05The map

Video-only inference

Possession from pixels: detect the ball and players, calibrate the pitch, project to the ground, then reason. Independent of any supplied tracking, and expensive at every step.

  • Computer vision
  • Detection
  • Calibration
Five stages between a frame and a point on the pitch, and certainty drains at each. The last step, from a ball position to a possession state, is still open at the end.
Inputs
Tactical or broadcast video
Stages
Detect, track, calibrate, project, spot
Wins
Sees contact; recovers what a tracker lost
Costs
Small objects, calibration, non-live footage

The design question

Can the evidence be rebuilt from the picture instead of trusted from a feed? A video-only design owns its own perception. That buys independence from someone else’s tracker, and pays for it with a chain of subproblems that each fail on their own.

The pipeline, stage by stage

Stage Must deliver Where it hurts
Ball detection The ball, most frames A handful of pixels in a tactical view
Player and team segmentation Who is on which side Similar kits, crowded boxes
Identity association The same player across time Occlusion, camera cuts
Pitch calibration Lines and keypoints to a camera model Few landmarks in view
Projection Pixel to pitch coordinate Valid on the ground plane only
Action spotting Pass, header, tackle, out Coarse in time
Live or replay Is this the match clock? Replays, close-ups, crowd shots

The errors multiply along the chain. A pipeline is never better than the product of its stages.

What breaks it in real football

The wide frame holds the whole match in one picture, so the ball is a few bright pixels, and so are a penalty spot and a white boot.
  • Detection is not semantics. A precise ball box says where the ball is. It does not say whether a deflection transferred controlled possession.
  • Geometry and time. The limits of projection and of broadcast timing have their own studies: Space is not time and One clock for the match.

Where it belongs

As one witness among three. A recovered ball position on frames where tracking has none, or a foot-to-ball contact at a closer angle, is one more measurement with its own confidence. How it is reconciled with the others is the single source of truth.

How to judge it

Measure each stage against its own ground truth before measuring possession. Ball recall on the frames the tracker missed is the number that matters. Recall on frames the tracker already has adds nothing.

Our read

Pixels recover what a tracker lost. They do not decide possession. A perfect detector still needs football semantics and a temporal decision behind it, so video earns its place as a witness, not as the system.

Sources

  1. Roboflow sports: soccer examplesgithub.com
  2. SoccerNet Game State Reconstructiongithub.com
  3. PnLCalib: points-and-lines pitch calibrationgithub.com
  4. T-DEED for team ball action spottinggithub.com