05The hard parts

Space is not time

Calibration says where a pixel lands on the pitch in one frame. It says nothing about when. Which calibration to trust, where it stops being true, and why geometry and timing are two separate gates.

  • Homography
  • Calibration
  • Multi-camera
A boot on the grass maps exactly. A ball in flight lands in the wrong place, because height is unseen. And cross-view evidence passes only when the space gate and the time gate are both open.
Spatial problem
Pixel to pitch, one frame at a time
Holds for
Points on the ground plane
A flying ball needs
Full camera geometry and synced views
Rule
Two gates, space and time, each with its own error

The design question

What has to be true before a point seen by one camera can be compared with a point seen by another? Two things, and they are independent: the geometry must be right in each frame, and the frames must show the same moment. What a homography is, and what the supplied ones contain, is in Homography: the bridge between pixels and pitch. This study is about the design choices: which calibration, and what to do where it fails. It is the second prerequisite of the single source of truth.

Calibration options

Method Strength Constraint
Supplied homography Available and documented Accuracy varies shot by shot
Pitch keypoint matching Simple physical correspondence Needs the right landmarks in view
Points and lines refinement Works with part of the pitch visible A misidentified line bends everything
Multi-frame consistency Stabilises a moving camera A cut breaks the smoothness
Full 3D camera model Enables true triangulation More unknowns, synchronised views

Why accuracy varies shot by shot

A wide shot shows many pitch landmarks and the camera is pinned down. A tight shot shows two, and the projected pitch is free to swing. The tight shot is also the one that shows the ball best.

The flat-earth limit

A homography relates the pitch plane to the image. It is exact for a boot on the grass and wrong for a ball in the air. Two homographies from two cameras do not triangulate a ball in flight when its height is unobserved. That needs real camera geometry and detections close enough in time to be the same ball.

What breaks it in real football

  • A fast pan. The camera model lags the picture, and projected players slide across the pitch.
  • A cross. The ball leaves the ground plane and its projected position runs away from its true one.
  • Good geometry, wrong moment. Reprojection error is small and the frames are half a second apart. A sprinting player has moved metres.

How to judge it

Project known pitch points into the image and measure the reprojection error in pixels, then in metres on the pitch, per shot type. Then count how much cross-view evidence survives when both gates are applied.

Our read

Treat calibration and synchronisation as two separate gates, each with its own error. Cross-view evidence passes only when both are open.

Sources

  1. PnLCalib: points-and-lines pitch calibrationgithub.com
  2. TVCalib: camera calibration for sports field registrationgithub.com
  3. OpenCV: camera calibration and 3D reconstructiondocs.opencv.org