How Much Velocity Does Off-Ball Space Value Need? A Broadcast-Viewport Benchmark
cs.CV · cs.LG
Quick verdict
Measures where velocity matters in broadcast off-ball analysis; rigorous paired benchmark, but built on three matches and the author's own reference surface.
Summary
This benchmark paper asks at which layer of broadcast-viewport off-ball football analysis player velocity actually changes the answer. Building on the author's off-screen imputation protocol (three Metrica matches, a 44 m panning viewport, Spearman-style pitch control, block-bootstrap CIs), it scores four velocity regimes — none, viewport-legal observed, true-for-visible, true-for-all — against a velocity-aware reference at three layers. Velocity is nearly useless for imputation (about −0.2 pp against a 12–14 pp control MAE), first-order for the surface (−1.5 to −1.8 pp, 11–15% of MAE), and an order of magnitude smaller for team verdicts (−0.12 to −0.19 pp). The useful velocity is the visible channel: perfect occluded-player velocity adds only 2–6% of the visible gain, and no last-seen decay policy exceeds that. Omitting velocity blurs the surface without time-averaged bias (≤0.4 pp per cell), whereas imputation error is a structured 5–9 pp bias against defending teams' deep zones. An eleven-setting noise ladder is ordered by per-axis velocity noise σv with break-even ≈1 m/s; eleven SoccerNet-GSR clips from one match through the author's pipeline measure σv=1.65 m/s yet recover 24–36% of the benefit because 43% of the error is frame-common and the residual heavy-tailed. Main limitations: three matches, one control-model family, a one-match real-broadcast axis, and 'cost' defined relative to the author's own reference surface.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Velocity matters at the control-surface layer (est−zero −1.48 to −1.79 pp of control MAE) but an order of magnitude less at team verdicts (−0.12 to −0.19 pp) and almost none at imputation (−0.19 to −0.25 pp).
- The useful velocity is the visible channel: perfect occluded-player velocity adds only 2–6% of the visible gain at the imputed positions, and no last-seen decay policy tested exceeds that.
- Omitting velocity blurs the surface without time-averaged bias (per-cell bias RMS ≤0.14 pp, max ≤0.4 pp), whereas imputation error is a structured 5–9 pp under-crediting of the defending team's deep zone.
- At a fixed 0.6 s window, observed velocity pays off when per-axis noise is below ≈1 m/s in i.i.d. terms, but the real pipeline at 1.65 m/s still recovered 24–36% because 43% of its error is frame-common and the residual heavy-tailed.
- Velocity's share of removable error grows with viewport width (7% at 36 m, 21% at 60 m), so pipelines should fix imputation on tight shots and velocity on wide ones.
Strengths
Decision-relevant layer decomposition
Section 4.1 and Table 1 quantify velocity's worth at imputation (−0.19 to −0.25 pp), surface (−1.48 to −1.79 pp) and verdict (−0.12 to −0.19 pp) with paired block-bootstrap CIs — a clean ordering a deployed pipeline can act on.
Matched-pair noise design isolates one axis
Section 4.4's σv-matched i.i.d./AR(1) pairs (e.g., 0.52 m vs 1.0 m, both σv=1.23, recover −0.36/−0.35) show per-axis velocity noise orders the ladder to first order, with paired differences (−0.03/−0.01 below break-even, +0.19/+0.45 above) quantifying the residual σ² dependence.
Real-broadcast error is decomposed, not just averaged
Reproducibility and honest scoping
Artifacts separated from findings
Weaknesses
Reference circularity
Thin empirical base
Real axis conditions on an easy subset
Headline ratios lack uncertainty
Partial replication of sweeps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks a narrow, well-posed question: in broadcast-viewport off-ball football analytics, at which processing layer does player velocity change the answer, and by how much? It inherits the author's off-screen imputation protocol (three Metrica matches, virtual panning camera, 44 m viewport) and a Spearman-style arrival-time control model, then scores four velocity regimes — zero, viewport-legal observed (est), true-for-visible (vis), true-for-all (oracle) — against a velocity-aware reference at three layers: imputation, control surface, and team verdicts. Central claims: (1) velocity is nearly useless for imputation, first-order for the surface, and ~10× smaller for verdicts (Table 1: surface est−zero −1.48/−1.68/−1.79 pp vs verdict −0.12 to −0.19 pp); (2) the useful velocity is the visible channel, perfect occluded velocity adding only 2–6% of the visible gain (Table 2: oracle−vis −0.04 to −0.09 pp); (3) velocity omission blurs (per-frame |e| 2.25–2.63 pp) but barely biases (cell-bias RMS 0.11–0.14 pp), while imputation is a structured 5–9 pp bias against the defending deep zone (§4.3, Fig. 1); (4) the noise ladder collapses onto one σv axis with break-even ≈1 m/s (§4.4); (5) real SoccerNet-GSR pipeline error at σv=1.65 m/s still recovers 24–36% because 43% of the variance is frame-common with a heavy-tailed residual (§4.5, Fig. 3b); (6) velocity's share of removable error rises with viewport width, 7% at 36 m to 21% at 60 m (Table 3).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Do you need player velocity for broadcast-based pitch control in football?
According to this benchmark, velocity matters mainly at the per-cell control-surface layer: omitting it costs 1.5–1.8 pp of control MAE (11–15% of the total). It matters ten times less for team-level share verdicts (−0.12 to −0.19 pp) and almost nothing for off-screen player imputation (about −0.2 pp). The useful velocity comes from visible players; occluded players' velocity adds only 2–6% even if it were perfect.
How accurate must broadcast tracking velocity be to be worth using?
The paper estimates a break-even at roughly 1 m/s of per-axis velocity noise for a 0.6 s finite-difference window in a Spearman-style control model; above that, using observed velocity is worse than using none. Real broadcast error measured at 1.65 m/s still recovered 24–36% of the benefit because much of the error is frame-common (43%) with a heavy-tailed residual rather than i.i.d. Gaussian.
What is the biggest error source in broadcast off-ball spatial analytics?
Not velocity. The paper finds that imputing off-screen players introduces a structured 5–9 pp under-crediting of the defending team's deep zone, while omitting velocity produces only transient blur (per-cell time-averaged bias of at most 0.4 pp). Threat-weighted verdicts are therefore exposed to imputation error, not velocity error.
What data was used in this study?
Three public Metrica Sports sample matches (first halves) for the simulated broadcast-viewport benchmark, plus eleven 30-second SoccerNet-GSR test clips from a single match processed through the author's own detection-tracking-calibration pipeline. Code, logs and the pipeline's tracks for those clips are released and archived on Zenodo (10.5281/zenodo.22310851).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–3 appear in the extraction only as embedded axis-label and caption text; the visual map structure could not be inspected, but every number cited in this review comes from body text, tables or captions present in the extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.