03Detection, calibration and synthetic

Spiideo SoccerNet SynLoc

A localisation benchmark built from synthetic images with exact answers: find every player in a wide-angle frame and say where they stand on the pitch, in metres.

  • Synthetic images
  • World coordinates
  • Single frame
Every player gets a ground point, and the answer is scored in metres, not pixels.
Modality
Synthetic and real images with player positions in world coordinates
Task
Single-frame athlete localisation, image to world
Possession labels
None
Video
Images, not sequences
Licence
Set by Spiideo on the dataset host
Access
Public development kit with data links
Paper
VISAPP 2025

What it is

Spiideo makes automated wide-angle camera systems for football. SynLoc is the localisation benchmark it contributes to SoccerNet: rendered scenes where each player’s true position on the pitch is known exactly, and a task that asks for those positions directly from a single image.

What it teaches

Signal What you get
Possession Nothing direct
Tracking Per-frame player positions on the pitch, no identity over time
Events Nothing
Calibration Camera-aware localisation: the projection is part of the problem

Where it wins

  • Exact labels. A renderer knows where every foot touches the ground. Human annotation of real footage never does.
  • The right metric. Scoring in world coordinates rewards the output a tracking pipeline actually needs.
  • Controlled variation. Cameras, stadiums and conditions can be varied systematically.

What it costs

  • The sim-to-real gap. Rendered players are cleaner than real ones. Validation on real footage is not optional.
  • No time. Single frames. No tracks, no ball ownership, no events.
  • Players only. It localises people, not the ball.

How it plugs into possession

It attacks the same quantity Alfheim measures: position error in metres. A box that is two pixels off at the far touchline of a wide-angle view is a metre off on the grass. Training a detector against world-coordinate truth shrinks that error at the source.

For possession, the gain is at the margins that decide it: which of two close players is nearer the ball, and whether anyone is within playing distance at all.

Our read

Synthetic data used for the one thing it is unbeatable at: exact ground truth for geometry. A detector that is scored in metres learns the right habits.

Sources

  1. Development kit and data links (sskit)github.com
  2. Paper reference (VISAPP 2025)github.com