Pith. sign in

REVIEW 2 major objections 6 minor 15 references

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

T0 review · 2 major / 6 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Ignoring off-screen players warps pitch-control metrics by 25 points; a training-free centroid-vote imputer halves that error.

desk verdict Solid open benchmark and a simple causal imputer that actually halves the ignore-baseline distortion on decision metrics; narrow data and fixed-width pan are the real limits, not the method math. read the letter →

arxiv 2607.11548 v1 pith:ACDKOA7U submitted 2026-07-13 cs.CV cs.LG

Seongjin Choi This is my paper · ORCID
classification cs.CVcs.LG
keywords playerimputationpitchcontrolbroadcastvideogame-statereconstructionspatialfootballanalyticstraining-freemethodsocclusioncentroidvoting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Broadcast football cameras routinely show only 10–16 of the 22 players, yet spatial metrics such as pitch control are written as if every player is known. This paper measures how badly the missing half distorts those metrics on open full-pitch tracking data with a simulated pan window, then shows that a simple online rule—each visible player votes for the full-team centroid by subtracting its running role offset—cuts the damage roughly in half without any training data or future frames. The same style of imputation, dropped into a real broadcast game-state pipeline, moves a possession-quality score by 15–17 points on two World Cup windows and flips the operational verdict class in one. The practical claim is that any video-based spatial analysis that skips an imputation layer is silently reporting camera framing rather than on-pitch reality, and that a zero-training floor already recovers most of the lost control share.

What carries the argument

Role-anchored centroid voting (B4): each visible player proposes the full-team centroid as its own position minus its EMA role offset; the average of those proposals becomes the imputed team centroid, and hidden players are placed at that centroid plus their own stored offset, attenuating the systematic left/right bias of the camera’s visible subset.

What would settle it

Re-run the same viewport-width sweep and B0-vs-B4 comparison on a larger multi-league full-pitch tracking corpus that also includes zoom and tilt; if B4’s share-error reduction collapses below roughly 30 percent of B0 or the long-occlusion share falls well below 50 percent, the central quantitative claim fails.

Watch

Extended reading notes

Core claim

On three open Metrica matches under a simulated broadcast viewport, the visible-only policy produces 25.1–26.9 percentage-point hidden-zone pitch-control error and 11.1–13.4 points of team control-share error. Role-anchored centroid voting halves the hidden-zone error (to 12.2–13.8 pp) and reduces share error to 28–48 percent of the ignore baseline at every viewport width from 36 m to 60 m, including the held-out match; short-occlusion median position error reaches 3.3–8.9 m, while 50–57 percent of hidden observations fall outside the 9.6 s regime of the closest learned prior work. On two real World Cup broadcast windows the same imputation style shifts Space-Creation Index by 15.6 and 17.2

Load-bearing premise

That a fixed-width horizontal pan window on three matches from one open provider, with online role offsets, is close enough to real broadcast occlusion statistics and within-half role stability for the reported error reductions to transfer to production video pipelines.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper quantifies how ignoring off-screen players in broadcast-based game-state reconstruction distorts team-level spatial metrics (pitch control, control share), using a simulated ball-following viewport over open Metrica full-pitch tracking (three matches, one held out). It then evaluates a ladder of causal, training-free online imputers that use only the current match. Role-anchored centroid voting (B4; Eq. 1) is reported as best overall on decision-relevant metrics: it roughly halves hidden-zone pitch-control MAE (to 12.2–13.8 pp from B0’s 25.1–26.9) and cuts control-share error to 28–48% of the ignore policy across viewport widths 36–60 m on all three matches. Occlusion-time stratification shows usable short-gap position error (3.3–8.9 m median for ≤9.6 s) while documenting that 50–57% of hidden samples fall outside the Graph Imputer’s 9.6 s protocol. An end-to-end GSR case study on two World Cup windows shows that a B4-inspired layer moves a Space-Creation Index by 15.6 and 17.2 points and flips the operational verdict class in one window.

Significance. The problem is real and under-measured: GSR pipelines and GS-HOTA score only visible players, yet pitch control and related metrics need the full roster. Establishing an open, reproducible distortion benchmark on public tracking data, with decision-relevant metrics and block-bootstrap intervals, is a genuine community contribution. The training-free online floor (especially B4), the explicit long-occlusion stratification against Graph Imputer’s protocol, the B3E/B3V ablation, held-out game 3, and public code/data are strengths that make the empirical claims checkable and useful as a baseline for learned methods. The real-broadcast SCI sensitivity result, though limited, correctly frames camera-framing dependence as a practical risk for possession-quality verdicts. If the reported reductions hold under the stated protocol, the paper sets a clear, low-compute floor that subsequent work should beat.

major comments (2)
  1. [Section 6] Section 6 (SCI case study): The claim that imputation “flips the verdict class” in window 2 rests on operational bins SCI ≥ +12 / +4 that the manuscript states were set in earlier internal work and are “not externally validated.” Window 2 moves −10.9 → +4.7, which sits on the dead/weak boundary. Without a threshold-sensitivity sweep (or an external calibration of the bins), the class-flip statement is load-bearing for the application contribution but only weakly supported. Either add a short sensitivity table over nearby thresholds or rephrase the contribution strictly as a continuous SCI shift of 15.6–17.2 points without class language.
  2. [§3, §5.1, §7] Section 3 viewport simulator and §7 Limitations: The 50–57% long-occlusion share and the transfer narrative for production GSR depend on a fixed-width horizontal ball-EMA pan with no zoom or tilt. The paper correctly discloses this, but the share is presented as a substantive finding relative to Graph Imputer’s 9.6 s protocol (Table 3 / §5.1). A minimal quantitative check—e.g., how the long-occlusion fraction and B0/B4 gaps change under a simple zoom schedule that varies W over a possession, or under a second provider if available—would make that comparative claim more robust. Absent that, the long-occlusion contribution should be scoped more tightly as “under this pan-only simulator.”
minor comments (6)
  1. [Table 1] Table 1 / game 3 share error: B2 and B3V edge B4 by 0.3 pp with a CI that includes zero. The text is honest; consider bolding co-best methods or adding a short note in the table caption so readers do not over-read “best overall” as uniformly best on every cell.
  2. [Section 4, Eq. (1)] Eq. (1): Clarify the evaluation-step rate of the EMA weight 0.1 (stated as per 5 fps step) and the seeding rule until three voters are available; a one-line pseudocode block would reduce implementation ambiguity for reusers of the open code.
  3. [Section 3, Table 1] Pitch-control setup (Section 3): Zero-velocity control is a deliberate isolation choice and is discussed in §7; flag it once more in the Table 1 caption so metric consumers do not treat the MAE magnitudes as velocity-aware pitch-control error.
  4. [Figures 1–2] Figure 1 / Figure 2: Ensure axis units and the viewport band are labeled in the figure itself (not only the caption), and that ghost markers in Fig. 2 are distinguishable in grayscale print.
  5. [Section 2] Related Work: FOOTPASS and SoccerNet-GSR are well placed; a brief pointer to whether any public GSR challenge submission already ships an imputation layer would help readers locate the practical gap.
  6. [Abstract / Introduction] Minor wording: “attenuating the viewport-induced subset bias” appears in both abstract and intro; one instance could be shortened. Also fix any residual spacing around em-dashes and “W = 44m” vs “W = 44 m” for consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical floor on open full-pitch tracking under a simulated viewport, scored against external ground truth rather than quantities defined from the method.

full rationale

The paper's central claims are measured reductions (hidden-zone pitch-control MAE, control-share error, position error) of training-free online imputers versus a visible-only baseline, obtained by applying a ball-EMA pan viewport to open Metrica full-pitch tracking and comparing imputed maps/positions to the true full-team maps. B4's voted-centroid recursion (Eq. 1) is an online self-consistent estimator of role offsets; it is not a definition that forces the reported MAE or share-error numbers, which are computed against held-out ground-truth coordinates. Hyperparameters are fixed round values developed on games 1–2 and frozen before game 3; SCI class thresholds were set in earlier internal work before the two World Cup windows. There is no fitted-input-called-prediction, no load-bearing self-citation uniqueness theorem, no ansatz smuggled via overlapping-author citation, and no renaming of a known result as a derivation. The evaluation is externally falsifiable on the released code and open tracking data. Score 0 is therefore the correct outcome.

Assumptions & free parameters 8 free parameters · 6 assumptions · 3 invented entities

The central empirical claims rest on a small set of modeling choices for the viewport, pitch-control map, and online role offsets, plus a handful of round hyperparameters fixed on development matches. No new physical entities are postulated; the main invented construct is the B4 voting estimator and the operational SCI bins used in the case study. Background pitch-control and GSR machinery are taken from prior literature.

free parameters (8)
  • last-seen decay time constant τ = 8 s
    B1 uses w = exp(−Δt/τ) with τ = 8 s; round value fixed during development on games 1–2, not grid-searched.
  • EMA offset update weight = 0.1 per 5 fps step
    B4/B3 offset EMA weight 0.1 per 5 fps evaluation step; fixed once on development matches.
  • three-voter threshold for B4 = 3 voters
    Voting path requires ≥3 offset-bearing players; otherwise fallback to B2/B1. Round operational choice.
  • velocity blend time constant = 1.5 s
    B3 blends constant-velocity extrapolation with weight exp(−Δt/1.5 s); fixed round value.
  • viewport pan EMA α = 0.06 per 25 fps frame
    Simulated camera follows exponentially smoothed ball with α = 0.06 per frame at 25 fps to mimic broadcast lag.
  • default viewport width W = 44 m (primary)
    Primary tables use W = 44 m chosen so average visible players (14.6–15.0) sit in the real-broadcast range; sensitivity also reported.
  • SCI operational class thresholds = +12 / +4
    space creation ≥ +12, weak progression +4 to +12, dead < +4; set in earlier internal work before these windows, not externally validated.
  • ghost lifetime and retirement rules (application) = 20 s / 2 s / 3 m
    Real-pipeline ghosts fade over 20 s lifetime; retire if new track within 2 s birth and 3 m of prediction; tops team to ≤11 players.
assumptions (6)
  • domain assumption Pitch control on a 3 m grid with arrival-time model and sigmoid contest, computed with zero velocities in all conditions, is an adequate decision-relevant spatial metric for scoring imputation.
    Section 3 Metrics; isolates position imputation from velocity estimation; magnitudes apply only to this position-only variant.
  • domain assumption A fixed-width full-height horizontal window panning with ball-EMA lag is a sufficient simulator of broadcast main-camera occlusion for benchmarking.
    Section 3 Viewport simulator; Limitations note missing zoom/tilt may change occlusion statistics.
  • domain assumption Player role offsets relative to a team centroid are stable enough within a half that storing and reapplying them recovers useful hidden positions.
    Core of B2–B4 (Section 4); B5’s poor fixed-template result is used as evidence of within-half drift, supporting dynamic offsets.
  • domain assumption Causal online methods using only the current match’s observations are the right comparison class for broadcast GSR pipelines that cannot wait for future frames or offline multi-match training.
    Framing in Introduction and Related Work vs bidirectional Graph Imputer.
  • ad hoc to paper Never-observed players remain absent from all partial-observation control maps (cold-start penalty shared across policies).
    Section 3 Metrics definition of the scored population and map metrics.
  • standard math Standard multi-object tracking, camera calibration, and pitch-control constructions from cited GSR/sports analytics literature hold as the pipeline backbone.
    SoccerNet-GSR, BoT-SORT, PnLCalib, Spearman pitch control cited as infrastructure (Sections 2, 6).
invented entities (3)
  • Role-anchored centroid voting (B4)
    purpose: Attenuate viewport-induced subset bias by having each visible player vote for the full-team centroid as position minus running role offset, then impute hidden players at voted centroid plus offset.
    Defined by Eq. (1) in Section 4; primary proposed method. Independent evidence is the open-benchmark error reductions vs B0–B3/B5, not an external physical prediction.
  • Space-Creation Index (SCI) operational bins
    purpose: Map possession windows to space-creation / weak-progression / dead-possession verdicts from changes in pitch-control shares.
    Used in Section 6 case study; thresholds from earlier internal work. Not independently validated outside the authors’ operational use.
  • Open broadcast-viewport distortion benchmark on Metrica independent evidence
    purpose: Quantify hidden-zone control MAE, share error, and position error under simulated camera occlusion with held-out match discipline.
    Section 3; the measurement instrument itself is a contribution. Reusable via public code/data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Training-Free Off-Screen Player Imputation for Broadcast-Based Spatial Football Analytics." pith.science (2026). https://pith.science/paper/ACDKOA7U

@misc{pith2026260711548,
  author       = {Pith},
  title        = {Pith review of: Training-Free Off-Screen Player Imputation for Broadcast-Based Spatial Football Analytics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ACDKOA7U}},
  note         = {Machine review of arXiv:2607.11548}
}
read the original 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 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 control-share error to 28-48% of the ignore policy at every viewport width from 36 m to 60 m in all three matches. For occlusions <=9.6 s -- the regime of the closest learned prior work -- it reaches binwise median position errors of 3.3-8.9 m; but 50-57% of hidden-player observations lie beyond that regime. Integrated end-to-end into a broadcast-video GSR pipeline, imputation moves 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 in one.

Figures

Figures reproduced from arXiv: 2607.11548 by the authors.

Figure 1
Figure 1. Left: imputation ladder at W = 44 m — three-game mean of hidden-zone control MAE and control-share error, all eight methods. Right: viewport-width sensitivity of the share error — B4 (solid) vs. B0 (dashed) across W = 36–60 m, all three games. B4’s absolute share-error gain is largest in the narrowest (hardest) viewports [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Benchmark visualisation (Metrica, W = 44 m). Left: ground-truth pitch control from all 22 players. Middle: control from visible players only (B0) — the hidden zone (outside the vertical viewport band) is systematically misassigned. Right: control with B4 imputation (ghost markers) — the hidden zone is largely restored without any training data. occlusion gap ≤2 s 2–9.6 s >9.6 s B4 median position error 3.3–3.7 m 7.2… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 1 canonical work pages

  1. [1]

    BoT-SORT: Robust associations multi- pedestrian tracking.arXiv preprint arXiv:2206.14651, 2022

    Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. BoT-SORT: Robust associations multi- pedestrian tracking.arXiv preprint arXiv:2206.14651, 2022

    arXiv 2022

  2. [2]

    TranSPORTmer: A holistic approach to trajectory understanding in multi-agent sports

    Guillem Capellera, Luis Ferraz, Antonio Rubio, Antonio Agudo, and Francesc Moreno-Noguer. TranSPORTmer: A holistic approach to trajectory understanding in multi-agent sports. In Asian Conference on Computer Vision (ACCV), 2024. arXiv:2410.17785

    arXiv 2024

  3. [3]

    Trajectory imputation in multi-agent sports with derivative-accumulating self-ensemble

    Han-Jun Choi, Hyunsung Kim, Minho Lee, Minchul Jeong, Chang-Jo Kim, Jinsung Yoon, and Sang-Ki Ko. Trajectory imputation in multi-agent sports with derivative-accumulating self-ensemble. InMachine Learning and Knowledge Discovery in Databases (ECML PKDD),

  4. [4]

    Beal, Tim Matthews, Joseph Early, Timothy J

    Gregory Everett, Ryan J. Beal, Tim Matthews, Joseph Early, Timothy J. Norman, and Sarva- pali D. Ramchurn. Inferring player location in sports matches: Multi-agent spatial imputation from limited observations. InProceedings of the 22nd International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 1643–1651, 2023. arXiv:2302.06569

    arXiv 2023

  5. [5]

    Falaleev and Ruilong Chen

    Nikolay S. Falaleev and Ruilong Chen. Enhancing soccer camera calibration through keypoint exploitation. In7th ACM International Workshop on Multimedia Content Analysis in Sports (MMSports), 2024

    2024

  6. [6]

    PnLCalib: Sports field registration via points and lines optimization.arXiv preprint arXiv:2404.08401, 2024

    Marc Gutiérrez-Pérez and Antonio Agudo. PnLCalib: Sports field registration via points and lines optimization.arXiv preprint arXiv:2404.08401, 2024

    arXiv 2024

  7. [7]

    Le, Peter Carr, Yisong Yue, and Patrick Lucey

    Hoang M. Le, Peter Carr, Yisong Yue, and Patrick Lucey. Data-driven ghosting using deep imitation learning. InMIT Sloan Sports Analytics Conference, 2017. 9

    2017

  8. [8]

    Metrica sports sample tracking data.https://github.com/metrica-sports/ sample-data, 2020

    Metrica Sports. Metrica sports sample tracking data.https://github.com/metrica-sports/ sample-data, 2020

    2020

Show all 15 references
  1. [9]

    FOOTPASS: A multi-modal multi-agent tactical context dataset for play-by-play action spotting in soccer broadcast videos.arXiv preprint arXiv:2511.16183, 2025

    Jeremie Ochin, Raphael Chekroun, Bogdan Stanciulescu, and Sotiris Manitsaris. FOOTPASS: A multi-modal multi-agent tactical context dataset for play-by-play action spotting in soccer broadcast videos.arXiv preprint arXiv:2511.16183, 2025

    2025

  2. [10]

    Connor, Paul Muller, Natalie Mackraz, Kris Cao, Pol Moreno, Pablo Sprechmann, Demis Hassabis, Ian Graham, William Spearman, Nicolas Heess, and Karl Tuyls

    Shayegan Omidshafiei, Daniel Hennes, Marta Garnelo, Zhe Wang, Adria Recasens, Eugene Tarassov, Yi Yang, Romuald Elie, Jerome T. Connor, Paul Muller, Natalie Mackraz, Kris Cao, Pol Moreno, Pablo Sprechmann, Demis Hassabis, Ian Graham, William Spearman, Nicolas Heess, and Karl T...

    2022 doi

  3. [11]

    Penn, Christl A

    Matthew J. Penn, Christl A. Donnelly, and Samir Bhatt. Continuous football player tracking from discrete broadcast data.Royal Society Open Science, 12(10):251175, 2025. doi: 10.1098/ rsos.251175

    2025

  4. [12]

    SoccerNet game state reconstruction: End-to-end athlete tracking and identification on a minimap

    Vladimir Somers, Victor Joos, Anthony Cioppa, Silvio Giancola, Seyed Abolfazl Ghasemzadeh, Floriane Magera, Baptiste Standaert, Amir Mohammad Mansourian, Xin Zhou, Shohreh Kasaei, Bernard Ghanem, Alexandre Alahi, Marc Van Droogenbroeck, and Christophe De Vleeschouwer. SoccerNe...

    2024 arXiv

  5. [13]

    Beyond expected goals

    William Spearman. Beyond expected goals. InMIT Sloan Sports Analytics Conference, 2018

    2018

  6. [14]

    Physics-based modeling of pass probabilities in soccer

    William Spearman, Austin Basye, Greg Dick, Ryan Hotovy, and Paul Pop. Physics-based modeling of pass probabilities in soccer. InMIT Sloan Sports Analytics Conference, 2017

    2017

  7. [15]

    NFL ghosts: A framework for evaluating defender positioning with conditional density estimation.arXiv preprint arXiv:2406.17220, 2024

    Ronald Yurko, Quang Nguyen, and Konstantinos Pelechrinis. NFL ghosts: A framework for evaluating defender positioning with conditional density estimation.arXiv preprint arXiv:2406.17220, 2024. 10

    2024 arXiv

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.