# camera-parameters(3) — intrinsics, distortion, and what they cost you

## NAME

camera parameters — the numbers that separate a picture from a measurement

---

## SYNOPSIS

```
        | fx  0  cx |
    K = |  0 fy  cy |      distortion: k1 k2 p1 p2 k3
        |  0  0   1 |
```

## DESCRIPTION

Intrinsics describe the camera itself: focal length in pixels (fx, fy),
principal point (cx, cy), and the lens's radial and tangential
distortion. They are properties of a lens at a focus setting, not of a
scene. Solve them once, per lens, and reuse them forever — as long as
nobody touches the zoom ring.

Extrinsics are where that camera was standing: a rotation and a
translation. Those change every time someone bumps the tripod.

## FIELD NOTES

- **Calibrate the lens you are going to use, at the focus you are going
  to use.** A borrowed intrinsics file is a guess with decimal places.
- **k1 does most of the work.** Fit k1 and k2 first, look at the
  residual, and only add p1/p2 and k3 if the image actually needs them.
  More parameters always fit your calibration images better and often
  fit the real world worse.
- **Cover the frame, especially the corners.** Distortion is a radial
  function; targets clustered in the middle leave it unconstrained where
  it matters most. Tilt the board. Fill the edges.
- **Reprojection error under half a pixel is a good day.** Above one and
  a half, something is wrong with the board, the focus, or the assumption
  that the sensor was not moving.
- **Autofocus and image stabilization will invalidate your solve
  silently.** Turn both off, tape the ring, note it in the runbook.
- **Undistort before every geometric step.** Homographies, point
  triangulation, and touch mapping all assume a pinhole camera. Feed one
  a distorted image and it will return a confident wrong answer.

## SEE ALSO

`man homography`, `man projection`, `fencepost`