Pith. sign in

REVIEW 2 major objections 5 minor 79 references

On-ball soccer events can be recovered from player trajectories alone by inferring a possession path with a masked conditional random field.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 23:54 UTC pith:HXIN624Z

load-bearing objection PathCRF is a solid system paper for inferring soccer possession paths from player-only tracking data, but its hard transition rules miss common tackles and interceptions, and the single-match evaluation leans on the constraint set. the 2 major comments →

arxiv 2602.12080 v2 pith:HXIN624Z submitted 2026-02-12 cs.LG

PathCRF: Ball-Free Soccer Event Detection via Possession Path Inference from Player Trajectories

Hyunsung Kim , Kunhee Lee , Sangwoo Seo , Sang-Ki Ko , Jinsung Yoon , Chanyoung Park This is my paper
classification cs.LG
keywords soccerevent detectionplayer tracking datapossession path inferenceconditional random fieldsequence labelingmulti-agent trajectory modelingsports analytics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that on-ball soccer events can be reconstructed from player tracking data alone, without ball tracking or manual annotation. It turns each moment of play into exactly one directed edge in a fully connected graph over the 22 players and four boundary-outside nodes: a self-loop means a player controls the ball, and a sender–receiver edge means a kick in flight. A neural encoder builds edge embeddings, and a dynamic masked conditional random field scores entire edge sequences while forbidding physically impossible transitions; events are then read off as the time steps where the selected edge changes. The claimed results are 69.64% accuracy in picking the correct edge among 676 candidates and a 75.69% F1 score for detecting control, kick, and out-of-play events, with downstream possession and pass analyses closely tracking the ground truth.

Core claim

The paper's central claim is that possession can be treated as a path through a state space of edges, and that this path is inferable from player movements alone. At each time step the model selects one edge; consecutive edges must obey a hand-coded allowed-transition set that only permits a state to continue, a player to kick from a self-loop, a receiver to take control or play one touch, or the ball to go out of play. The dynamic masked CRF computes emission and transition scores from edge embeddings, masks everything outside the allowed set, and decodes the highest-scoring globally consistent sequence. This yields a possession path with no illegal transitions, and the paper reports that c

What carries the argument

The possession path: a sequence over the 676 directed edges of a fully connected dynamic graph whose nodes are the 22 players plus four outside nodes; a self-loop denotes ball control and a directed edge denotes a kick in flight. The neural backbone encodes each snapshot into edge embeddings, and the dynamic masked CRF computes an emission score for every candidate edge and a transition score for every candidate pair of consecutive edges. Transitions outside the allowed set are masked with a large negative score, so the decoding step—exact dynamic programming over the sparse allowed graph—returns only physically consistent paths. Work done by this machinery: it converts the event-detection p

Load-bearing premise

The load-bearing premise is that every possession change in real play is covered by the hand-coded transition set—continuation, kick from control, reception or one-touch kick, and out-of-play—so any tackle, interception, or loose-ball recovery that transfers the ball directly between players is forced into an artificial kick edge, and stoppages and restarts are excluded from the model's windows; if that premise is not physically true, the 0% violation rate and the event score

What would settle it

Take a held-out match segment annotated with a clear tackle, interception, or loose-ball recovery in which possession moves directly from player A to player B without any kick edge. If the allowed-transition set is complete, the true event must be representable and PathCRF should recover it; if the model cannot express the event or its accuracy on such actions is far below the reported average, the constraint set is not covering the game's actual possession physics. A simpler quantitative check is to score the inferred sequences against a richer event taxonomy that includes direct possession c

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If this holds, event data can be generated for any match that already has player tracking, removing ball-tracking hardware and manual annotation as prerequisites for event-based analysis.
  • Standard downstream analytics can run on the detected events without human labeling: the paper reports close agreement on event heatmaps, a team possession share of 62.64% versus 62.13% ground truth, and pass networks with small errors in node degrees and edge weights.
  • A semi-automated annotation workflow becomes viable, because many predicted events are near misses on player identity while still matching the true event's time and location; recall reaches roughly 80% within one second and four meters.
  • Because the allowed-transition constraint is enforced during training and decoding, the inferred possession path contains no teleportation artifacts, which removes a failure mode that per-frame edge classifiers exhibit.
  • The modeling recipe does not depend on a ball trajectory at all, so it could scale to lower-tier and youth competitions where dense multi-camera ball tracking has been too expensive to deploy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The hand-coded transition set only allows possession to change through a kick from a self-loop and a reception by the named receiver; this presupposes that tackles, interceptions, and loose-ball pickups do not transfer control directly between players. A natural extension is to enrich the allowed set with those possession-winning actions and measure whether the CRF still trains and decodes cleanly
  • The same edge-sequence formulation could transfer to other possession sports such as basketball, handball, or hockey, where the physical transition rules would need to be redefined but the structured-inference skeleton would carry over.
  • Because tracking is downsampled to 5 Hz before inference, the temporal resolution of any detected event is limited to roughly 200 milliseconds; recovering true event times within a frame would require interpolation or a finer sampling rate.
  • Possession paths inferred from player trajectories could serve as weak supervision for tactic and role models in leagues with no event data, effectively widening the pool of matches available for data-driven analysis.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes PathCRF, a framework for detecting on-ball soccer events using only player tracking data. It models each time step as a fully connected dynamic graph over 22 players plus four outside nodes, casts the possession state as an edge in this graph, and uses a neural backbone with a masked CRF to select a globally consistent edge sequence via Viterbi decoding. Events are extracted whenever the selected edge changes. Experiments on the Sportec Open DFL dataset report 69.64% edge accuracy, 75.69% event F1, and 0% violation rate, and the authors show that downstream metrics such as possession shares, heatmaps, and pass networks closely match ground truth. The paper claims that this substantially reduces the need for manual event annotation and ball-tracking infrastructure.

Significance. If the central claim holds, the paper makes a useful practical contribution: it is the first to my knowledge to formulate possession inference as a dynamic-edge CRF over player trajectories and to demonstrate that several downstream analytics can be approximated without ball tracking. The manuscript is generally reproducible: it uses a public dataset, releases source code, compares against multiple baselines, and includes ablations on backbone and loss components. The main value is in the problem formulation and the structured-inference machinery. However, the significance is currently limited by the evaluation on a single test match and by a load-bearing incompleteness in the hand-defined transition constraint set, which I discuss below.

major comments (2)
  1. [§2.3.2 (allowed-transition set A) and §3.1/§2.4 (ground-truth construction)] The set A only permits identity continuation, kick edges from a self-loop, reception edges to the named receiver, and out-of-play exits. It explicitly forbids direct possession changes such as (u,u)→(v,v) (a tackle or loose-ball recovery) and (u,v)→(w,w) with w≠v (an interception of a pass intended for v). These are common events in real soccer. Because the ground-truth edge labels are generated from SPADL events plus RDP-detected ball touches and then mapped to the three simplified categories (control/kick/out-of-play), the ground-truth sequences satisfy A by construction. Consequently, the reported 0% violation rate is definitional rather than evidence that A covers the true possession dynamics, and the 75.69% F1 is measured only on events representable inside A. The model cannot emit a direct gain; it will either insert a spurious kick edge or fail to detect the possession change. The
  2. [§3.1/Table 1 and §3.4/Table 2] The entire evaluation is based on a single test match (1,831 events, one match in Table 1). No error bars, confidence intervals, or per-match breakdowns are given. This is particularly concerning because the gains of the proposed Dynamic Masked CRF over Static Masked CRF are modest at the event level (75.69 vs. 73.83 F1) and edge-level differences are within fractions of a percent (69.64 vs. 69.54). With one test match, one cannot tell whether the dynamic transition scoring and the specific constraint set are reliable advantages or artifacts of a single game. Please report results across multiple held-out matches (or at least a small cross-validation) with variance estimates, or provide a clear justification for why a single-match evaluation is sufficient for the paper's claims.
minor comments (5)
  1. [§3.1] Typo: "we general ground-truth edge labels" should be "we generate ground-truth edge labels".
  2. [Appendix A.1] The text says "a large negative score (e.g., 10^{-4})" but should be "−10^4" to match the main text and to be negative.
  3. [§4.2] Typo: "the our prediction" should be "our prediction".
  4. [§4.4] Wording: "relatively strict criterion of1 sand4 m" should read "of 1 s and 4 m".
  5. [Reference [11]] The proceedings name is misspelled: "Learning Representationsn" should be "Learning Representations".

Circularity Check

1 steps flagged

No material circularity: edge accuracy and F1 are evaluated against external ground truth; only the 0% violation rate is a by-construction consistency metric.

specific steps
  1. self definitional [Appendix A.1 (Eq. 15 and P(e) definition); Sec. 3.3 (violation rate); Table 2]
    "to enforce the hard constraints defined in Section 2.3.2, we iterate only over the set of allowed previous edges P(e)={e′∈E:(e′,e)∈A} ... we measure the violation rate, defined as the percentage of illegal transitions in the predicted edge sequence."

    The forward and Viterbi recursions (Eqs. 15, 19) restrict every considered path to transitions in A. Therefore any decoded sequence satisfies A by construction, making the reported 0.00% violation rate a logical consequence of the masking definition rather than an empirical property of the model. This metric is self-definitional: the thing being measured is exactly the constraint that was hard-coded. It does not affect the external validity of the edge-accuracy or F1 results, which are computed against ground-truth labels, but the paper's use of 'logically consistent' should not be read as independent evidence of physical completeness of A.

full rationale

PathCRF's central claims are supervised predictions: possession edges and events are trained and evaluated against ground-truth labels constructed from external event/tracking data (Sec. 3.1), and the reported 69.64% edge accuracy and 75.69% F1 are empirical comparisons to those external labels. The allowed-transition set A is a hand-specified prior, not a parameter fitted to the evaluation targets, so the model's predictions do not reduce to its inputs by construction. The only by-construction quantity is the violation rate: because the forward/Viterbi recursions restrict to A, 0% violations are guaranteed by design. This is a minor definitional metric, not a load-bearing circular prediction. The possible incompleteness of A for real soccer transitions (tackles, interceptions, loose-ball gains) is a correctness/coverage concern rather than circularity, since the model is not fitting its own evaluation labels. Self-citations such as Ball Radar and ELASTIC are used as architectural or preprocessing components and do not themselves encode the target result.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 2 invented entities

The central claim rests on domain assumptions about possession being representable as a single edge per time step and about the completeness of the hand-coded transition constraints. No new physical entities are proposed beyond the four outside nodes. Several hyperparameters and preprocessing thresholds are not reported.

free parameters (6)
  • Auxiliary loss weights lambda1, lambda2 = not reported
    Eq. 10 weights are chosen by hand; the appendix only reports them set to zero in ablations.
  • Transition mask value = -10^4
    Section 2.3.2 uses a fixed large negative score to mask illegal transitions; the exact value is arbitrary.
  • Window length and stride = 10-second windows, stride 5 frames
    Section 3.1 fixes input context length and overlap; different values may change results.
  • Downsampling rate = 25 to 5 FPS
    Section 3.1 reduces temporal resolution for tractability; event timing is later upsampled.
  • Backbone hyperparameters = not specified
    Hidden sizes, attention heads, and layer counts for PPE-FPE and PE-SABs are not listed in the paper.
  • RDP and Needleman-Wunsch parameters = not specified
    Section 3.1 uses RDP and Needleman-Wunsch to refine ball-touch labels; thresholds and gap penalties are not given.
axioms (5)
  • domain assumption Ball possession state can be inferred from player trajectories alone.
    The entire ball-free framework relies on this premise (Sections 1 and 2.1).
  • ad hoc to paper The allowed transition set A contains all physically legal possession transitions; all others are impossible.
    Section 2.3.2 defines A by hand and calls other transitions 'impossible'; it excludes tackles and loose-ball recoveries where control changes without a kick edge.
  • domain assumption At each time step the ball is possessed by exactly one edge (self-loop or sender-receiver).
    Section 2.1 simplifies the state space, ignoring contested balls, deflections, and simultaneous possession.
  • domain assumption Only in-play segments are considered; outside-node self-loops are absorbing.
    Section 2.4 and footnote 2 restrict training/inference to in-play windows, excluding restarts and out-of-play transitions beyond (u,o)->(o,o).
  • standard math Viterbi and forward algorithms compute exact CRF normalization and decoding.
    Appendix A uses standard dynamic programming; this is unproblematic background.
invented entities (2)
  • Four outside nodes O (left, right, top, bottom) no independent evidence
    purpose: Represent the ball leaving the pitch as an absorbing out-of-play edge.
    Modeling abstraction; there is no falsifiable prediction outside the paper.
  • Possession path (sequence of selected edges) no independent evidence
    purpose: Latent output that defines events whenever the selected edge changes.
    A construct of the framework; no independent evidence outside the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 18803 in / 14448 out tokens · 141237 ms · 2026-08-02T23:54:03.050100+00:00 · methodology

0 comments
read the original abstract

Despite recent advances in AI, event data collection in soccer still relies heavily on labor-intensive manual annotation. Although prior work has explored automatic event detection using player and ball trajectories, ball tracking also remains difficult to scale due to high infrastructural and operational costs. As a result, comprehensive data collection in soccer is largely confined to top-tier competitions, limiting the broader adoption of data-driven analysis in this domain. To address this challenge, this paper proposes PathCRF, a framework for detecting on-ball soccer events using only player tracking data. We model player trajectories as a fully connected dynamic graph and formulate event detection as the problem of selecting exactly one edge corresponding to the current possession state at each time step. To ensure logical consistency of the resulting edge sequence, we employ a Conditional Random Field (CRF) that forbids impossible transitions between consecutive edges, where emission and transition scores are dynamically computed from edge embeddings produced by a socio-temporal backbone architecture. During inference, the most probable edge sequence is obtained via Viterbi decoding, and events such as ball controls or passes are detected whenever the selected edge changes between adjacent time steps. Experiments show that PathCRF produces accurate, logically consistent possession paths, enabling reliable downstream analyses while substantially reducing the need for manual event annotation. The source code is available at https://github.com/hyunsungkim-ds/pathcrf.git.

Figures

Figures reproduced from arXiv: 2602.12080 by Chanyoung Park, Hyunsung Kim, Jinsung Yoon, Kunhee Lee, Sang-Ki Ko, Sangwoo Seo.

Figure 1
Figure 1. Figure 1: Overall architecture of PathCRF. Empirically, in Section 3.5, we observe that this attention-based temporal module achieves performance comparable to its Bi-LSTM [30] counterpart. We attribute this behavior to the nature of our task: soccer possession dynamics are largely driven by short-term continuity, where the most recent context is typically the most informative. As a result, both recurrent models wit… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of the ground-truth possession path [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of kernel density estimation (KDE) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: 5-minute timeline of home team’s possession shares [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparison of pass networks constructed [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Event recall under varying tolerance thresholds. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

79 extracted references · 6 linked inside Pith

  1. [1]

    Alcorn and Anh Nguyen

    Michael A. Alcorn and Anh Nguyen. baller2vec++: A look-ahead multi-entity transformer for modeling coordinated agents.arXiv preprint arXiv:2104.11980, 2021

    Pith/arXiv arXiv 2021

  2. [2]

    Alcorn and Anh Nguyen

    Michael A. Alcorn and Anh Nguyen. baller2vec: A multi-entity transformer for multi-agent spatiotemporal modeling.arXiv preprint arXiv:2102.03291, 2021

    Pith/arXiv arXiv 2021

  3. [3]

    Prediction of the ball location on the 2d plane in football using optical tracking data.Academic Platform Journal of Engineering and Smart Systems, 10(1):1–8, 2022

    Anar Amirli and Hande Alemdar. Prediction of the ball location on the 2d plane in football using optical tracking data.Academic Platform Journal of Engineering and Smart Systems, 10(1):1–8, 2022

    2022

  4. [4]

    Robertson, Joshua W

    Gabriel Anzer, Kilian Arnsmeyer, Pascal Bauer, Joris Bekkers, Ulf Brefeld, Jesse Davis, Nicolas Evans, Matthias Kempe, Samuel J. Robertson, Joshua W. Smith, and Jan Van Haaren. Common data format (CDF): A standardized format for match-data in football (soccer).arXiv preprint arXiv:2505.15820, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    An integrated dataset of spatiotemporal and event data in elite soccer.Scientific Data, 12(195), 2025

    Manuel Bassek, Robert Rein, Hendrik Weber, and Daniel Memmert. An integrated dataset of spatiotemporal and event data in elite soccer.Scientific Data, 12(195), 2025

    2025

  6. [6]

    Putting team formations in association football into context.Journal of Sports Analytics, 9(6):39–59, 2023

    Pascal Bauer, Gabriel Anzer, and Laurie Shaw. Putting team formations in association football into context.Journal of Sports Analytics, 9(6):39–59, 2023

    2023

  7. [7]

    Matthews

    Alina Bialkowski, Patrick Lucey, Peter Carr, Yisong Yue, Sridha Sridharan, and Iain A. Matthews. Large-scale analysis of soccer matches using spatiotemporal tracking data. InIEEE International Conference on Data Mining, 2014

    2014

  8. [8]

    Event detection in football: Improving the reliability of match analysis.PLOS ONE, 19(4), 2024

    Jonas Bischofberger, Arnold Baca, and Erich Schikuta. Event detection in football: Improving the reliability of match analysis.PLOS ONE, 19(4), 2024

    2024

  9. [9]

    Bradley and Jack D

    Paul S. Bradley and Jack D. Ade. Are current physical match performance metrics in elite soccer fit for purpose or is the adoption of an integrated approach needed? International Journal of Sports Physiology and Performance, 13(5):656–664, 2018

    2018

  10. [10]

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

    Guillem Capellera, Luis Ferraz, Antonio Rubio, Antonio Agudo, and Francesc Moreno-Noguer. TranSPORTmer: A holistic approach to trajectory understand- ing in multi-agent sports. InProceedings of the 17th Asian Conference on Computer Vision, 2024

    2024

  11. [11]

    JointDiff: Bridging continuous and discrete in multi-agent trajectory generation

    Guillem Capellera, Luis Ferraz, Antonio Rubio, Alexandre Alahi, and Antonio Agudo. JointDiff: Bridging continuous and discrete in multi-agent trajectory generation. InProceedings of the 14th International Conference on Learning Representationsn, 2026

    2026

  12. [12]

    Unified uncertainty-aware diffusion for multi-agent trajectory modeling

    Guillem Capellera, Antonio Rubio, Luis Ferraz, and Antonio Agudo. Unified uncertainty-aware diffusion for multi-agent trajectory modeling. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    2025

  13. [13]

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

    Han-Jun Choi, Hyunsung Kim, Minho Lee, Minchul Jeong, Changjo Kim, Jin- sung Yoon, and Sang-Ki Ko. Trajectory imputation in multi-agent sports with derivative-accumulating self-ensemble. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases, 2025

    2025

  14. [14]

    Fan R. K. Chung.Spectral Graph Theory, volume 92 ofCBMS Regional Conference Series in Mathematics. American Mathematical Society, 1997

    1997

  15. [15]

    A network-based ap- proach to evaluate the performance of football teams

    Paolo Cintia, Salvatore Rinzivillo, and Luca Pappalardo. A network-based ap- proach to evaluate the performance of football teams. InECML PKDD Workshop on Machine Learning and Data Mining for Sports Analytics, 2015

    2015

  16. [16]

    Wong, and Rui Sousa Mendes

    Filipe Manuel Clemente, Fernando Manuel Lourenço Martins, Dimitris Kala- maras, Daniel P. Wong, and Rui Sousa Mendes. General network analysis of national soccer teams in FIFA World Cup 2014.International Journal of Perfor- mance Analysis in Sport, 15(1):80–96, 2015

    2014

  17. [17]

    Actions speak louder than goals: Valuing player actions in soccer

    Tom Decroos, Lotte Bransen, Jan Van Haaren, and Jesse Davis. Actions speak louder than goals: Valuing player actions in soccer. InProceedings of the 25th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2019

    2019

  18. [18]

    Player vectors: Characterizing soccer players’ playing style from match event streams

    Tom Decroos and Jesse Davis. Player vectors: Characterizing soccer players’ playing style from match event streams. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases, volume 11908 ofLecture Notes in Computer Science, pages 569–584. Springer, 2019

    2019

  19. [19]

    Automatic discovery of tactics in spatio-temporal soccer match data

    Tom Decroos, Jan Van Haaren, and Jesse Davis. Automatic discovery of tactics in spatio-temporal soccer match data. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2018

    2018

  20. [20]

    SoccerMix: Representing soccer actions with mixture models

    Tom Decroos, Maaike Van Roy, and Jesse Davis. SoccerMix: Representing soccer actions with mixture models. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases, 2020

    2020

  21. [21]

    Howie Huang

    Ding Ding and H. Howie Huang. A graph attention based approach for trajectory prediction in multi-agent sports games.arXiv preprint arXiv:2012.10531, 2020

    Pith/arXiv arXiv 2012

  22. [22]

    David Douglas and Thomas Peucker. Algorithms for the reduction of the number of points required to represent a digitized line or its caricature.Cartographica: The International Journal for Geographic Information and Geovisualization, 10(2):112– 122, 1973

    1973

  23. [23]

    Waitzman, and Luís A

    Jordi Duch, Joshua S. Waitzman, and Luís A. Nunes Amaral. Quantifying the performance of individual players in a team activity.PLoS ONE, 5(6):e10937, 2010

    2010

  24. [24]

    Beal, Tim Matthews, Joseph Early, Timothy J

    Gregory Everett, Ryan J. Beal, Tim Matthews, Joseph Early, Timothy J. Norman, and Sarvapali D. Ramchurn. Inferring player location in sports matches: Multi- agent spatial imputation from limited observations. InProceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems, 2023

    2023

  25. [25]

    Semi-supervised generative models for multiagent trajectories

    Dennis Fassmeyer, Pascal Fassmeyer, and Ulf Brefeld. Semi-supervised generative models for multiagent trajectories. InAdvances in Neural Information Processing Systems 35, 2022

    2022

  26. [26]

    Where will they go? Predict- ing fine-grained adversarial multi-agent motion using conditional variational autoencoders

    Panna Felsen, Patrick Lucey, and Sujoy Ganguly. Where will they go? Predict- ing fine-grained adversarial multi-agent motion using conditional variational autoencoders. InEuropean Conference on Computer Vision, 2018

    2018

  27. [27]

    Semi-automated offside technology, 2022

    FIFA. Semi-automated offside technology, 2022. https://inside.fifa.com/ innovation/world-cup-2022/semi-automated-offside-technology, Accessed: Jan 21, 2026

    2022

  28. [28]

    Grammar- constrained decoding for structured NLP tasks without finetuning

    Saibo Geng, Martin Josifoski, Maxime Peyrard, and Robert West. Grammar- constrained decoding for structured NLP tasks without finetuning. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023

    2023

  29. [29]

    A survey on recent advances in sequence labeling from deep learning models.arXiv preprint arXiv:2011.06727, 2020

    Zhiyong He, Zanbo Wang, Wei Wei, Shanshan Feng, Xianling Mao, and Sheng Jiang. A survey on recent advances in sequence labeling from deep learning models.arXiv preprint arXiv:2011.06727, 2020

    Pith/arXiv arXiv 2011

  30. [30]

    Long short-term memory.Neural Computation, 9(8):1735–1780, 1997

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory.Neural Computation, 9(8):1735–1780, 1997

    1997

  31. [31]

    Lexically constrained decoding for sequence gener- ation using grid beam search

    Chris Hokamp and Qun Liu. Lexically constrained decoding for sequence gener- ation using grid beam search. InProceedings of the 55th Annual Meeting of the Association for Computational Linguistics, 2017

    2017

  32. [32]

    S. K. Hong and Jae-Gil Lee. DTranNER: Biomedical named entity recognition with deep learning-based label-label transition model.BMC Bioinformatics, 21(1):53, 2020

    2020

  33. [33]

    Bidirectional LSTM-CRF models for sequence tagging.arXiv preprint arXiv:1508.01991, 2015

    Zhiheng Huang, Wei Xu, and Kai Yu. Bidirectional LSTM-CRF models for sequence tagging.arXiv preprint arXiv:1508.01991, 2015

    Pith/arXiv arXiv 2015

  34. [34]

    Event2Tracking: Reconstructing multi- agent soccer trajectories using long-term multimodal context

    Harry Hughes, Michael Horton, Xinyu Wei, Harshala Gammulle, Clinton Fookes, Sridha Sridharan, and Patrick Lucey. Event2Tracking: Reconstructing multi- agent soccer trajectories using long-term multimodal context. InProceedings of the 39th AAAI Conference on Artificial Intelligence, 2025

    2025

  35. [35]

    Imputing multi-agent trajectories from event and snapshot data in soccer

    Geonhee Jo, Miru Hong, Han-Jun Choi, Minho Lee, Pascal Bauer, and Sang-Ki Ko. Imputing multi-agent trajectories from event and snapshot data in soccer. InProceedings of the 34th ACM International Conference on Information and Knowledge Management, 2025

    2025

  36. [36]

    Ade, Andy Laws, Mark Evans, and Paul S

    Wonwoo Ju, Dominic Doran, Richard Hawkins, Antonio Gómez-Díaz, Andres Martin-Garcia, Jack D. Ade, Andy Laws, Mark Evans, and Paul S. Bradley. Con- textualised peak periods of play in English Premier League matches.Biology of Sport, 39(4):973–983, 2022

    2022

  37. [37]

    Ball trajectory inference from multi-agent sports contexts using set transformer and hierarchical bi-LSTM

    Hyunsung Kim, Han-Jun Choi, Chang Jo Kim, Jinsung Yoon, and Sang-Ki Ko. Ball trajectory inference from multi-agent sports contexts using set transformer and hierarchical bi-LSTM. InProceedings of the 29th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2023

    2023

  38. [38]

    ELASTIC: Event-tracking data synchronization in soccer with- out annotated event locations

    Hyunsung Kim, Hoyoung Choi, Sangwoo Seo, Tom Boomstra, Jinsung Yoon, and Chanyoung Park. ELASTIC: Event-tracking data synchronization in soccer with- out annotated event locations. InECML PKDD Workshop on Machine Learning and Data Mining for Sports Analytics, 2025

    2025

  39. [39]

    Contextual sprint classification in soccer based on deep learning

    Hyunsung Kim, Gun-Hee Joe, Jinsung Yoon, and Sang-Ki Ko. Contextual sprint classification in soccer based on deep learning. InIJCAI Workshop on Intelligent Technologies for Precision Sports Science, 2024

    2024

  40. [40]

    SoccerCPD: Formation and role change-point detection in soccer matches using spatiotemporal tracking data

    Hyunsung Kim, Bit Kim, Dongwook Chung, Jinsung Yoon, and Sang-Ki Ko. SoccerCPD: Formation and role change-point detection in soccer matches using spatiotemporal tracking data. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022

    2022

  41. [41]

    Lafferty, Andrew McCallum, and Fernando C

    John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. InProceedings of the 18th International Conference on Machine Learning, 2001

    2001

  42. [42]

    Neural architectures for named entity recognition

    Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, Kazuya Kawakami, and Chris Dyer. Neural architectures for named entity recognition. InProceed- ings of the 2016 Conference of the North American Chapter of the Association for Hyunsung Kim et al. Computational Linguistics: Human Language Technologies, 2016

    2016

  43. [43]

    Kosiorek, Seungjin Choi, and Yee Whye Teh

    Juho Lee, Yoonho Lee, Jungtaek Kim, Adam R. Kosiorek, Seungjin Choi, and Yee Whye Teh. Set Transformer: A framework for attention-based permutation- invariant neural networks. InProceedings of the 36th International Conference on Machine Learning, 2019

    2019

  44. [44]

    Constrained decoding for computationally efficient named entity recognition taggers

    Brian Lester, Daniel Pressel, Amy Hemmeter, Sagnik Ray Choudhury, and Srini- vas Bangalore. Constrained decoding for computationally efficient named entity recognition taggers. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, 2020

    2020

  45. [45]

    EvolveGraph: Multi- agent trajectory prediction with dynamic relational reasoning

    Jiachen Li, Fan Yang, Masayoshi Tomizuka, and Chiho Choi. EvolveGraph: Multi- agent trajectory prediction with dynamic relational reasoning. InAdvances in Neural Information Processing Systems 33, 2020

    2020

  46. [46]

    Divergence measures based on the Shannon entropy.IEEE Transac- tions on Information Theory, 37(1):145–151, 1991

    Jianhua Lin. Divergence measures based on the Shannon entropy.IEEE Transac- tions on Information Theory, 37(1):145–151, 1991

    1991

  47. [47]

    Individual ball possession in soccer.PLOS ONE, 12(7), 2017

    Daniel Link and Martin Hoernig. Individual ball possession in soccer.PLOS ONE, 12(7), 2017

    2017

  48. [48]

    A network theory analysis of football strategies

    Javier López Peña and Hugo Touchette. A network theory analysis of football strategies. InSports Physics: Proceedings of the 2012 Euromech Physics of Sports Conference, pages 517–528, Palaiseau, France, 2012

    2012

  49. [49]

    Smith, and Yejin Choi

    Ximing Lu, Sean Welleck, Peter West, Liwei Jiang, Jungo Kasai, Daniel Khashabi, Ronan Le Bras, Lianhui Qin, Youngjae Yu, Rowan Zellers, Noah A. Smith, and Yejin Choi. NeuroLogic A*esque Decoding: Constrained text generation with lookahead heuristics. InProceedings of the 2022 Conference of the North Ameri- can Chapter of the Association for Computational ...

    2022

  50. [50]

    Matthews, and Yaser Sheikh

    Patrick Lucey, Alina Bialkowski, Peter Carr, Stuart Morgan, Iain A. Matthews, and Yaser Sheikh. Representing and discovering adversarial team behaviors using player roles. InIEEE Conference on Computer Vision and Pattern Recognition, 2013

    2013

  51. [51]

    Matthews

    Patrick Lucey, Dean Oliver, Peter Carr, Joe Roth, and Iain A. Matthews. Assessing team strategy using spatiotemporal data. InProceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2013

    2013

  52. [52]

    Xuezhe Ma and Eduard H. Hovy. End-to-end sequence labeling via bi-directional LSTM-CNNs-CRF. InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics, 2016

    2016

  53. [53]

    What players do with the ball: A physically constrained interaction modeling

    Andrii Maksai, Xinchao Wang, and Pascal Fua. What players do with the ball: A physically constrained interaction modeling. InIEEE Conference on Computer Vision and Pattern Recognition, 2016

    2016

  54. [54]

    Leapfrog diffusion model for stochastic trajectory prediction

    Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  55. [55]

    Game-changing IoT sensors in 2022 World Cup soccer balls de- liver real-time performance data, 2022

    Monnit Inc. Game-changing IoT sensors in 2022 World Cup soccer balls de- liver real-time performance data, 2022. https://www.monnit.com/blog/game- changing-iot-sensors-in-2022-world-cup-soccer-balls/. Accessed: Jan 21, 2026

    2022

  56. [56]

    DAG-Net: Double attentive graph neural network for trajectory forecasting

    Alessio Monti, Alessia Bertugli, Simone Calderara, and Rita Cucchiara. DAG-Net: Double attentive graph neural network for trajectory forecasting. InProceedings of the 25th International Conference on Pattern Recognition, 2020

    2020

  57. [57]

    A general method applicable to the search for similarities in the amino acid sequence of two proteins.Journal of Molecular Biology, 48(3):443–453, 1970

    Saul Needleman and Christian Wunsch. A general method applicable to the search for similarities in the amino acid sequence of two proteins.Journal of Molecular Biology, 48(3):443–453, 1970

    1970

  58. [58]

    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 Re- casens, 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 Tuyls. Multiagent off-screen behavior prediction in football.Scientific Reports...

    2022

  59. [59]

    Constraining linear-chain CRFs to regular languages

    Sean Papay, Roman Klinger, and Sebastian Padó. Constraining linear-chain CRFs to regular languages. InProceedings of the 10th International Conference on Learning Representations, 2022

    2022

  60. [60]

    Multi-modal soccer scene analysis with masked pre-training

    Marc Peral, Guillem Capellera, Luis Ferraz, Antonio Rubio, and Antonio Agudo. Multi-modal soccer scene analysis with masked pre-training. InIEEE/CVF Winter Conference on Applications of Computer Vision, 2026

    2026

  61. [61]

    Temporally accurate events detection through ball possessor recognition in soccer

    Marc Peral, Guillem Capellera, Antonio Rubio, Luis Ferraz, Francesc Moreno- Noguer, and Antonio Agudo. Temporally accurate events detection through ball possessor recognition in soccer. InProceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 2025

    2025

  62. [62]

    An iterative procedure for the polygonal approximation of plane curves.Computer Graphics and Image Processing, 1(3):244–256, 1972

    Urs Ramer. An iterative procedure for the polygonal approximation of plane curves.Computer Graphics and Image Processing, 1(3):244–256, 1972

    1972

  63. [63]

    Jones, and Ming Xu

    Jinchang Ren, James Orwell, Graeme A. Jones, and Ming Xu. Tracking the soccer ball using multiple fixed cameras.Computer Vision and Image Understing, 113(5):633–642, 2009

    2009

  64. [64]

    Matthews

    Long Sha, Patrick Lucey, Yisong Yue, Peter Carr, Charlie Rohlf, and Iain A. Matthews. Chalkboarding: A new spatiotemporal query paradigm for sports play retrieval. InProceedings of the 21st International Conference on Intelligent User Interfaces, 2016

    2016

  65. [65]

    Fine-grained retrieval of sports plays using tree-based alignment of trajectories

    Long Sha, Patrick Lucey, Stephan Zheng, Taehwan Kim, Yisong Yue, and Sridha Sridharan. Fine-grained retrieval of sports plays using tree-based alignment of trajectories. InProceedings of the 11th ACM International Conference on Web Search and Data Mining, 2018

    2018

  66. [66]

    Kochenderfer, Jiajun Wu, and Nick Haber

    Fan-Yun Sun, Isaac Kauvar, Ruohan Zhang, Jiachen Li, Mykel J. Kochenderfer, Jiajun Wu, and Nick Haber. Interaction modeling with multiplex attention. In Advances in Neural Information Processing Systems 35, 2022

    2022

  67. [67]

    Fast structured decoding for sequence models

    Zhiqing Sun, Zhuohan Li, Haoqing Wang, Di He, Zi Lin, and Zhi-Hong Deng. Fast structured decoding for sequence models. InAdvances in Neural Information Processing Systems 32, 2019

    2019

  68. [68]

    Embedded-state latent conditional random fields for sequence labeling

    Dung Thai, Sree Harsha Ramesh, Shikhar Murty, Luke Vilnis, and Andrew McCal- lum. Embedded-state latent conditional random fields for sequence labeling. In Proceedings of the 22nd Conference on Computational Natural Language Learning, 2018

    2018

  69. [69]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems 30, pages 5998–6008, 2017

    2017

  70. [70]

    Automatic event detection in football using tracking data.Sports Engineering, 25(18), 2022

    Ferran Vidal-Codina, Nicolas Evans, Bahaeddine El Fakir, and Johsan Billingham. Automatic event detection in football using tracking data.Sports Engineering, 25(18), 2022

    2022

  71. [71]

    Andrew J. Viterbi. Error bounds for convolutional codes and an asymptotically optimum decoding algorithm.IEEE Transactions on Information Theory, 13(2):260– 269, 1967

    1967

  72. [72]

    Take your eyes off the ball: Improving ball-tracking by focusing on team play.Computer Vision and Image Understanding, 119:102–115, 2014

    Xinchao Wang, Vitaly Ablavsky, Horesh Ben Shitrit, and Pascal Fua. Take your eyes off the ball: Improving ball-tracking by focusing on team play.Computer Vision and Image Understanding, 119:102–115, 2014

    2014

  73. [73]

    Masked conditional random fields for sequence labeling

    Tianwen Wei, Jianwei Qi, Shenghuan He, and Songtao Sun. Masked conditional random fields for sequence labeling. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2021

    2021

  74. [74]

    Multi-camera 3d ball tracking framework for sports video.IET Image Processing, 14(15):3751–3761, 2020

    Wanneng Wu, Min Xu, Qiaokang Liang, Li Mei, and Yu Peng. Multi-camera 3d ball tracking framework for sports video.IET Image Processing, 14(15):3751–3761, 2020

    2020

  75. [75]

    Uncover- ing the missing pattern: Unified framework towards trajectory imputation and prediction

    Yi Xu, Armin Bazarjani, Hyung-Gun Chi, Chiho Choi, and Yun Fu. Uncover- ing the missing pattern: Unified framework towards trajectory imputation and prediction. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    2023

  76. [76]

    Yeh, Alexander G

    Raymond A. Yeh, Alexander G. Schwing, Jonathan Huang, and Kevin Murphy. Diverse generation for multi-agent sports games. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019

    2019

  77. [77]

    ROLAND: Graph learning framework for dynamic graphs

    Jiaxuan You, Tianyu Du, and Jure Leskovec. ROLAND: Graph learning framework for dynamic graphs. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2022

    2022

  78. [78]

    Generating multi-agent trajectories using programmatic weak supervision

    Eric Zhan, Stephan Zheng, Yisong Yue, Long Sha, and Patrick Lucey. Generating multi-agent trajectories using programmatic weak supervision. InProceedings of the 7th International Conference on Learning Representations, 2019

    2019

  79. [79]

    Shuai Zheng, Sadeep Jayasumana, Bernardino Romera-Paredes, Vibhav Vineet, Zhizhong Su, Dalong Du, Chang Huang, and Philip H. S. Torr. Conditional random fields as recurrent neural networks. In2015 IEEE International Conference on Computer Vision, 2015. PathCRF: Ball-Free Soccer Event Detection via Possession Path Inference A Forward and Viterbi Algorithms...

    2015