Lenses & opticsExplainer

Omnidirectional camera lens distortion: model spherical rigs

Which projection model an omnidirectional camera lens follows, why the pinhole picture fails past 120 degrees, and how to calibrate and validate the fit.

LumaSpec Vision9 min read Share on LinkedIn

On this page 9 sections
  1. 01Why one focal length fails past 120 degrees
  2. 02Worked example: a 147.4° diagonal lens on a 1/1.8-inch sensor
  3. 03What the data sheet tells you before you capture anything
  4. 04The projection models you will meet
  5. 05Calibrating the model: frames, fit and acceptance
  6. 06When the task is geometry, not imagery
  7. 07Fields under 120 degrees: stay rectilinear
  8. 08Before you commit: the checklist
  9. 09Questions

A spherical multi-camera rig covers a full sphere with several wide lenses, each past 140° of diagonal field. Past about 120° the flat pinhole picture stops describing the image: one focal length cannot reproduce both the center and the corner. This guide gives you the decision rule by field angle, the projection models you will meet, and a calibration procedure you can run: read the data sheet, fit candidate models from checkerboard frames, then accept or reject the fit on a residual threshold. The same rule tells you when a plain rectilinear lens is still the right part.

Why one focal length fails past 120 degrees

A rectilinear lens maps scene angle θ to image radius through the pinhole relation:

r = f·tan(θ)

The tangent diverges at a 90° half-angle, so 180° is a hard wall. The practical wall arrives earlier. At a 140° diagonal the pinhole focal length falls to 0.18× the sensor diagonal, and the corner samples each scene degree 8.5 times more densely than the center — a starved middle and a wasted rim. Designs that wide stop being buildable with usable correction long before the geometry formally breaks.

Fisheye families remove the wall by mapping the angle itself. The equidistant model, the default for wide lenses, spends image radius linearly on scene angle:

r = f·θ (θ in radians)

The working split: below about 110° diagonal, stay pinhole with a polynomial distortion fit. From 110° to 150°, fit an angular model. Above 150°, angular mapping with per-lens calibration is not optional.

Check what focal length a given field forces on your sensor with the calculator below.

Lens focal length, field of view and working distance
Units
Solve for
Focal length
– mm
Field of view
– mm
Working distance
– mm
Magnification
– x
Angle of view
– °
Pixel on the part
– mm

Thin-lens model: m = sensor width / field of view, f = WD × m / (1 + m). Accurate to a few percent above about 150 mm; check the lens's minimum object distance.

Find lenses with this focal length Open the full calculator

Worked example: a 147.4° diagonal lens on a 1/1.8-inch sensor

Inputs: a per-lens diagonal field of 147.4°, the figure spherical-rig data sheets quote for each lens; a 1/1.8-inch sensor, whose standard image diagonal is 8.9 mm; half-angle 147.4° / 2 = 73.7°.

The rules, in words: a rectilinear lens covering half-angle θ on sensor diagonal d needs f = (d/2) / tan(θ/2); an equidistant lens needs f = (d/2) / (θ/2), θ in radians.

Rectilinear attempt:

f = 4.45 mm / tan(73.7°) = 4.45 mm / 3.42 = 1.30 mm

That is 0.15× the sensor diagonal, shorter than any corrected wide lens is built.

Equidistant attempt:

θ = 73.7° × (π / 180) = 1.286 rad f = 4.45 mm / 1.286 = 3.46 mm

A 3.5 mm-class focal length is an ordinary design value, which is why lenses this wide exist in the equidistant family.

Sampling check: pinhole pixels per degree scale with sec²θ. At the corner, sec²(73.7°) = 1 / cos²(73.7°) = 1 / 0.0788 = 12.7, so a rectilinear design would put 12.7 times more pixels on each rim degree than on each center degree.

Decision: past roughly 120° diagonal, treat the lens as angular and fit the equidistant family. If the maker states a focal length near 3.5 mm for this field on this format, the numbers close in the angular model; no rectilinear design reaches this field.

What the data sheet tells you before you capture anything

Spherical-rig data sheets quote each lens as three angles — in this class, 147.4° diagonal, 113.4° horizontal and 94.8° vertical, at apertures between F2.5 and F4 — and they carry a warning that the figures are rough because distortion is high. Two exact identities tell you which family those numbers live in.

Flat-image identity: tan(θd/2) = √(tan²(θh/2) + tan²(θv/2)). Feeding the sheet in, tan(56.7°) = 1.523 and tan(47.4°) = 1.088, so √(1.523² + 1.088²) = 1.871 and 2·atan(1.871) = 123.8° — about 24° short of the quoted diagonal. The fields do not close in tangent space.

Angular-image identity: when radius is proportional to angle, θd = √(θh² + θv²). Check: √(56.7² + 47.4²) = √5462 = 73.9°, against the quoted half-diagonal of 73.7° — agreement within 0.2°. The fields close in angle space.

Decision: the quoted fields are angular, so start from the equidistant family before a single frame is captured. Aperture then follows depth-of-field and aberration needs, as in aperture, f-number and depth of field; the rough per-axis numbers come from distortion bending the axes, so trust the closure test rather than the absolute values.

The projection models you will meet

Model Image radius Typical reach You fit Found on
Pinhole + polynomial f·tanθ plus a radial polynomial to about 120° diagonal f, k1, k2, tangential terms standard machine vision lenses
Equidistant (Kannala-Brandt) f·θ 150° diagonal and beyond f plus an odd polynomial in θ most fisheyes and spherical rigs
Equisolid 2f·sin(θ/2) about 150-200° f plus a correction term prime photographic fisheyes
Orthographic f·sinθ hemisphere, 180° hard limit f plus a correction term ultra-wide and mirror designs
Unified sphere / division sphere mapping, or r_d = r / (1 + ω·r²) to about 190° one to five parameters catadioptric and very wide central cameras
Spline radial map a table of θ(r) knots any knot positions vendor toolchains for spherical rigs

The equidistant family dominates wide lenses because it spreads pixels evenly across the field and stays well-conditioned at the rim. A widely fitted variant multiplies the angle by a four-term odd polynomial, θd = θ·(1 + k1θ² + k2θ⁴ + k3θ⁶ + k4θ⁸) — the model the OpenCV fisheye module documents and fits. Unified sphere and division models trade a little accuracy for one or two parameters; a spline map buys exactness at the cost of a knot table, which is what a maker's B-spline distortion file is.

How each model's radii turn back into pixel coordinates is the undistortion step, covered in fisheye undistortion: how the pixel mapping works; the rig-level join on a sphere is fisheye undistortion and field-of-view loss: stitch on a sphere. For board-level fisheye builds, image circle and back focal length decide which sensor a lens can serve at all — fisheye lens selection for M12 board cameras walks that side, and the lenses topic page gathers the related optics guides.

Calibrating the model: frames, fit and acceptance

  1. Lock exposure and iris first. An auto-iris stepping its f-number between frames changes the aberration balance mid-sequence — see automatic iris, exposure and gain control.
  2. Lay out a planar checkerboard or ChArUco target, flat to about 0.1 mm per 100 mm, and detect corners to sub-pixel.
  3. Capture 25-40 frames: sweep the board through all four quadrants, tilt it from 0° to 60°, and push it into the extreme rim of every corner. The rim cells carry the model choice.
  4. Fit candidates in order: pinhole plus polynomial, equidistant plus odd polynomial, unified sphere, spline. Change family when the residual map says so; do not keep stacking orders onto the wrong mapping.
  5. Judge on overall RMS and on the corner cells. A healthy equidistant fit lands at 0.2-0.4 px RMS; if the error concentrates in the corners, change family rather than adding orders.
  6. Save one parameter set per lens, dated and tied to its aperture — save camera calibration parameters — and keep the files consistent when they feed a stitching or stereo exchange, as in importing stereo calibration parameters across software formats. Refit after any lens swap, focus change or shock.

When the task is geometry, not imagery

If the deliverable is per-pixel range on scenes with structure, skip the projection solve entirely: a depth camera returns range directly. Time-of-flight sensing cameras such as the MV-DT01WMU output depth without any lens-model fit, and the 3D sensors range covers laser profilers and structured-light cameras that measure shape the same direct way.

3D sensorsHikrobot MV-DT01WMU ToF sensing cameraMV-DT01WMU
Type
Depth / ToF camera
Price on requestView and quote

Fields under 120 degrees: stay rectilinear

Part location, code reading and gauging rarely need more than 110°, and a rectilinear lens keeps pixels metric: straight lines stay straight and polynomial distortion corrects to a fraction of a percent. The widest fixed C-mount lenses in the machine vision lens catalog span 8 to 16 mm across 2/3-inch to 1.1-inch sensors.

Specification LS-FFL-1582 LS-FFL-0602 LS-FFL-0567
Flange back 17.526 mm in air – 17.526 mm in air
Back focus 11.5 mm in air – 22.4 mm (min)
Aperture 1:2.8~16 F2.4 Manual iris F2.8-22
Focal length 8 mm 8.5 mm 16 mm
Distortion <0.1% <0.7% –
Resolution 10 MP 10 Megapixel 20 Megapixel
Format – 1" 1.1"
M.O.D. – 0.1 m 0.1 m
Filter thread – M77 x P0.75 M35.5 x P0.5
Price $309.00 $429.00 $679.00

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

Where a data sheet prints distortion outright, the guesswork disappears: the LS-FFL-0567 specifies under 0.3% distortion at 16 mm for 1.1-inch, 20 MP sensors.

Machine vision lenses16 mm, 1.1", 20 MP, C-mount machine vision lens, F2.8-22, <0.3% distortion (LS-FFL-0567)LS-FFL-0567
Type
Machine vision fixed
Sensor format
1.1"
Mount
C-mount
Resolution
16-25 MP
$679.00View partBuild a system with it

The pinhole side of that fit — solving intrinsics when some are already known — is known-intrinsics camera calibration. To pair a focal length with a sensor, use the lens finder or start at the lens step of the configurator.

Before you commit: the checklist

  • Half-angle above 55°, diagonal past 110°: fit an angular model, not pinhole plus polynomial.
  • Sheet fields close in angle space within about 1°: the equidistant family is consistent before capture.
  • 25-40 frames, rim of all four corners covered, tilts to 60°, iris locked at the working aperture.
  • RMS under 0.5 px with rim cells near or under 1 px on the residual map.
  • One dated parameter set per lens, aperture recorded, refit scheduled after swaps or shock.
  • Geometry deliverable with structure in the scene: consider measuring range directly instead of fitting a lens.

Run the closure test on your own data sheet first: it is one line of arithmetic, and it usually settles the family before any capture.

Questions

Does one calibration cover every lens in a spherical rig?

No. Fit each lens separately: focal length, radial terms and principal point vary lens to lens within one rig, and each lens looks at its own patch of the sphere. Share the model family across lenses and solve the extrinsic transforms between cameras; the stitching step then joins the per-lens files. Refit after any lens swap or shock.

How many frames does a wide-lens calibration need?

Plan on 25-40. Below roughly 15 frames the focal length and the highest radial coefficient trade against each other: the summary error looks fine while the rim wanders. Include tilted views between 0° and 60° and frames that push the board into the extreme rim of each corner; frontal-only boards condition the focal length but not the angular terms.

What reprojection error is good enough for a fisheye fit?

Treat 0.5 px RMS as the acceptance line, and read the residual map by cell: a 0.4 px average can hide 1.5 px at the rim. Errors clustered in the corners point to the wrong model family; a ring of error at one radius points to a missing radial term. Both are fixed by changing the model, not by capturing more frames.

Why don't the quoted horizontal, vertical and diagonal fields close under one formula?

On a heavily distorted lens the per-axis fields are measured along image axes that distortion bends, so data sheets mark them rough. Test closure in angle space: the diagonal half-angle should equal √(h² + v²) within about 1°. In tangent space the same three numbers miss by tens of degrees, which is itself the evidence that the lens is angular.

Does a bigger sensor change which model fits?

No. The radial mapping is a property of the lens; the sensor only decides how much of that mapping you sample. The same fisheye on a larger format sees a wider angular slice of the same radius-versus-angle curve, and the fitted coefficients describe the lens, not the sensor. Format changes the field you use, never the model.