Training-Free Off-Screen Player Imputation for Broadcast-Based Spatial Football Analytics

Jul 16, 2026

Spatial football metrics like pitch control assume position access to all 22 players, but broadcast main-camera views reveal only 10–16 at any moment. This work quantifies the distortion using an open, reproducible benchmark: a simulated broadcast viewport applied to full-pitch tracking data. A visible-only baseline inflates hidden-zone pitch-control error to 25.1–26.9 points and yields 11.1–13.4 points mean absolute control-share error. Training-free online imputations improve on this floor, with role-anchored centroid voting halving hidden-zone error and reducing control-share error to 28–48% of the ignore policy, and enabling decision-impact changes in downstream Space-Creation Index.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages10

Reading Time25 min

Introduction

Contributions

Related Work

arXiv :2607 . 11548v1 [ cs .CV] 13 Jul 2026 Training-Free Off-Screen Player Imputation for Broadcast-Based Spatial Football Analytics Seongjin Choi∗ Abstract Spatial football metrics such as pitch control assume access to the positions of all 22 players, yet the most widely available source of positional data—the broadcast main camera—shows only 10–16 of them at any moment. We quantify the resulting distortion with an open, reproducible benchmark: a simulated broadcast viewport applied to open full-pitch tracking data (Metrica Sports; three matches, one held out from method development) . Ignoring off-screen players—the visible-only baseline implied whenever a video-based game-state-reconstruction (GSR) pipeline adds no imputation layer—inflates hidden-zone pitch-control error to 25.1–26.9 percentage points and produces a mean absolute control-share error of 11 .1–13.4 points across the three matches. We then evaluate a ladder of training-free, online imputation baselines that use only observations from the match being analysed. The best overall on these decision-relevant metrics, role-anchored centroid voting—each visible player votes for the full-team centroid by subtracting its running role offset, attenuating the viewport-induced subset bias—roughly halves hidden-zone error (to 12.2–13.8 points) and cuts the control-share error to 28–48% of the ignore policy at every viewport width from 36 m to 60 m in all three matches (4.5–4.7 points at W = 44 m) . In the short-occlusion regime (≤9.6 s) covered by the closest learned prior work, our training-free method reaches binwise median position errors of 3.3–8.9 m; 50–57% of hidden-player observations under the simulated viewport, however, lie beyond that regime, outside the 9.6 s sequence protocol of that prior work. We integrate the method end-to-end into a broadcast-video GSR pipeline and show that imputation changes a downstream possession-quality score (Space-Creation Index) by 15.6 and 17.2 points on two real World Cup broadcast windows, flipping the verdict class—under prespecified operational thresholds—in one of them. 1 Introduction Video-based game state reconstruction (GSR) turns ordinary broadcast footage into player coordinates on a common pitch template, promising tracking-level analytics without stadium hardware [12] . But the broadcast main camera pans and zooms with the ball: in our World Cup broadcast clips only 10–16 of the 22 players are visible in a typical frame, consistent with the 9–12 reported for broadcast video by Ochin et al. [9] and the 12.8±3.7 of Omidshafiei et al. [10] . Team-level spatial metrics—pitch control [13, 14], space value, block compactness—are then computed over a biased subset of players: the team whose defenders sit off-screen silently cedes control of the hidden half of the pitch, not because of anything that happened on the pitch, but because of where the camera pointed. This distortion is routinely acknowledged and rarely measured. The GSR evaluation protocol itself (GS-HOTA [12]) scores only the players that are visible; what happens to downstream teamlevel metrics when the invisible remainder is ignored has, to our knowledge, not been quantified ∗ Independent researcher. ORCID: [https://orcid.org/0009-0001-3193-7424](https://orcid.org/0009-0001-3193-7424. Code)[. Code](https://orcid.org/0009-0001-3193-7424. Code) and data: [https://github](https://github). com/nowayfootball/offscreen-impute. Contact: nowayfootball@gmail .com. on open data. The one closely related study, DeepMind’s Graph Imputer [10], demonstrated the problem and a learned solution, but on 105 proprietary Premier League tracking matches, with a bidirectional model that consumes future observations inside 9.6-second windows — neither reproducible nor applicable in a strictly online setting. This paper makes three contributions: 1. A reproducible distortion benchmark. We simulate a broadcast viewport (ball-tracking pan with lag, width W ) over o

FAQs

Why does broadcast-based spatial football analytics distort pitch control measurements?

Because the broadcast main camera pans and zooms with the ball, only a subset of players is visible in each frame, so metrics computed over visible players ignore the hidden portion of the pitch.

How is the distortion caused by off-screen players quantified in this paper?

The paper introduces a reproducible benchmark that simulates a broadcast viewport over open full-pitch tracking data and evaluates imputation policies using hidden-zone pitch-control error and related team-control metrics.

What training-free imputation baseline performs best, and what improvements does it bring?

Role-anchored centroid voting is best, where each visible player votes for the full-team centroid after subtracting its running role offset. It roughly halves hidden-zone error and reduces control-share error to 28–48% of the visible-only ignore policy across tested viewport widths.

Jul 16, 2026

Upload by Theodora