Graph-to-Grid (G2G): Continuous-Coordinate Feature Painting for Soccer Pass Surfaces
cs.LG · cs.CV
Quick verdict
Careful pre-specified study: painting player features bilinearly on the pitch grid improves pass-surface selection by ~0.25 nats; limits honestly reported.
Summary
The paper evaluates an entity-to-grid interface for dense soccer pass surfaces: rather than binning each player into a raster cell, per-player features — and, in the full model, embeddings from a message-passing graph encoder — are scattered bilinearly onto the pitch grid at measured continuous coordinates, and a SoccerMap-style convolutional core is trained end to end through the sparse endpoint loss. On 53,628 passes from all 64 matches of the 2022 World Cup (48/8/8 match split), the full front end improves selection NLL over the identical raster-only core by 0.251 nats in eight-fold cross-validation (ahead in 64 of 64 matches), 0.292 nats under per-arm learning-rate tuning over five seeds, and 0.239 nats after retraining on seven public Bundesliga matches from another provider. Thirteen pre-specified studies attribute roughly three quarters of the gain to painting the nine raw features with no encoder, and about half of the painted package to bilinear placement itself; offset channels, a four-times-finer raster, an attention painter and a raster-free decoder do not close the gap, while the original SoccerMap and a canonical U-Net core both benefit. Limitations are candid: the shipped comparison is retrospective and the G2G arm opened the test set twice; frozen transfer across the provider boundary loses the likelihood advantage; the model is more sensitive to injected tracking noise; the value head is weak; and no calibration comparison survives correction.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
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
- Bilinear painting of per-player features onto the pitch grid improves selection NLL over the identical raster-only core by 0.251 nats in eight-fold cross-validation (ahead in 64/64 matches), 0.292 under five-seed per-arm tuning, and 0.239 after retraining on seven external Bundesliga matches.
- Roughly three quarters of the gain comes from painting the nine raw features with no learned encoder (Draw − C = −0.186 nats in CV); the learned encoder plus message passing adds a further −0.065 nats, together resolved over 64 clusters.
- About half of the painted package's selection gain is attributable to bilinear placement at continuous coordinates: nearest-cell painting costs +0.084 to +0.093 nats across both painted arms.
- The effect is not a parameter-count artifact and replicates on the original SoccerMap core (original_splat −0.295 nats) and a canonical U-Net (U_raw − U_raster −0.242); offset channels (−0.044), a half-metre raster (−0.047, unresolved), an attention painter (+0.235) and a raster-free decoder (+0.219) do not close the gap.
- Frozen zero-shot transfer across the provider boundary loses the selection-likelihood advantage (G2G − C = +0.079, behind in 7/7 matches) while localisation survives, and injected 1 m tracking noise compresses the lead over the original control from +0.52 to +0.090 ± 0.025 nats.
Strengths
Genuine pre-registration and audit discipline
13 studies with signed commitments, hash-locked predictions and per-study metric families; the invalid-at-8-clusters bootstrap p-values are explicitly demoted in favour of the exact 256-pattern max-T test, attainable floors (0.0078/0.0156) are printed, and the double test opening is disclosed with the discarded cycle printed (Sec 5.3, 5.5, 7)
Attribution is decomposed, not asserted
The wide control matches G2G within 0.25% of parameters; nearest-vs-bilinear (+0.084/+0.093 nats), sub-cell offset channels (−0.044), a half-metre raster (−0.047, unresolved) and a parameter-matched core (null) price each alternative against the same core (Studies 2, 3, 8, 9)
The effect is not tied to one core or one corpus
Honest nulls and negative results
Scope discipline on what the surfaces mean
Weaknesses
The headline comparison is retrospective and the holdout was opened twice
Eight-cluster inference limits resolution of key attribution contrasts
Cross-validation is confirmatory only in a qualified sense
Deployment validity is thin despite honest reporting
Value surfaces are near-null across all arms
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: what does writing per-player features onto the pitch grid at continuous (bilinear) coordinates buy, relative to hard cell binning, for dense pass-surface prediction? The G2G family nests four arms around a shared convolutional core: raster-only SoccerMap-v2 (C), the same core plus painted raw features (Draw), plus a per-node MLP encoder (D), plus relational message passing (G2G). Four heads (completion PP, selection PS over 7,072 cells, two value branches PE) are supervised only at the one observed endpoint per pass. The central claim (Sec 1.1, Table 1): painting improves selection NLL over the identical raster-only core by −0.251 nats [−0.263, −0.239] in eight-fold cross-validation over all 64 World Cup matches (ahead in 64/64), −0.292 under per-arm learning-rate tuning over five seeds, and −0.239 after retraining on seven public Bundesliga/2. Bundesliga matches from another provider. Claims are explicitly scoped to observed-endpoint prediction, not counterfactual validity (Sec 4.5, 7) — a commendable discipline.
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
What is Graph-to-Grid (G2G) in soccer analytics?
G2G is a family of models that predicts dense pass surfaces over a soccer pitch. Its key ingredient is feature painting: each player's features (position, velocity, team, role) are written onto the pitch grid at the player's exact continuous coordinates via bilinear splatting, instead of being counted into one-meter raster cells. A convolutional network then reads the grid and outputs completion, selection and value surfaces, trained end to end from the single observed pass endpoint.
How much does continuous-coordinate painting improve pass prediction?
On 53,628 passes from all 64 matches of the 2022 World Cup, it improves the selection negative log-likelihood by about 0.25 nats over an identical raster-only network - on a geometric average the model places roughly 29% more probability on the actual pass destination. It wins in every match under eight-fold cross-validation, in every match under per-arm tuning over five seeds, and in every match after retraining on seven public German league matches.
Does the G2G model work on data from a different tracking provider?
Only after retraining. Retrained from scratch on seven public German league (IDSSE) matches, the ordering and effect size return (-0.239 nats, 7 of 7 matches). Applied frozen, zero-shot, across the provider boundary, the full G2G loses its selection-likelihood advantage (behind the raster-only core in all seven matches), though its localisation accuracy survives - consistent with the learned encoder being fitted to the source provider's data distribution.
Is most of the gain from the graph neural network or from the painting itself?
From the painting. Painting just the nine raw player features with no learned encoder carries about three quarters of the total gain, and bilinear placement at continuous coordinates accounts for roughly half of that. The learned encoder plus message passing adds the remaining quarter (about 0.065 nats), which is small relative to its seed-to-seed variation on the eight-match holdout.
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.
- Extraction marked several sections as omitted (parts of Secs 3, 5.5, 6.3-6.6, Sec 8, and Statements/Declarations); the data- and code-availability statements were therefore not visible and codeAvailable=false reflects absence of evidence, not evidence of absence.
- Figures 1-5 are referenced but not analyzable from text; several protocol details live only in the online supplement (Tables S1-S41), which was not available.
- The abstract in the metadata is truncated mid-sentence; all claims above were verified against the body text.
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.