favour lenses with a printed distortion limit and mounts that hold focus. Start in the lens finder, run the chain through the configurator at the lens step, and pick the sensor first in the area scan camera finder when resolution decides the tolerance. If a profile keeps drifting on a live cell, reach us through the contact page and we will walk the budget with you.