{"id":"281e3806-5a93-48a4-9053-85c0b6e7256a","arxiv_id":"2506.04365","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"PLUCC fuses player-segmentation context with multiscale image features to localize the puck, reporting 12.2% higher average precision than the tested baselines and a new rink-space error metric.","lead":"This paper describes a computer vision model called PLUCC that finds the hockey puck in broadcast video by also looking at where players are on the ice. It reports large accuracy gains over two standard object detectors and introduces a new way to measure puck-location error in real rink coordinates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported SOTA margin is not trustworthy because the Gaussian sigma hyperparameter is selected on the test set, inflating PLUCC's numbers relative to untuned baselines.","rationale":"The reader's verdict is CONDITIONAL, and I agree that a conditional stance is appropriate. I depart from the reader's weakest_assumption: the most load-bearing issue is not whether the frozen player detector/SAM2 masks are accurate, but the test-set model selection in Section 4.6.2. A noisy context signal would reduce the method's advantage, but the paper's core comparison could survive if the advantage is real; test-set tuning directly corrupts the reported margin for the headline claim and is not a matter of mechanism. The σ=5 choice changes AP5 by 6 points (Table 1), which is large relative to the claimed 7.4% context gain and the 12.3% overall margin. The fix is straightforward: select σ on validation and re-report. Because the reader already conditioned acceptance on additional evidence, my concern strengthens the condition rather than changing the verdict; hence UNCHANGED. I also note that the context-encoder attribution is confounded by added capacity, since the FCN baseline lacks the GateBlocks, but that is secondary to the test-set selection issue.","tokens_in":12700,"tokens_out":5355,"duration_ms":50611,"concrete_test":"Re-run the σ ablation in Section 4.6.2 using only the validation split (the same split used for LR reduction in Section 4.2), select σ by best validation mAPτ, then evaluate the chosen model once on the held-out test split; recompute the margin vs Faster-RCNN in Table 1. If the validation-selected PLUCC test mAPτ advantage drops below ~12.3% (or AP5 below 7.4%), the SOTA claim is not established. Ideally, also report the same validation-selection procedure for FCN-ResNet152.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—PLUCC outperforms Faster-RCNN by 12.3% mAPτ—depends on a hyperparameter chosen using the test set. Section 4.2 fixes σ=5 and refers to Section 4.6.2; Section 4.6.2 states the σ ablation was evaluated 'on the test set' and σ=5 was selected as best across thresholds. Table 1 shows the choice is material: PLUCC σ=5 gives mAPτ 83.5/AP5 82.2, while PLUCC σ=15 gives 81.6/76.2. The baselines (Faster-RCNN, YOLOv5) are not given this test-set tuning, and the FCN baseline is also trained with σ=5 rather than selected on validation. Hence the reported 12.3% margin (and the 7.4% AP5 context-encoder gain) may reflect selection on the evaluation set, not a genuine method advantage. This is more fundamental than the upstream-context concern: even if player masks are perfectly informative, the headline numbers are not an unbiased estimate until the σ choice is made on validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PLUCC, a single-frame heatmap-based puck detector that fuses standard RGB features with an RGB player-segmentation context image through a gated decoder, and introduces RSLE, a homography-based rink-space evaluation metric. The method is evaluated on a proprietary VIP-PuckDataset against Faster-RCNN, YOLOv5, and an FCN-ResNet152 baseline, with reported gains of 12.3% in mAP_tau over Faster-RCNN and large improvements in rink-space AP. The core hypothesis, that player position and gaze provide a useful puck prior, is plausible, and the match-disjoint train/test split is a strength. However, the central SOTA claim is not yet supported because the key Gaussian sigma hyperparameter is selected on the test set, the comparison omits all puck-specific methods discussed in the paper, and the homography underlying RSLE is not validated.","tokens_in":12908,"tokens_out":5295,"duration_ms":47242,"significance":"If the results hold, the main contribution is evidence that pretrained player segmentation, encoded as RGB context, improves single-frame puck localization under challenging broadcast conditions, and the RSLE metric would be a useful complement to pixel-space AP. Strengths include the match-disjoint split, the explicit ablation of the Gaussian target, and the attempt to address perspective bias with a rink-space metric. However, the reported effect sizes are not yet credible because the key hyperparameter is selected on the test set, the comparison omits puck-specific methods, and the homography basis of RSLE is unvalidated. The architecture is a reasonable fusion of existing components, and the application is well motivated, but the paper currently overstates the strength of the evidence.","major_comments":[{"comment":"The sigma value used for the main PLUCC results is selected on the test set. The text states that the sigma ablation was evaluated 'on the test set' and Figure 6 selects sigma=5, while Table 1 shows the choice is material: PLUCC sigma=5 gives mAP_tau 83.5 and AP5 82.2, whereas PLUCC sigma=15 gives mAP_tau 81.6 and AP5 76.2. The baselines were not given comparable test-set tuning, so the headline 12.3% mAP_tau gain over Faster-RCNN and the 7.4% AP5 context-encoder gain may reflect selection on the evaluation set rather than a genuine method advantage. Please select sigma on validation and report both tuned and untuned results.","section":"Section 4.6.2 and Table 1"},{"comment":"The comparison omits all puck-specific methods discussed in the paper, specifically PuckNet [49], Yang [55], Li et al. [24], Sarkhoosh et al. [37], and Pidaparthy et al. [29]. The abstract and Section 5 call PLUCC 'state-of-the-art' and claim it surpasses 'previous baseline methods', but only generic object detectors and one heatmap FCN are evaluated. Please add these methods where they can be run on the same data, or clearly state and justify why they are excluded; otherwise the SOTA claim is not supported.","section":"Section 2.1 and Table 1"},{"comment":"The RSLE comparison relies on per-frame homographies from Shang et al. [39], but the paper reports no validation of those homography estimates on VIP-PuckDataset. If H is inaccurate, the reported AP_r and RSLEavg values are not meaningful and the abstract's 25% RSLE improvement is unverifiable. Please report homography registration error on the test set and evaluate how sensitive RSLE is to homography errors.","section":"Section 4.4.2 and Table 2"},{"comment":"No error bars or multiple-seed runs are reported. Differences such as PLUCC mAP_tau 83.5 vs FCN-ResNet152 79.6 and AP_r 43.59 vs 41.85 are reported as point estimates; without variance, the claimed 3.9% context-encoder gain and 1.74% rink-space gain cannot be assessed. Please provide mean and standard deviation over at least three seeds and a significance test for the central comparisons.","section":"All experiments (Tables 1-4)"},{"comment":"The context encoder relies on a pretrained player detector followed by SAM 2 segmentation, but the paper does not report the accuracy of these upstream models on VIP-PuckDataset. Table 3 shows that a network trained only on context images achieves only 6.0 mAP_tau, so the claimed benefit of context depends on the reliability of the frozen upstream masks. Please report the player detector and segmentation accuracy on the test set, and include an ablation that perturbs or corrupts the context masks to establish robustness.","section":"Section 3.1.2 and Table 3"}],"minor_comments":[{"comment":"There are typos: 'F or evaluation' in the abstract and 'do to similar colours' in the Figure 1 caption.","section":"Abstract and Figure 1 caption"},{"comment":"The text calls sigma the variance of the Gaussian label, but the exponent uses 2*sigma^2, which means sigma is actually the standard deviation; please clarify.","section":"Equation (8)"},{"comment":"The variable list repeats 'xrink' twice ('xrink and xrink'); the second should be 'yrink'.","section":"Equation (12)"},{"comment":"The 'pretrained detector' used to generate player bounding boxes is never cited or named; please specify the detector architecture and checkpoint.","section":"Section 3.1.2"},{"comment":"The text says 'using billinear interpolation' in the decoder description; it should be 'bilinear'.","section":"Section 4.2"},{"comment":"The abstract reports a 12.2% average precision improvement while Section 4.5 reports 12.3%; please harmonize these numbers.","section":"Abstract and Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"Editor: I want to flag that the homography method [39] is from the same group as the current paper. That is not by itself a problem, but it reinforces the need for independent validation of the RSLE metric. The proprietary dataset also limits reproducibility; the authors should at least release code and, if possible, a public benchmark or a data-access agreement. The 'state-of-the-art' language in the abstract and conclusion is premature given the missing puck-specific baselines and the test-set hyperparameter selection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a sincere engineering paper with one genuinely new idea and one clear flaw in the evidence. The new idea is PLUCC's fusion of player-segmentation context with feature-pyramid features through a static gating decoder, plus the RSLE rink-space metric. The writing is clear, and the ablations are thoughtful. The context-only FCN producing only 6 mAP but consistently peaking along the players' line of sight is the best thing in the paper: it shows the proposed context cue actually carries the intended signal.\n\nThe stress-test concern holds up. Section 4.6.2 selects sigma=5 on the test set, and Table 1 shows the choice matters: PLUCC with sigma=5 gets 83.5 mAP, with sigma=15 gets 81.6. The baselines are not given the same test-set tuning, so the reported 12.3% margin over Faster-RCNN includes selection bias. That does not make the method fake, but it makes the headline number untrustworthy as an unbiased comparison.\n\nThe other soft spots are real but secondary. The baselines omit prior puck-specific methods like PuckNet, Pidaparthy, and Yang, so 'state-of-the-art' is not supported even setting aside the sigma issue. RSLE depends on a homography from a self-cited method, but the homography accuracy on this dataset is never validated. No error bars are reported. The dataset is proprietary and no code is released, so independent verification is not possible. None of this is circular or incoherent; the architecture is not fitting the metric in a hidden way.\n\nIf the authors redo the hyperparameter selection on validation, add at least one puck-specific baseline, and report variance across runs, the core claim would be much easier to assess. As it stands, the paper is a plausible contribution to sports vision, not a demonstrated SOTA result. The right venue is peer review with a referee who will insist on those fixes. I would not cite the headline numbers, but I would cite the architecture and RSLE once they are properly supported.","headline":"A credible context-cue architecture and a well-motivated metric, but the 12.3% SOTA claim rests on test-set hyperparameter tuning and a comparison set too thin to support it.","tokens_in":13452,"tokens_out":1489,"would_cite":false,"duration_ms":15985,"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":"By feeding a network a picture of only the players, PLUCC boosts single-frame puck detection by 12.3% mean average precision over the best generic baseline.","keywords":["ice hockey puck detection","contextual cues","player segmentation","gated feature fusion","Gaussian heatmap","rink-space localization error","homography evaluation","single-frame detection"],"falsifier":"Replace the player-segmentation context image in the PLUCC pipeline with a deliberately uninformative version (e.g., random masks or masks of players shifted to the opposite side of the rink) and measure mAP on the VIP-PuckDataset test set; if the score does not drop materially, the reported gains do not come from player context. A complementary test is to compare PLUCC's accuracy on frames where the puck is far from any player against frames where players surround the puck, since the contextual cue should be weakest in the former.","tokens_in":12500,"feed_emoji":"🏒","tokens_out":12657,"duration_ms":98122,"temperature":0.7,"pith_summary":"This paper tries to establish that player behavior is a strong, learnable cue for locating a hockey puck in broadcast video: players consistently turn their bodies and gaze toward the puck, and a network that is explicitly shown a segmented player image can use that cue to localize the puck where appearance alone fails. The proposed system, PLUCC, processes the full-resolution frame and a half-resolution image containing only the segmented players through two encoders, fuses their multi-scale features with a learned gating mechanism, and outputs a Gaussian heatmap of puck position. On the proprietary VIP-PuckDataset test set, PLUCC reports a 12.3% mean average precision improvement over the strongest generic detector tested (Faster-RCNN), and a new rink-space metric, RSLE, shows its detections land within one puck radius more than twice as often as that baseline. If correct, this means single-frame, context-driven detection is a viable path to accurate puck tracking without multi-frame temporal methods or expensive player-tracking hardware.","feed_headline":"Player gaze cues lift puck detection accuracy by 12 percent","feed_subtitle":"The 12 percent edge comes from teaching the network to read player orientation instead of just puck pixels.","key_machinery":"The load-bearing mechanism is a gated feature-fusion decoder operating on two encoders: a ResNet-152 feature pyramid that processes the full-resolution RGB frame, and a context encoder that processes a half-resolution RGB image containing only the segmented players, produced by a frozen player detector followed by an off-the-shelf segmentation model. At each fusion stage the decoder concatenates corresponding-scale features from the two encoders and applies a static per-channel gating vector (a sigmoid over learned weights) before a convolution block, letting the context selectively amplify or suppress visual channels. Training uses Gaussian heatmap labels centered on the puck with a KL-divergence loss, and a 1% context-driven dropout of the raw frame pushes the network to lean on player cues. For evaluation, RSLE computes a per-frame homography into standardized rink coordinates so that pixel distances are converted to physical meters, removing the perspective bias of broadcast camera angles.","core_discovery":"On the paper's own terms, the central discovery is that player context is a learnable prior strong enough to carry puck localization: PLUCC, which fuses RGB player segmentations with standard image features in a gated multi-scale decoder, outperforms every baseline detector at every distance threshold, with mean average precision of 83.5 versus 71.3 for Faster-RCNN and 79.6 for the context-free FCN-ResNet152. The largest gains appear at the strictest threshold (AP5 of 82.2 versus 69.6), which the authors attribute to the context encoder implicitly capturing player orientation and position. An ablation supports this interpretation: a network trained only on the player-segmentation image predicts heatmaps that peak in the direction of players' gaze, even though its absolute accuracy is low, and a 1% training dropout of the raw frame raises AP5 by 3.5 points and lowers average rink-space error from 1.36 to 1.05 meters. The paper also introduces RSLE, a homography-based metric that projects image detections into NHL rink coordinates, and under that metric PLUCC achieves 43.6% average precision within one puck radius compared with 18.6% for the best generic baseline.","pith_inferences":["Editorial inference: the same context-encoder design may transfer to other broadcast sports in which a small object's position is implied by player orientation and spacing, such as lacrosse or rugby, but the reported 12-point gain should be expected to shrink if players do not reliably orient toward the object.","Editorial inference: the paper does not report the upstream player detector's or segmenter's accuracy on this dataset; using ground-truth player masks instead of the frozen pipeline would reveal how much of the measured gain comes from the context signal itself rather than from good upstream segmentation.","Editorial inference: RSLE's sensitivity to homography estimation error is not quantified; a natural test is to perturb the per-frame homography matrices and measure how much APr and RSLEavg change.","Editorial inference: at 6.03 FPS including preprocessing, the frozen player detector and segmenter dominate runtime; swapping them for faster lightweight models could make the system real-time, with unknown but testable cost to the context signal's usefulness."],"forward_implications":["PLUCC produces a detection per frame with no temporal input, so it can be deployed in streaming settings and will not inherit error propagation from multi-frame tracking methods.","Because the output is a Gaussian heatmap, the detections plug directly into the heatmap-based sports-ball tracking pipeline that motivated the label design.","The 1% context-driven dropout is a cheap training change that buys 3.5 points of AP5 and 0.31 meters of rink-space accuracy, making the model more robust to occlusion and visually masked pucks.","The RSLE metric gives hockey analytics a way to compare detectors in physical rink coordinates, which the authors argue is fairer than pixel-space distances that are distorted by camera perspective."],"supporting_citations":[{"why":"Faster-RCNN is the strongest generic object-detection baseline and the main comparison anchor for the reported 12.3% mAP gain.","marker":"[36]"},{"why":"FCN-ResNet152 is the context-free architectural counterpart used to isolate the contribution of the context encoder and gating.","marker":"[53]"},{"why":"Supplies the segmentation model that produces the player-mask context image from player detections.","marker":"[32]"},{"why":"Provides the Gaussian heatmap label formulation and the sports-ball detection baseline that motivates the training target.","marker":"[43]"},{"why":"Documents the correlation between player location and puck location that motivates using player context as a prior.","marker":"[29]"},{"why":"PuckNet is the temporal-context method whose player-density assumption PLUCC extends with explicit player segmentation, and whose limitations motivate single-frame context.","marker":"[49]"},{"why":"Provides the per-frame homography estimation technique on which the RSLE rink-space evaluation is built.","marker":"[39]"},{"why":"The channel reweighting idea in squeeze-and-excitation is the stated inspiration for the decoder's static gating.","marker":"[21]"}],"fun_headline_variants":["Puck detection gets 12% boost from player gaze cues","Player orientation is key to finding hockey puck in video","Contextual model improves puck localization by 12%","AI reads player gaze to locate puck more accurately","Player context lifts puck detection average precision 12%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on the pretrained player detector and segmenter producing masks that reliably encode player position and body direction under the exact broadcast conditions in the test set; if these upstream models fail on a meaningful fraction of frames, the context encoder receives noise and the reported advantage could shrink.","fun_headline_variants_meta":{"raw":{"variants":["Puck detection gets 12% boost from player gaze cues","Player orientation is key to finding hockey puck in video","Contextual model improves puck localization by 12%","AI reads player gaze to locate puck more accurately","Player context lifts puck detection average precision 12%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000351,"raw_usage":{"total_tokens":1968,"prompt_tokens":1055,"completion_tokens":913,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":671,"completion_tokens_details":{"reasoning_tokens":836}},"tokens_in":671,"tokens_out":913,"duration_ms":8990,"temperature":1.0,"reasoning_tokens":836,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:44:10.167837+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the player-segmentation context image in the PLUCC pipeline with a deliberately uninformative version (e.g., random masks or masks of players shifted to the opposite side of the rink) and measure mAP on the VIP-PuckDataset test set; if the score does not drop materially, the reported gains do not come from player context. A complementary test is to compare PLUCC's accuracy on frames where the puck is far from any player against frames where players surround the puck, since the contextual cue should be weakest in the former.","supporting_citations":[{"cited_title":"V ote from the Center: 6 DoF Pose Estimation in RGB-D Images by Radial Keypoint V oting, 2022","cited_arxiv_id":null,"evidence_quote":"FCN-ResNet152 is the context-free architectural counterpart used to isolate the contribution of the context encoder and gating."},{"cited_title":"Widely Applicable Strong Baseline for Sports Ball Detection and Tracking","cited_arxiv_id":"2311.05237","evidence_quote":"Provides the Gaussian heatmap label formulation and the sports-ball detection baseline that motivates the training target."},{"cited_title":"Keep your eye on the puck: Automatic hockey videography","cited_arxiv_id":null,"evidence_quote":"Documents the correlation between player location and puck location that motivates using player context as a prior."},{"cited_title":"Puck localization and multi-task event recognition in broad- cast hockey videos","cited_arxiv_id":null,"evidence_quote":"PuckNet is the temporal-context method whose player-density assumption PLUCC extends with explicit player segmentation, and whose limitations motivate single-frame context."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the per-frame homography estimation technique on which the RSLE rink-space evaluation is built."}],"review_version":1}