Pith. sign in

REVIEW 4 major objections 5 minor 37 references

By keeping each player-role slot as a separate entity throughout encoding—temporal attention within a slot, spatial attention across slots—ME-DST raises player-centric ball action spotting F1 to 0.778 on the FOOTPASS validation set, 10.3 po

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

2026-08-04 22:41 UTC pith:E4RFA4DU

load-bearing objection A solid, well-ablated extension of DST for player-centric spotting, with a big F1 jump that is plausible but not yet cleanly attributed to the entity-preserving design. the 4 major comments →

arxiv 2608.01696 v1 pith:E4RFA4DU submitted 2026-08-03 cs.CV cs.AI

Entity-Aware Sequence Transduction for Player-Centric Ball Action Spotting

Ruifeng Wang , Di Yang , Jiangtao Wang This is my paper
classification cs.CV cs.AI
keywords player-centric ball action spottingsequence transductionmulti-entity encoderfactorized attentionrole embeddingssoccer video understandingFOOTPASS benchmark
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.

The paper argues that the strongest official baseline for player-centric ball action spotting—a Denoising Sequence Transduction model—hurts itself by flattening all 26 player-role slots into one frame-level vector before temporal modeling. ME-DST keeps each role slot as a distinct entity through encoding: temporal attention tracks each player's history, spatial attention exchanges information across players at each frame, and learnable role embeddings give each slot an explicit identity. On the FOOTPASS validation set, ME-DST reports Micro F1 0.778, beating the TAAD+DST baseline by 10.3 percentage points. The ablations attribute the largest single share of this gain to role embeddings, and show that the richer tactical features help only when the encoder preserves player-level structure. The paper's central claim is that explicit entity structure is a useful inductive bias for multi-agent sports event understanding.

Core claim

ME-DST claims that player-centric ball action spotting should be modeled as sequence transduction over distinct player-role entities rather than over a flattened frame representation. The encoder processes a tensor indexed by frame and role slot, applies temporal self-attention independently to each role trajectory, then spatial self-attention across role slots at each frame, and adds a learnable role embedding to every slot. This factorization is designed so within-player temporal evolution and cross-player context are modeled separately instead of being mixed in one flat sequence. The paper reports Micro F1 of 0.778 on FOOTPASS validation, 10.3 percentage points above the official TAAD+DST

What carries the argument

The load-bearing component is the factorized multi-entity encoder. It keeps a T×26×22 entity tensor (frames × role slots × features), applies temporal attention separately to each of the 26 role-slot trajectories, then spatial attention across role slots at each frame, with a learnable role embedding added to each slot before attention. Temporal attention preserves per-player history; spatial attention lets players exchange context; the role embedding provides stable identity that the flattened DST encoder lacks.

Load-bearing premise

The evaluation assumes that predicted role slots can be reliably converted to jersey numbers by interpolating shirt-number annotations across frames, which the paper itself says can be unreliable near substitutions, tactical changes, or long gaps without a visible shirt number.

What would settle it

Run ME-DST on the FOOTPASS validation set with ground-truth jersey identities supplied instead of the interpolated role-to-jersey mapping; if Micro F1 stays near 0.778, the gains are robust to identity mapping, and if it drops substantially, the reported improvement partly reflects annotation propagation rather than entity modeling.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Preserving the role-slot axis through encoding is itself worth 10.3 F1 points over the official TAAD+DST baseline on FOOTPASS validation.
  • Role identity, not spatial interaction, is the dominant component: dropping role embeddings costs 16.7 F1 points, while dropping spatial attention costs only 0.3.
  • Tactical feature richness only pays off when the encoder keeps player structure: 8-dim features raise ME-DST from 0.758 to 0.778 but slightly lower the flat encoder's score.
  • The optional global temporal refinement layer should be omitted; removing it improves F1 from 0.736 to 0.778.
  • Entity-aware encoding helps most where visual evidence is scarce: recall on ball-hidden events rises from 0.391 to 0.480, and Cross/Header/Block improve by 0.216/0.211/0.173 F1.

Where Pith is reading between the lines

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

  • If the role-to-jersey interpolation is unreliable, the reported 0.778 may overstate the model's true attribution accuracy on substitution-heavy or occluded stretches; a test with ground-truth identities would separate entity modeling from annotation propagation.
  • The same factorized temporal-then-spatial design may transfer to other multi-agent sequence transduction tasks where identities are stable, such as tactical forecasting in other team sports; the paper leaves this untested.
  • Because the visual logits come from separately trained TAAD branches, upstream detection errors are baked into the sequence model, so the entity encoder's contribution is measured on top of fixed perception; joint training is a natural next experiment.
  • A parameter-matched flat-versus-multi-entity comparison would sharpen the claim that structure, not capacity, drives the gain; the paper acknowledges its ablations do not fully control for parameter count.

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

4 major / 5 minor

Summary. The manuscript introduces ME-DST, an extension of the Denoising Sequence Transduction (DST) framework for player-centric ball action spotting on the FOOTPASS benchmark. The proposed encoder keeps the role-slot axis explicit, applies factorized temporal and spatial attention, adds learnable role embeddings and tracking-derived tactical features, and fuses X3D-L and Swin3D-S visual logits. On the three-match FOOTPASS validation set, ME-DST achieves Micro F1 0.778, improving over the official TAAD+DST baseline (0.675) by 10.3 percentage points. Ablation studies attribute large gains to role embeddings and to removing a global refinement layer, and per-class analysis shows broad improvements across the eight action classes.

Significance. If the central attribution holds, the paper provides a useful and timely insight: preserving explicit entity structure in sequence transduction improves actor-attributed action spotting. The manuscript is generally clear, includes code availability, and reports component ablations and per-class numbers. Its main weakness is that the headline claims about the source of the gain are not yet fully supported by controlled comparisons, and the evaluation rests on a single small validation set with no error bars. These issues are acknowledged in the Limitations, but they are load-bearing for the conclusions rather than incidental.

major comments (4)
  1. [Results, Table 4; Discussion/Limitations] The claim that preserving the entity axis is central is confounded by differences in model capacity. The multi-entity encoder adds per-role projections, spatial attention, and a different representation shape relative to the flat DST encoder, but no parameter counts are reported. The manuscript itself states that the ablations 'do not fully disentangle architectural changes from differences in model capacity' and calls for parameter-matched comparisons. Because the +7.0 F1 gap between flat DST (0.708) and multi-entity (0.778) in Table 4 is central support for the main claim, the missing parameter-matched control is load-bearing. Please report parameter counts and compare against a flat DST encoder matched in capacity, or alternatively soften the attribution claim substantially.
  2. [Results, Table 5] The role-embedding ablation (0.736 to 0.569) is performed in the Base configuration that still includes global temporal refinement, whereas the final best model is the no-refinement configuration (0.778). Since removing global refinement changes F1 by +4.2 points, its interaction with role embeddings is unknown. The abstract and conclusion attribute the gain to role identity, yet there is no direct ablation of role embeddings in the deployed configuration. Please report the '– role embeddings' ablation with global refinement removed, preferably for both block depths, and discuss any interaction.
  3. [Experimental Setup; Ablations] The best configuration (no global refinement) was selected using the same three-match FOOTPASS validation set on which all results are reported, and no multiple seeds or error bars are given. With only three validation matches and several hyperparameters (fusion weight, smoothing sigma, number of blocks, refinement choice, confidence threshold) selected on that set, the headline +10.3 improvement may be optimistic and could reflect selection effects. Please provide standard deviations across multiple seeds, or a separate development/test split, and a sensitivity analysis for the main hyperparameters.
  4. [Inference; Limitations] The evaluation protocol requires the correct jersey number for a true positive, but the predicted role index is converted to a jersey number using shirt-number interpolation that the authors describe as 'unreliable near substitutions, changes in tactical assignment, or extended periods without a visible shirt number.' This mapping is load-bearing for the absolute F1 values. Please quantify the reliability of the role-to-jersey conversion (e.g., conversion accuracy on substituted vs. non-substituted events) and report the sensitivity of the final F1 to mapping errors. This issue affects baselines too, but it is still essential for interpreting the reported numbers.
minor comments (5)
  1. [Methods, Tactical Feature Representation] The text contains an unresolved cross-reference: 'eight engineered tactical descriptors (Section )' with no section number. Please fix.
  2. [Methods, Eq. (1)] Please define Z_X3D and Z_Swin explicitly and clarify how the 9-dimensional per-role logits (eight actions plus background) are reduced or used in the 22-dimensional input vector.
  3. [Results, Table 1 vs. Table 6] Table 1 compares against the official TAAD+DST baseline, while Table 6 compares against a reproduced X3D-L+DST implementation. Please state more prominently that the per-class gains in Table 6 are relative to the reproduced baseline, not the official one, to avoid ambiguity.
  4. [Results, Table 5] For the '1 block + spatial + role' rows, it is not explicitly stated whether the no-refine variant also uses the same spatial attention and role embeddings. Please make the configuration descriptors unambiguous.
  5. [Figure 2] The caption says percentages are row-normalized but also describes a rightmost missed-events column and a bottom unmatched-predictions row. Clarify how the row-normalization is computed for these cells and whether the bottom row is column-normalized.

Circularity Check

0 steps flagged

No circular derivation; empirical evaluation with acknowledged confounds.

full rationale

ME-DST is an empirical systems paper. The central claim (0.778 Micro F1, +10.3 pp over TAAD+DST) is established by training and evaluating on the FOOTPASS validation set under the official protocol; it is not derived from the model definition. The ablations compare model variants by measured F1, not by algebraically forcing the outcome. No parameter is fitted to the reported metric and then renamed as a prediction: role embeddings, tactical features, and fused logits are model inputs, while the predicted action, frame, and role come from the trained decoder. There are no load-bearing self-citations: the cited baselines and backbones are external prior work by other groups. The paper explicitly acknowledges that the ablations 'do not fully disentangle architectural changes from differences in model capacity' and calls for parameter-matched comparisons; that is an internal-validity limitation, not circular reasoning. The shirt-number interpolation caveat affects evaluation fidelity for both systems and is a data-quality concern, not a definitional equivalence. No equation in the paper reduces the reported result to its inputs by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim rests on an empirical evaluation with several hand-chosen hyperparameters and dataset assumptions. No new physical or conceptual entities are introduced. The most important free choices are the fusion weighting, Gaussian smoothing, block count, and the validation-driven removal of the global refinement layer.

free parameters (6)
  • Fusion weight alpha = 0.6 for X3D, 0.4 for Swin (Eq. 1)
    Weighted average of visual logits is chosen without reported sensitivity analysis; it affects all downstream results.
  • Gaussian smoothing sigma = 1.5
    Applied to Swin3D-S logits before fusion; value hand-set, not varied.
  • Number of factorized blocks = 2
    Selected via ablation (1 vs 2 blocks) on the validation set.
  • Global temporal refinement = removed
    Removing the layer raises validation F1 from 0.736 to 0.778; this selection is made on the same validation set used for reporting.
  • Tactical feature dimensionality = 8-dim
    Eight handcrafted features chosen over the 3-dim variant; only helps in the multi-entity encoder.
  • Confidence threshold = 0.15
    Official evaluation protocol threshold, not a free choice, but it heavily influences the F1 numbers.
axioms (5)
  • domain assumption FOOTPASS ground-truth annotations (frame, team, jersey, class) are correct
    The evaluation treats the dataset labels as ground truth; any errors propagate directly into reported F1.
  • domain assumption Role-slot to jersey mapping via shirt-number interpolation is valid
    The model predicts a role index, and evaluation converts it to jersey number using interpolated shirt numbers. The paper acknowledges this can fail near substitutions or occlusion.
  • domain assumption TAAD visual logits are a sufficient representation of the video
    The sequence model never sees raw video, only precomputed logits from X3D-L and Swin3D-S branches; errors in those branches are inherited.
  • domain assumption The 3-match validation set is representative of test performance
    All conclusions and model selections are based on three validation matches; no test set results are reported.
  • standard math Standard Transformer attention and Pre-LN are used as described
    The method relies on standard multi-head attention, layer norm, and positional encodings from the cited literature.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Entity-Aware Sequence Transduction for Player-Centric Ball Action Spotting." pith.science (2026). https://pith.science/paper/E4RFA4DU

@misc{pith2026260801696,
  author       = {Pith},
  title        = {Pith review of: Entity-Aware Sequence Transduction for Player-Centric Ball Action Spotting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E4RFA4DU}},
  note         = {Machine review of arXiv:2608.01696}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Player-centric ball action spotting requires temporally precise event detection together with actor attribution in crowded, partially observed multi-agent sports videos. Existing Denoising Sequence Transduction (DST) baselines treat the player-role dimension as part of a flattened frame-level representation, which weakens the inductive bias for modeling player-specific temporal evolution and inter-player interactions. To address this limitation, we propose Multi-Entity Denoising Sequence Transduction (ME-DST). ME-DST keeps the role-slot dimension throughout encoding. It uses temporal attention to model the history of each role slot, and spatial attention to exchange information across role slots at each frame. This factorized design gives the model a direct structure for separating within-player evolution from inter-player context. We also add learnable role embeddings, tracking-derived tactical features, and fused visual predictions from X3D-L and Swin3D-S. Experiments on the FOOTPASS dataset show that ME-DST reaches a Micro F1 of 0.778. This improves the strongest official TAAD+DST baseline by 10.3 percentage points. Controlled ablations show that preserving the entity axis and encoding role identity are central to this gain. These results suggest that explicit entity modeling is an effective inductive bias for player-centric sports event understanding.

Figures

Figures reproduced from arXiv: 2608.01696 by Di Yang, Jiangtao Wang, Ruifeng Wang.

Figure 1
Figure 1. Figure 1: Overview of ME-DST. (1) Fused TAAD logits from X3D-L and Swin3D-S are concatenated with tracking and tactical features to form the entity tensor X ∈ R T ×26×22. (2) Role- and time-aware entity embeddings are processed by factorized temporal and spatial attention and pooled into the memory M ∈ R T ×512. (3) An autoregressive decoder predicts the action, frame, and role of each player-centric event. The main… view at source ↗
Figure 2
Figure 2. Figure 2: Event-level confusion matrices for the reproduced X3D-L+DST baseline and ME-DST. Rows denote ground-truth action classes and columns denote predicted action classes. Diagonal cells are offi￾cial class-aware true positives. Off-diagonal cells show residual player-aware temporal matches with an incorrect predicted class. The rightmost column reports missed ground-truth events, and the bottom row reports pred… view at source ↗
Figure 3
Figure 3. Figure 3: Baseline errors corrected by ME-DST. Yellow dashed boxes mark the annotated actor; green, red, and blue boxes show the ground truth, reproduced X3D-L+DST baseline, and ME-DST results, respectively. (A) A Header is corrected from pass by ME-DST. (B) A Cross mistaken as a Shot is corrected (C) A rare Block is discovered by ME-DST. (D) A Throw-in by the line is corrected from pass. flat sequence formulation a… 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

37 extracted references · 33 canonical work pages · 1 internal anchor

  1. [1]

    & Ghanem, B

    Giancola, S., Amine, M., Dghaily, T. & Ghanem, B. SoccerNet: A scalable dataset for action spotting in soccer videos. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops1711–1721 (2018)

    work page 2018

  2. [2]

    Cioppa, A.et al.SoccerNet 2023 challenges results.Sports Eng.27, 24; 10.1007/s12283- 024-00466-4 (2024)

    work page doi:10.1007/s12283- 2023

  3. [3]

    Deliège, A.et al.SoccerNet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Work- shops(2021)

    work page 2021

  4. [4]

    & Ghanem, B

    Giancola, S. & Ghanem, B. Temporally-aware feature pooling for action spotting in soccer broadcasts. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops4490– 4499 (2021)

    work page 2021

  5. [5]

    & Sivic, J

    Arandjelović, R., Gronat, P., Torii, A., Pajdla, T. & Sivic, J. NetVLAD: CNN architecture forweaklysupervisedplacerecognition.InProc. IEEE Conf. Comput. Vis. Pattern Recognit. 5297–5307 (2016)

    work page 2016

  6. [6]

    Cioppa, A.et al.A context-aware loss function for action spotting in soccer videos. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit.(2020)

    work page 2020

  7. [7]

    & Fatahalian, K

    Hong, J., Zhang, H., Gharbi, M., Fisher, M. & Fatahalian, K. Spotting temporally precise, fine-grained events in video. InProc. Eur. Conf. Comput. Vis.33–51 (2022)

    work page 2022

  8. [8]

    Xarles, A., Escalera, S., Moeslund, T. B. & Clapés, A. ASTRA: An action spotting TRAns- formerforsoccervideos.InProc. 6th Int. Workshop Multimedia Content Anal. Sports93–102 (2023). 15

    work page 2023

  9. [9]

    & Hérault, R

    Denize, J., Liashuha, M., Rabarisoa, J., Orcesi, A. & Hérault, R. COMEDIAN: Self- supervised learning and knowledge distillation for action spotting using transformers. In Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis. Workshops530–540 (2024)

    work page 2024

  10. [10]

    Soares, J. V. B., Shah, A. & Biswas, T. Temporally precise action spotting in soccer videos using dense detection anchors. InProc. IEEE Int. Conf. Image Process.2796–2800; 10.1109/ICIP46576.2022.9897256 (2022)

    arXiv 2022

  11. [11]

    Giancola, S.et al.SoccerNet 2022 challenges results. InProc. 5th Int. ACM Workshop Multimedia Content Anal. Sports75–86; 10.1145/3552437.3558545 (2022)

    work page arXiv 2022

  12. [12]

    Tan, M. & Le, Q. V. EfficientNetV2: Smaller models and faster training. InProc. 38th Int. Conf. Mach. Learn.10096–10106 (2021)

    work page 2021

  13. [13]

    A Boosted Model Ensembling Approach to Ball Action Spotting in Videos: The Runner-Up Solution to CVPR'23 SoccerNet Challenge

    Wang, L., Guo, H. & Liu, B. A boosted model ensembling approach to ball action spotting in videos: The runner-up solution to the CVPR 2023 SoccerNet Challenge.arXiv2306.05772 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  14. [14]

    Cioppa, A.et al.SoccerNet 2024 challenges results.arXiv2409.10587 (2024)

    Pith/arXiv arXiv 2024

  15. [15]

    Xarles, A., Escalera, S., Moeslund, T. B. & Clapés, A. T-DEED: Temporal-discriminability enhancer encoder-decoder for precise event spotting in sports videos. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops(2024)

    work page 2024

  16. [16]

    Giancola, S.et al.SoccerNet 2025 challenges results.arXiv2508.19182 (2025)

    Pith/arXiv arXiv 2025

  17. [17]

    & Manitsaris, S

    Ochin, J., Chekroun, R., Stanciulescu, B. & Manitsaris, S. FOOTPASS: A multi-modal multi-agent tactical context dataset for play-by-play action spotting in soccer broadcast videos.arXiv2511.16183 (2025)

    work page arXiv 2025

  18. [18]

    & Van Gool, L

    Singh, G., Choutas, V., Saha, S., Yu, F. & Van Gool, L. Spatio-temporal action detec- tion under large motion. InProc. IEEE/CVF Winter Conf. Appl. Comput. Vis.6009–6018 (2023)

    work page 2023

  19. [19]

    & Manitsaris, S

    Ochin, J., Devineau, G., Stanciulescu, B. & Manitsaris, S. Game state and spatio-temporal action detection in soccer using graph neural networks and 3D convolutional networks. In Proc. 14th Int. Conf. Pattern Recognit. Appl. Methods636–646 (2025)

    work page 2025

  20. [20]

    & Manitsaris, S

    Ochin, J., Chekroun, R., Stanciulescu, B. & Manitsaris, S. Beyond pixels: Leveraging the language of soccer to improve spatio-temporal action detection in broadcast videos. InAd- vanced Concepts for Intelligent Vision Systems: 22nd International Conference, ACIVS 2025552–563; 10.1007/978-3-032-07343-3_44 (Springer, 2026)

    work page doi:10.1007/978-3-032-07343-3_44 2026

  21. [21]

    X3D: Expanding architectures for efficient video recognition

    Feichtenhofer, C. X3D: Expanding architectures for efficient video recognition. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit.200–210 (2020)

    work page 2020

  22. [22]

    Liu, Z.et al.Video Swin Transformer. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit.3202–3211 (2022)

    work page 2022

  23. [23]

    Neural Inf

    Vaswani, A.et al.Attention is all you need.Adv. Neural Inf. Process. Syst.30(2017)

    work page 2017

  24. [24]

    Cho, K.et al.Learning phrase representations using RNN encoder–decoder for statistical machine translation. InProc. 2014 Conf. Empir. Methods Nat. Lang. Process.1724–1734 (2014)

    work page 2014

  25. [25]

    InSports Analytics: Third International Con- ference, ISACE 2026170–187 (Springer, 2026)

    Song, K.et al.Decoding defensive coverage responsibilities in American football using fac- torized attention-based transformer models. InSports Analytics: Third International Con- ference, ISACE 2026170–187 (Springer, 2026). 16

    work page 2026

  26. [26]

    & Madeleine, P

    Ibh, M., Grasshof, S., Witzner, D. & Madeleine, P. TemPose: A new skeleton-based transformer model designed for fine-grained motion recognition in badminton. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops(2023)

    work page 2023

  27. [27]

    & Snoek, C

    Gavrilyuk, K., Sanford, R., Javan, M. & Snoek, C. G. M. Actor-transformers for group activity recognition. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit.839–848 (2020)

    work page 2020

  28. [28]

    Li, S.et al.GroupFormer: Group activity recognition with clustered spatial-temporal trans- former. InProc. IEEE/CVF Int. Conf. Comput. Vis.13668–13677 (2021)

    work page 2021

  29. [29]

    & Kitani, K

    Yuan, Y., Weng, X., Ou, Y. & Kitani, K. M. AgentFormer: Agent-aware transformers for socio-temporal multi-agent forecasting. InProc. IEEE/CVF Int. Conf. Comput. Vis. 9813–9823 (2021)

    work page 2021

  30. [30]

    & Claudel, C

    Mohamed, A., Qian, K., Elhoseiny, M. & Claudel, C. Social-STGCNN: A social spatio- temporal graph convolutional neural network for human trajectory prediction. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit.14424–14432 (2020)

    work page 2020

  31. [31]

    & Balasubramanian, V

    Santra, S., Chudasama, V., Wasnik, P. & Balasubramanian, V. N. Precise event spotting in sports videos: Solving long-range dependency and class imbalance. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit.3163–3172 (2025)

    work page 2025

  32. [32]

    Cabado, B.et al.Beyond the Premier: Assessing action spotting transfer capability across diverse domains. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops 3386–3398 (2024)

    work page 2024

  33. [33]

    Dalal, A.et al.Action anticipation from SoccerNet football video broadcasts. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. Workshops(2025)

    work page 2025

  34. [34]

    Xiong, R.et al.On layer normalization in the Transformer architecture. InProc. Int. Conf. Mach. Learn.(2020)

    work page 2020

  35. [35]

    & Wojna, Z

    Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J. & Wojna, Z. Rethinking the Inception ar- chitecture for computer vision. InProc. IEEE Conf. Comput. Vis. Pattern Recognit.(2016)

    work page 2016

  36. [36]

    Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. InProc. 3rd Int. Conf. Learn. Represent.(2015)

    work page 2015

  37. [37]

    & Guestrin, C

    Chen, T., Xu, B., Zhang, C. & Guestrin, C. Training deep nets with sublinear memory cost. arXiv1604.06174 (2016). 17

    Pith/arXiv arXiv 2016

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.