{"id":"31bf46e5-39d3-4748-ae26-2c1a312f928d","arxiv_id":"2506.23843","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"EFPI matches player tracking data to 65 football formation templates using linear sum assignment, with position scaling and a stability parameter, to label formations and player roles.","lead":"This paper presents EFPI, a method that labels team formations and player positions by matching tracking data to 65 predefined templates with an optimal assignment algorithm. It may be useful for automated tactical analysis because it is transparent, open source, and works with multiple tracking data providers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The per-frame linear scaling of player positions is the load-bearing step: it can change the cost ranking across templates, and the paper gives no evidence that the rescaled geometry preserves the true formation.","rationale":"The reader's weakest assumption is the scaling normalization, and I agree that it is the crucial unvalidated step. The paper's headline contribution is not the Hungarian assignment (Wei et al.) nor the template set (mplsoccer), but the scaling that is claimed to fix illogical assignments. If that scaling is not guaranteed to preserve the relevant geometry, then EFPI's formation and role labels are just the cheapest template after an arbitrary transform. The manuscript provides no error analysis, no baseline comparison, and no synthetic experiment, so the condition for the central claim to hold is untested. I do not see an internal contradiction or a demonstrable counterexample in the text; the method is clearly described and open-source, which is why the appropriate verdict remains conditional rather than reject. The concrete test above would settle whether the normalization preserves true formation identity in realistic game states.","tokens_in":6235,"tokens_out":5416,"duration_ms":61508,"concrete_test":"Use the open Sportec tracking data from Bassek et al. (2025), which the paper itself uses in Codeblock 1, and construct a ground-truth set of formation/role labels for at least one full match (e.g., from lineups plus manual annotation of formation changes). Run EFPI in four configurations: (a) no scaling, (b) the paper's scaling to the template set's maximum width and length, (c) scaling to each candidate template's own width and length, and (d) separate z-score standardization of x and y. Report per-frame formation-label accuracy and role-label accuracy by game state (defensive low block, mid-block, high press, transition). If configuration (b) does not dominate (a) and (d) on held-out frames, the scaling step is not 'largely resolving' the assignment problem.","verdict_should_be":"UNCHANGED","load_bearing_attack":"EFPI's only novel ingredient is a per-frame scaling of player positions to the maximum width/length of the template set before the linear-sum assignment in Eq. 1. This normalization changes the relative geometry of the player cloud, and the assignment cost is not invariant under it: the x- and y-scale factors depend on the current bounding box, so the same real formation can map to different normalized shapes in a compact low block versus a stretched high press. The paper does not compare the scaled and unscaled cost surfaces, does not ablate the scaling, and reports no accuracy against ground-truth formation/role labels. The single illustrative frame in Figure 2 and the qualitative Portugal-Switzerland case (Figures 4-5) cannot establish that the minimum-cost template after scaling corresponds to the true formation. Since the claim that scaling 'largely resolves' Poppeliers's illogical-assignment problem is the load-bearing bridge from a generic template-matching idea to a usable tactical-analytics tool, this is the step that must be validated before the method is accepted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces EFPI, a rule-based method for football formation recognition and per-player role labeling from spatiotemporal tracking data. For each frame or time segment, player coordinates are linearly scaled to the maximum width and length of a fixed set of 65 static formation templates, a linear sum assignment solver matches players to template positions for each template, and the lowest-cost template is selected. An optional stability parameter suppresses formation changes when the relative cost improvement is below a threshold. The method is illustrated on a single frame and on Portugal's 2022 World Cup match against Switzerland, and the implementation is released through the open-source unravelsports Python package.","tokens_in":6442,"tokens_out":3176,"duration_ms":40396,"significance":"If the scaling step can be shown to preserve the relevant geometry of player arrangements, EFPI would provide a simple, transparent, and computationally cheap baseline for dynamic formation and role labeling, complementing supervised and change-point approaches such as those of Pleuler (2024) and Kim et al. (2022). The open-source implementation and support for multiple tracking-data providers through Kloppy are clear practical strengths, as is the use of an openly available dataset for the code example. However, the paper currently offers no quantitative evidence for its central claim that the method 'functions effectively': there is no ground-truth comparison, no error metric, no sensitivity analysis, and no baseline comparison. The contribution is therefore a plausible algorithmic recipe whose practical validity remains unestablished.","major_comments":[{"comment":"The per-frame scaling of player positions to the maximum width and length of the template set is the load-bearing novelty of the method, but it is neither justified nor validated. The scaling factors depend on the current bounding box of the team, so the same real formation can map to different normalized shapes in a compact low block versus a stretched high press, and the assignment cost in Eq. (1) is not invariant under this normalization. The paper provides no comparison of scaled versus unscaled assignment costs, no ablation of the scaling, and no sensitivity analysis. The single illustrative frame in Figure 2 cannot establish that the lowest-cost template after scaling corresponds to the true formation. The claim that scaling 'largely resolves the illogical assignment problem pointed out by Poppeliers (2025)' therefore needs quantitative support, for example by comparing assignment accuracy against ground-truth formation and role labels on a labeled dataset.","section":"Section 2.1"},{"comment":"The central claim that EFPI 'functions effectively on individual frames' is not supported by any quantitative experiment. The paper reports no accuracy, precision, recall, or error rates for formation identification or player-role assignment, and it compares against no baseline method such as Wei et al. (2013), Pleuler (2024), or SoccerCPD. The qualitative Portugal-Switzerland case study (Figures 4-5) and the single-frame example (Figure 2) cannot substantiate the claim. Because the paper's stated goal is a usable tactical-analytics tool, the authors should evaluate EFPI on the openly available dataset cited as [1], using ground-truth labels (e.g., starting lineups and known formation annotations), and report frame-level and segment-level metrics against at least one baseline.","section":"Entire manuscript (especially Abstract and Figures 2, 4, 5)"},{"comment":"The stability parameter is a free parameter, yet the paper provides no sensitivity analysis. The only demonstration is a single 24-second sequence in Figure 3, which is insufficient to show that the threshold reliably separates genuine formation changes from noise. Additionally, Eq. (3) uses C_{t-1} for the cost of the previously assigned template but does not specify what happens if that template was not the lowest-cost template at time t-1, or how the parameter interacts with the 'change after possession' option shown in Codeblock 1. A systematic robustness study over a range of epsilon values, and over multiple matches and game situations, is needed before the stability mechanism can be considered a validated component.","section":"Section 2.3, Eq. (3)"},{"comment":"The segmentation procedure that averages player positions over a segment (Eq. (2)) is presented without validation. Averaging over a full period or a five-minute window can obscure genuine tactical transitions and is sensitive to substitutions and out-of-possession excursions. The statement that segments are 'partitioned by attacking and defending periods' is not operationalized, and it is unclear how this interacts with possession data. The paper should compare frame-level vs segment-level assignments and report how often the segment-level label differs from the majority of frame-level labels, especially in transition-heavy intervals.","section":"Section 2.2"}],"minor_comments":[{"comment":"The caption refers to '31231' while the text and figure describe '31213'; the caption also contains the typo 'identifeied'. Please correct both.","section":"Figure 2 caption"},{"comment":"The expression should be written as (C_{t-1} - C_t)/C_t > epsilon, and each symbol should be defined at first use, including the distinction between the cost of the previously assigned template and the cost of the best template at time t.","section":"Eq. (3)"},{"comment":"The phrase 'partition these segments by attacking and defending periods' needs a precise definition, including how attacking and defending are determined from tracking or event data.","section":"Section 2.2"},{"comment":"Reference [7] is incomplete: no publication venue, arXiv identifier, or year-consistent citation details are given. Please provide a complete citation or remove it if it is not publicly available.","section":"References"},{"comment":"The heading 'Model Configuration and Training' is a misnomer because the method has no training phase. Also, the 'change_threshold' parameter is not explicitly connected to the stability parameter epsilon in the text.","section":"Codeblock 1"},{"comment":"There are several typographical and phrasing issues, including 'the the' in Section 2.4, 'individually players' in the introduction, and variable naming inconsistencies (N and q in Eq. (2)). A careful proofread is recommended.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central algorithm is simple and the open-source release is a positive feature, but the paper currently resembles a software note rather than a validated research contribution. I chose major_revision rather than reject because the missing validation is, in principle, fixable within the manuscript's scope: the authors already cite an open dataset and an open package, so adding a quantitative evaluation with ground-truth labels, sensitivity analyses, and baselines is feasible. If such validation is not added, the paper should not be accepted. I would also ask the editor to consider whether the journal's standards for empirical evaluation in applied ML are met by the current qualitative demonstration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nShort version: this is a clearly written, incremental method paper that ships code, but its central claim—that the scaling step fixes the assignment problem—is supported only by two qualitative examples. If it lands in a venue, it should be after a real evaluation.\n\nWhat's new: EFPI applies linear sum assignment (Wei et al.'s idea) to formation labeling with two additions: per-frame scaling of player positions to the template collection's max width/length before cost computation, and a hysteresis threshold on cost differences to stabilize labels across segments. The implementation is open source and works across several tracking data providers via Kloppy. That is genuinely useful for practitioners who want a transparent, low-parameter alternative to supervised classifiers like Pleuler's.\n\nThe paper's main illustration is honest about the problem: Figure 2 shows the unscaled assignment mislabeling a center back as a defensive midfielder, and the scaled version gives the expected 4411 and more sensible labels. That's a nice demonstration that the scaling can help in at least this frame. The Portugal-Switzerland case study is also a reasonable qualitative use-case.\n\nThe soft spot is exactly what the stress-test note says: the scaling is load-bearing and unvalidated. The scale factors depend on the current bounding box, so the same nominal formation can map to different normalized shapes depending on whether the team is compact or stretched. The cost surface is not invariant under this transformation, so there is no guarantee the lowest-cost template after scaling corresponds to the true formation. The paper provides no ablation of scaling, no comparison against unscaled assignment on a labeled dataset, no baseline (e.g., Wei et al., or a simple majority classifier), and no sensitivity analysis for epsilon or the scaling itself. The phrase 'largely resolves' is not backed by evidence. Also, the evaluation is entirely qualitative; there is no ground-truth formation/role labeling, so accuracy is unknown. The stability parameter is reasonable but also unexamined—how sensitive are segment-level labels to epsilon?\n\nThese are fixable. The method is simple enough that a careful comparison on a few matches with manually annotated formations and roles would settle it. I'd like to see that before trusting the results, but the paper is coherent and the code is public, so it deserves a proper review rather than a desk reject.\n\nMy recommendation: if you're refereeing, return it with a request for quantitative validation and sensitivity analyses. If you're just scanning the literature, it's worth a skim for the scaling idea, but don't cite it as a validated method until those numbers exist.","headline":"Incremental template-matching method with a promising scaling fix that is not yet validated by any quantitative experiment.","tokens_in":6914,"tokens_out":2074,"would_cite":false,"duration_ms":24140,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that scaling player positions to template dimensions before linear-sum assignment lets a single transparent method label both formations and individual roles from spatiotemporal tracking data.","keywords":["football analytics","formation recognition","player role labeling","linear sum assignment","Hungarian algorithm","spatiotemporal tracking data","template matching","tactical analysis"],"falsifier":"Take a sample of match frames with human-annotated formations and roles, run EFPI with and without scaling, and compare agreement. If the scaled version disagrees with the annotations systematically in compressed or stretched situations—teams packed deep, attacking in wide spaces, or playing narrow—then the scaling does not preserve the geometry needed for correct assignment.","tokens_in":6048,"feed_emoji":"⚽","tokens_out":5827,"duration_ms":60296,"temperature":0.7,"pith_summary":"EFPI claims that a team's formation and each outfield player's role can be read off from tracking data in one step, by matching players to static formation templates. The method first scales actual player positions to the width and length of the template set, then solves a linear sum assignment for each of 65 templates and selects the lowest-cost one. The paper argues that this pre-scaling is what fixes the tactically wrong assignments that unscaled template matching produces, and that the same procedure works on single frames, possessions, or arbitrary time windows. The contribution is a transparent alternative to supervised role classifiers, needing no training labels, with an optional stability parameter to suppress noise in transition moments.","feed_headline":"Scale positions first: formation and role labels in one pass","feed_subtitle":"A pre-scaling step lets one algorithm label both formation and player roles from tracking data.","key_machinery":"The mechanism is the linear sum assignment problem, solved with the Hungarian algorithm, applied to a cost matrix of Euclidean distances between players and template positions. Before any cost is computed, all player coordinates are scaled to the maximum width and length of the 65-template set, which turns the matching into a shape comparison rather than a raw proximity fit. Assignment is run separately for each template, the lowest-cost template wins, and the player-to-position matching from that template supplies the role labels. An optional stability parameter $\\epsilon$ blocks a template change unless the new cost is at least $\\epsilon$ percent lower than the previous one.","core_discovery":"On the paper's own terms, the central claim is that scaling player coordinates to the bounding dimensions of the formation templates before cost computation makes template matching tactically coherent. In the paper's illustrative frame, unscaled matching labels a center back as a defensive midfielder and a central midfielder as a winger; after scaling, the same frame is labeled 4411 with roles the author treats as correct. The paper states that this scaling largely resolves the illogical assignment problem identified in earlier work, and that formation labels and player position labels can therefore be produced instantly and simultaneously for any time segment, from single frames to full periods.","pith_inferences":["The scaling step is implicitly a claim that formation geometry is scale-invariant across match situations; a natural extension would test rotation or shear alignment, which the paper does not do.","Frame-level labels from EFPI could serve as weak supervision for role-classification models, but that use would require checking EFPI's own accuracy against annotated ground truth first.","The stability threshold and segment length interact: a fixed $\\epsilon$ may be too lax for long segments and too strict for short ones, so an adaptive threshold is a testable refinement.","A systematic comparison of EFPI's formation labels with human-annotated formations across varied match states would reveal whether the scaling truly resolves illogical assignments or only on the illustrated frame."],"forward_implications":["A single algorithm outputs both the formation label and individual role labels for every outfield player, without training data or manual annotation.","Because player locations can be averaged over any segment, the same code labels single frames, possessions, 5-minute windows, or full periods.","The stability parameter $\\epsilon$ suppresses spurious formation switches during transitions while still allowing genuine changes when the cost improvement is large enough.","Templates for 8, 9, and 10 outfield players let the method keep working when red cards or injuries reduce a team's numbers.","The open-source implementation consumes standardized tracking data, so it can be applied to data from a wide range of providers."],"supporting_citations":[{"why":"Establishes the treatment of formation recognition as an optimal assignment of players to formation positions, which EFPI builds on.","marker":"[12]"},{"why":"Identifies the illogical-assignment failure that the paper's scaling step is designed to resolve.","marker":"[7]"},{"why":"Provides the library of 65 static formation templates that define the candidate shapes and position labels.","marker":"[8]"},{"why":"Supplies the Hungarian algorithm used to solve the linear sum assignment problem.","marker":"[5]"},{"why":"Provides the computational routine for linear sum assignment used in the implementation.","marker":"[10]"},{"why":"Supplies the openly available spatiotemporal tracking data used to demonstrate the method.","marker":"[1]"},{"why":"Hosts the open-source implementation of EFPI that the paper describes.","marker":"[2]"}],"fun_headline_variants":["Scale first, match later: formation and roles in one pass","Pre-scaling coordinates fixes assignment: labels both formation and roles","One scaling step before matching: formation and roles resolved","Scaling positions before assignment yields accurate formation and role labels","EFPI: scale coordinates first for instant formation and role labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that linearly scaling all player positions to the widest and longest formation template preserves the relative geometry that makes the lowest-cost template the tactically correct one in every game situation; this normalization is introduced without a sensitivity analysis.","fun_headline_variants_meta":{"raw":{"variants":["Scale first, match later: formation and roles in one pass","Pre-scaling coordinates fixes assignment: labels both formation and roles","One scaling step before matching: formation and roles resolved","Scaling positions before assignment yields accurate formation and role labels","EFPI: scale coordinates first for instant formation and role labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1257,"prompt_tokens":833,"completion_tokens":424,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":341}},"tokens_in":449,"tokens_out":424,"duration_ms":5118,"temperature":1.0,"reasoning_tokens":341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:29:31.256362+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a sample of match frames with human-annotated formations and roles, run EFPI with and without scaling, and compare agreement. If the scaled version disagrees with the annotations systematically in compressed or stretched situations—teams packed deep, attacking in wide spaces, or playing narrow—then the scaling does not preserve the geometry needed for correct assignment.","supporting_citations":[{"cited_title":"Large-scale analysis of formations in soccer","cited_arxiv_id":null,"evidence_quote":"Establishes the treatment of formation recognition as an optimal assignment of players to formation positions, which EFPI builds on."},{"cited_title":"Automatic discovery of buildup tactics using football match data","cited_arxiv_id":null,"evidence_quote":"Identifies the illogical-assignment failure that the paper's scaling step is designed to resolve."},{"cited_title":"mplsoccer: A python package for plotting soccer/football charts in matplotlib","cited_arxiv_id":null,"evidence_quote":"Provides the library of 65 static formation templates that define the candidate shapes and position labels."},{"cited_title":"The Hungarian method for the assignment problem","cited_arxiv_id":null,"evidence_quote":"Supplies the Hungarian algorithm used to solve the linear sum assignment problem."},{"cited_title":"SciPy 1.0: fundamental algorithms for scientific computing in Python","cited_arxiv_id":null,"evidence_quote":"Provides the computational routine for linear sum assignment used in the implementation."},{"cited_title":"An integrated dataset of spatiotemporal and event data in elite soccer","cited_arxiv_id":null,"evidence_quote":"Supplies the openly available spatiotemporal tracking data used to demonstrate the method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Hosts the open-source implementation of EFPI that the paper describes."}],"review_version":1}