{"id":"bab668a5-66bf-4ff8-8903-ad137c9df347","arxiv_id":"2607.10998","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Temporal Feature Distillation plus Transformer Gate Shift lets Vision Transformers spot precise sports events with 10–80% labels, matching full supervision on two of four benchmarks.","lead":"This paper shows that standard self-distillation (like DINO) fails for precise sports-event spotting because it erases the tiny frame-to-frame motion cues that mark event boundaries. By aligning backbone features instead of projection heads, adding a gated temporal shift module, and using a supervised warm-up, the method reaches strong accuracy with far fewer labels.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The 4.54-point FSPerf@10% claim rests on adapted SSL baselines that share the authors' ViT-S+TGS backbone, so the gain may overstate TFD's contribution relative to true prior art.","rationale":"The reader's weakest-assumption correctly flags that backbone-feature alignment is only empirically justified (Fig. 4 is a single-clip illustration). That concern is real but secondary: even if backbone features are more boundary-sensitive, the quantitative claim that TFD is 4.54 points better than \"the strongest competing approach\" is only secure if the competitors are fairly instantiated. The controlled re-implementation protocol of §5.2 deliberately equalizes the backbone, which is good science for isolating the loss, yet it simultaneously weakens the external claim made in the abstract. Because the paper already acknowledges the adaptation caveat, the issue does not overturn the internal evidence that TFD+TGS works; it does, however, keep the verdict CONDITIONAL until either (a) native-baseline numbers are supplied or (b) code is released so others can verify the margin. Error bars remain desirable but are not the load-bearing gap. No stronger internal inconsistency was found.","tokens_in":15122,"tokens_out":597,"duration_ms":5931,"concrete_test":"Re-run the four SSL baselines (DINO, Teeti, Kwon, Zhou) under their original published architectures / temporal modules (or at least on plain ViT-S without TGS) with the same 10% FSPerf split and three seeds; if any native baseline reaches ≥48 mAP, the 4.54-point claim no longer holds against true prior art.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim (Abstract / Table 1) is that under 10% labels on FSPerf, TFD+TGS reaches 50.29 mAP vs. 45.75 for the strongest competitor (ASTRM). Section 5.2 states that all ViT-based SSL/USL baselines (DINO, Teeti et al., Kwon et al., Zhou et al.) were re-implemented on the identical ViT-S+TGS backbone, resolution, clip length and schedule \"to reduce confounding factors.\" Consequently the 4.54-point margin is measured against methods whose original architectures and temporal modules were replaced by the authors' own TGS. The paper never reports the same SSL objectives on their native backbones (or on a plain ViT without TGS). If those native versions already close most of the gap, the headline improvement attributed to Temporal Feature Distillation (Eq. 9 + warm-up) is inflated. The fully-supervised ASTRM number is left untouched, but the SSL comparison that underpins the \"strongest competing approach\" claim is not.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper addresses label-efficient Precise Event Spotting (PES) in sports videos. It argues that DINO-style projection-head self-distillation is poorly suited to PES because it over-smooths adjacent-frame representations and suppresses boundary-sensitive motion cues. The authors propose Temporal Feature Distillation (TFD), which instead aligns ℓ2-normalized backbone features from a student–EMA-teacher pair (Eqs. 8–11), after a supervised warm-up and cosine ramp-up of the distillation weight. They also introduce Transformer Gate Shift (TGS), a multi-scale element-wise gated temporal shift module for ViTs (Eqs. 1–7), and Temporal Motion Augmentation (TMA) that preserves high-motion patches and applies short temporal masks. Experiments on Tennis, FSComp, FSPerf, and FineDiving under 10–80% labels report consistent gains; the headline result is +4.54 mAP over ASTRM at 10% labels on FSPerf, with 80%-label performance matching or exceeding fully supervised 100% baselines on two of four datasets (supp.).","tokens_in":15494,"tokens_out":1713,"duration_ms":30257,"significance":"PES is a practically important, annotation-heavy sports-video task, and semi-supervised methods for it remain underexplored relative to image SSL and coarse action recognition. The core insight—that projection-level invariance can destroy the temporal discriminability PES needs—is well motivated and directly supported by the inter-frame cosine-similarity analysis in Figure 4 and the attention maps in Figure 5. Isolating the distillation objective on a shared ViT-S+TGS backbone (Section 5.2) is a sound experimental choice for comparing learning objectives. Ablations in Table 2 cleanly separate TGS, TFD, and TMA. If the gains hold under fuller statistical reporting and clearer baseline disclosure, the work is a useful, task-aware contribution to label-efficient fine-grained video localization, with reusable modules (TGS, TMA) beyond the specific distillation recipe.","major_comments":[{"comment":"Table 1 / Abstract: The main quantitative claims (including the +4.54 mAP FSPerf@10% result and the 80%-vs-100% claim) are reported as point estimates only, despite the protocol of three random seeds (Section 5.2). Without standard deviations or a significance test, it is hard to judge whether several of the smaller margins (e.g., Tennis@80% 93.04 vs ASTRM 92.97; FineDiving@80% 66.57 vs ASTRM 66.19 / Kwon 67.92) are reliable. Please add mean±std (or at least std) to Table 1 and state whether the headline improvements remain significant.","section":null},{"comment":"Section 5.2 and Table 1: All ViT-based SSL/USL baselines (DINO, Teeti et al., Kwon et al., Zhou et al.) are re-implemented on the authors’ ViT-S+TGS backbone, resolution, clip length, and schedule. This is appropriate for isolating the learning objective, but the paper never reports those same SSL objectives on their native backbones or on plain ViT without TGS. Table 2 shows that TGS alone accounts for large gains (e.g., ViT-S FSPerf@10%: 28.61 → 41.37). Please either (i) add a controlled “SSL objective on ViT-S without TGS” column, or (ii) explicitly qualify that the SSL ranking is under a shared TGS-enhanced backbone and that absolute gains relative to published native implementations may differ.","section":null},{"comment":"Abstract / Section 5.3: The claim that “with only 80% labeled data, it matches or surpasses the fully supervised 100% baseline on two of the four datasets” cannot be verified from the main text; 100% fully supervised numbers are deferred to the supplementary material, and Table 1 stops at 80%. For a claim featured in the abstract and Figure 1, the 100% supervised reference numbers (at least for ASTRM / E2E / the authors’ own fully supervised ViT-S+TGS) should appear in the main paper, with the same evaluation protocol.","section":null},{"comment":"Section 3.4 / Eq. (9) and the free-parameter set: The central modeling assumption—that backbone features after TGS are intrinsically more boundary-sensitive than projection-head outputs—is supported empirically by Figure 4 on one Tennis clip, but the method depends on several unablated schedule and module choices (5-epoch warm-up, 10-epoch cosine ramp of λ to 1, TGS deltas {1,3,5} and 3-way channel split, TMA 16×16 / top-9 / mask length 3, dual EMA decays). A short sensitivity study on warm-up length and λ schedule (or a statement that defaults transfer without retuning across the four datasets) is needed to show that the recipe is not brittle to these free parameters.","section":null}],"minor_comments":[{"comment":"Table 3: ViT-S+TGS costs ~1000 GFLOPs and 867 FPS vs ~40–45 GFLOPs / multi-thousand FPS for the CNN PES baselines. The paper notes that CNN/ViT families are not directly comparable, but a short discussion of when the accuracy–cost trade-off is justified for PES deployment would help readers.","section":null},{"comment":"Table 2, clip-length ablation: L=100 is suboptimal on FSPerf (several other lengths gain +5.5 to +8.7 mAP). The choice is justified as a community default, but the main results on FSPerf may understate the method; consider reporting the best L per dataset in the supplement or noting the sensitivity more prominently.","section":null},{"comment":"Figure 1 caption and Abstract: “matches or surpasses the fully supervised 100% baseline on 2 of 4 datasets” should name which two datasets and which 100% reference model, once those numbers are moved into the main paper.","section":null},{"comment":"Section 3.3 / Figure 3: TMA’s motion-map construction (frame difference, top-2% pixels per patch, keep top-9 of 16×16) is clear, but the probability 0.5 and mask settings appear only in Section 5.1; a brief pointer in Section 3.3 would improve reproducibility from the method section alone.","section":null},{"comment":"Related Work: Liu et al. [11] is correctly distinguished as multimodal; a one-sentence note on whether any pure-RGB SSL PES baseline existed before this work would sharpen the novelty claim.","section":null},{"comment":"Typos / polish: “generalisability” (Section 2.3) is fine in British English but inconsistent with “generalization” elsewhere; “The first attempts” capitalization in Section 2.3; ensure δ vs 𝛿 notation is uniform between text and equations.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The skeptic note overstates the confounding for the +4.54 headline: that margin is vs fully supervised ASTRM (45.75), not vs an adapted SSL method. The shared-backbone SSL comparison is actually the right control for isolating TFD vs other objectives; DINO collapsing even with TGS strengthens rather than weakens the paper’s thesis. I would not reject on that ground. Main remaining risks are statistical reporting and the abstract’s 80%-vs-100% claim living only in the supplement. Fit for a solid CV / sports-video venue after minor revision; not a landmark theoretical result, but a careful task-driven SSL design with useful empirical evidence."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: for frame-level event spotting, DINO-style projection invariance is the wrong inductive bias, and aligning backbone features after a short supervised warm-up works better. Figure 4 makes that concrete—DINO keeps inter-frame cosine near 1.0 across ground-truth boundaries while TFD drops sharply. That is the real contribution.\n\nWhat is new is the combination, not any single module. TGS is a clean multi-scale, element-wise gate-shift for ViT tokens (deltas 1/3/5, full-channel gates, residual). TFD is ordinary cosine alignment on ℓ2-normalized backbone features (Eq. 9) with a 5-epoch warm-up and cosine ramp of λ. TMA is a sensible motion-map + temporal-mask view generator. None of these is revolutionary, but they are well-motivated for PES and the ablations in Table 2 isolate them cleanly under 10% labels.\n\nTable 1 shows consistent gains across four public benchmarks and four label ratios, averaged over three seeds. At 80% labels the method matches or beats the 100% fully-supervised baselines on two datasets; that is practically useful for sports analytics. The fully-supervised ASTRM comparison is fair. The SSL comparison is less so: Section 5.2 states that DINO, Teeti, Kwon, Zhou etc. were all re-run on the authors’ own ViT-S+TGS backbone “to reduce confounding factors.” So the 4.54-point FSPerf@10% margin (50.29 vs 45.75) is measured against methods whose original temporal modules were replaced by TGS. That overstates pure TFD credit relative to true prior art. Missing error bars on the main table and the usual SSL hyper-parameter list (warm-up length, ramp, EMA, TMA grid) are secondary but real.\n\nMath is elementary and correct; citations cover the PES and SSL literature without heavy self-citation. Code is promised anonymously. This is for people who actually ship sports-event systems or work on fine-grained video SSL; it will not change general video understanding. It deserves a serious referee—ask for native-backbone SSL numbers and error bars, then accept with minor revision. I would read the camera-ready and probably cite the backbone-vs-projection observation.","headline":"Solid label-efficient PES recipe whose core insight (backbone features, not DINO heads) is real; the 4.54-point headline is partly inflated by shared-backbone SSL re-implementations.","tokens_in":16052,"tokens_out":585,"would_cite":true,"duration_ms":5958,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Aligning backbone features instead of projection heads preserves the subtle motion cues precise sports event spotting needs, so models reach high accuracy with far fewer frame labels.","keywords":["precise event spotting","semi-supervised learning","sports video understanding","temporal feature distillation","vision transformers","self-distillation","motion-aware augmentation"],"falsifier":"Train the same TGS-ViT backbone with pure DINO-style projection alignment (no backbone-feature loss, no warm-up) under 10 percent labels on the four datasets; if that projection-only model matches or exceeds TFD mAP and still shows sharp inter-frame similarity drops at ground-truth event boundaries, the claim that backbone alignment is required fails.","tokens_in":16052,"feed_emoji":"🎾","tokens_out":983,"duration_ms":20342,"temperature":0.7,"pith_summary":"Precise Event Spotting asks a model to mark the exact frame of a sports action—such as a stroke, pass, or dive entry—even when neighboring frames look almost the same. Standard image self-distillation forces representations to stay invariant across views, which washes out the tiny motion differences that mark true event boundaries. This paper argues that a different objective, Temporal Feature Distillation, keeps those boundary-sensitive cues by matching multi-scale backbone features rather than projection-head outputs, after a short supervised warm-up that first teaches the model what events look like. A gated multi-scale shift module injects temporal motion into ordinary Vision Transformers, and motion-aware augmentations force the model to focus on moving objects instead of static backgrounds. On four fine-grained sports benchmarks the method consistently beats supervised and semi-supervised baselines, matching full-supervision performance with only 80 percent of the labels on two datasets and delivering multi-point gains under 10 percent labels.","feed_headline":"4.5-point mAP gain at 10% labels for sports event spotting","feed_subtitle":"Backbone feature alignment keeps the boundary cues that DINO-style methods erase.","key_machinery":"Temporal Feature Distillation (TFD): a semi-supervised objective that aligns temporally structured backbone features (not projection-head outputs) between a strongly augmented student and a weakly augmented EMA teacher, so motion-sensitive and boundary-aware cues survive for frame-level localization.","core_discovery":"Direct DINO-style projection alignment is ineffective for Precise Event Spotting because it over-smooths adjacent frames and treats subtle motion as noise; aligning ℓ2-normalized backbone features from a temporally enhanced Vision Transformer, after supervised warm-up and ramp-up, preserves boundary sensitivity and yields large gains under limited frame-level labels.","pith_inferences":["The same backbone-feature alignment idea may transfer to other fine-grained temporal localization tasks outside sports, such as surgical phase detection or industrial process monitoring, where adjacent frames are visually similar.","If projection-level invariance systematically erases boundary cues, future video SSL methods for localization may need to default to intermediate-feature matching rather than final-head matching.","The need for a short supervised warm-up suggests a broader principle: when the semantic space of a short clip is narrow, pure self-distillation collapses distinctions that later task labels require.","Dataset-specific optimal clip length (Tennis versus figure skating) implies temporal context length should be treated as a first-class hyperparameter when porting the method to new sports or camera setups."],"forward_implications":["Semi-supervised PES can match or exceed fully supervised 100 percent baselines using only 80 percent labeled frames on some sports datasets.","Ordinary Vision Transformers become competitive for PES once multi-scale gated temporal shifting supplies motion information without a dedicated video architecture.","Motion-aware temporal augmentations that emphasize moving objects improve consistency training more than standard image augmentations in sports video settings.","Low-label regimes (around 10 percent annotations) benefit most, with multi-point mAP gains over prior SSL and supervised methods.","Pure unsupervised pretraining is insufficient when small event cues occupy only a few pixels; a supervised warm-up is required first."],"fun_headline_variants":["Temporal distillation lifts PES mAP 4.54 points at 10% labels","Backbone feature alignment keeps boundary cues DINO erases","Semi-supervised PES matches full labels at 80% data","Temporal Feature Distillation preserves motion for frame spotting","Warm-up plus backbone alignment beats projection distillation"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method rests on the premise that matching backbone features after a brief supervised warm-up will keep the small motion cues that projection-head methods erase, without proving those features are intrinsically more boundary-sensitive beyond the four sports datasets tested.","fun_headline_variants_meta":{"raw":{"variants":["Temporal distillation lifts PES mAP 4.54 points at 10% labels","Backbone feature alignment keeps boundary cues DINO erases","Semi-supervised PES matches full labels at 80% data","Temporal Feature Distillation preserves motion for frame spotting","Warm-up plus backbone alignment beats projection distillation"]},"model":"grok-4.5","effort":"low","cost_usd":0.003258,"raw_usage":{"total_tokens":1093,"prompt_tokens":777,"num_sources_used":0,"completion_tokens":66,"cost_in_usd_ticks":32580000,"prompt_tokens_details":{"text_tokens":777,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":250,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":777,"tokens_out":66,"duration_ms":3254,"temperature":1.0,"reasoning_tokens":250,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T07:45:13.617027+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the same TGS-ViT backbone with pure DINO-style projection alignment (no backbone-feature loss, no warm-up) under 10 percent labels on the four datasets; if that projection-only model matches or exceeds TFD mAP and still shows sharp inter-frame similarity drops at ground-truth event boundaries, the claim that backbone alignment is required fails.","supporting_citations":[],"review_version":1}