Pith. sign in

REVIEW 3 major objections 4 minor 54 references

SoccerHigh: A Benchmark Dataset for Automatic Soccer Video Summarization

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SoccerHigh introduces the first public benchmark for soccer video summarization, pairing 237 full matches with official league summaries aligned shot-by-shot.

desk verdict The dataset is a genuine and useful contribution; treat the baseline numbers as indicative, not authoritative, and the gold standard needs reliability evidence. read the letter →

arxiv 2509.01439 v1 pith:NMNEKASF submitted 2025-09-01 cs.CV cs.AIcs.MM

Artur Díaz-Juan , Coloma Ballester , Gloria Haro This is my paper
classification cs.CVcs.AIcs.MM
keywords soccervideosummarizationsportsunderstandingdatasetbenchmarkshotalignmenthighlightgenerationTransformerbaselineF1@TNet
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper presents SoccerHigh, a public dataset for soccer video summarization: 237 full matches from Spain, France, and Italy, each paired with an official league-produced highlight summary and aligned to the corresponding broadcast video at shot level. The authors claim this is the first public dataset built specifically for the task of soccer video summarization. They support the claim with a semi-automated annotation pipeline that reduces manual effort, a baseline model reaching an F1 of 0.3956, and a new length-constrained metric, F1@T, that evaluates summaries without forcing a fixed output duration. If the benchmark holds, researchers gain a reproducible way to train and compare models that turn broadcast footage into coherent highlight reels capturing both game events and editorial context.

What carries the argument

The dataset's construction rests on a two-stage alignment pipeline: shot boundary detection that intentionally oversegments, followed by retrieval of the best-matching broadcast segment for each summary shot using temporally averaged frame features, with DINOv2 yielding the best alignment F1 of 0.8407. The baseline model combines a frozen pretrained backbone (VideoMAEv2 giant), a two-layer Transformer encoder over 60-second chunks, and an anchor-free head that predicts per-frame classification scores plus boundary offsets, with Non-Maximum Suppression used at inference. Evaluation uses F1@T: predicted shots are ranked by confidence and selected until the cumulative duration matches the groun

What would settle it

Take a random sample of, say, 20 matches, have two independent annotators redo the manual refinement stage, and compute pairwise IoU between the resulting alignments; low agreement (e.g., mean IoU below 0.7) would show the ground truth is not a stable gold standard. A second check: remove the manual refinement and train/evaluate the baseline on purely automatic alignments; a large performance drop would indicate the benchmark's numbers depend on the authors' subjective corrections.

Watch

Extended reading notes

Core claim

The central claim is that a usable, public, shot-level-aligned dataset of soccer matches and their official summaries can be constructed with a semi-automated pipeline: segment the summary into shots, retrieve the best-matching segment of the full broadcast using frame-level features, then manually refine the correspondences. With DINOv2 features and a kNN-based oversegmentation, the automatic alignment reaches an F1 of 0.8407 and finds 92.5% of relevant shots at a low overlap threshold, leaving a small manual correction burden. The paper further argues that a Transformer-based baseline with a classification-and-regression head, trained on 60-second chunks with VideoMAEv2 features and MixUp

Load-bearing premise

The manually refined shot-to-broadcast alignments, produced by the authors without inter-annotator agreement testing, are treated as correct ground truth; if these alignments are noisy or systematically biased, all reported accuracies and the benchmark's label quality are affected.

Editorial extensions

If this is right

  • Any team can now train or fine-tune a soccer summarization model on 237 match-summary pairs with fixed train/validation/test splits and compare against the baseline's F1 of 0.3956.
  • The alignment pipeline, with 92.5% recall at coarse overlap, makes it practical to expand the dataset to more leagues and seasons without full manual annotation.
  • The F1@T metric removes the need to force models to output a fixed percentage of the match, aligning evaluation with how editors actually vary summary length.
  • Because summaries span pre-match, half-time, and post-match content, the benchmark captures editorial choice, not just goal detection, pushing models toward narrative coherence.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next step is to test whether the alignment pipeline transfers to other team sports or to user-generated highlight clips, where shot structure is less regular.
  • The benchmark's ground truth is a single editorial summary per match; using it for training may bake in each league's editorial biases, and multi-annotator or multi-edition summaries would be a stronger gold standard.
  • The F1@T protocol could be adopted by general video summarization research as a way to compare models whose summary lengths differ from the reference without imposing a fixed ratio.
  • Since the authors discarded out-of-context summary shots as annotation noise, the dataset's definition of a 'good' summary itself encodes their human judgment; a model trained on it will inherit that judgment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This paper introduces SoccerHigh, a dataset for soccer video summarization consisting of 237 matches from La Liga, Ligue 1, and Serie A, paired with official league summaries aligned at shot level to SoccerNet broadcast videos. The paper also proposes a semi-automated annotation pipeline (shot boundary detection, feature-based alignment, and manual refinement), a baseline model using frozen frame features, a Transformer encoder, classification and boundary-regression heads with NMS, and a length-constrained evaluation metric F1@T. The baseline achieves a test F1 of 0.3956 and an F1@T of 0.3883. The central claim is that this is the first public dataset for soccer video summarization.

Significance. If the dataset and its annotations are reliable, SoccerHigh would fill a real gap: it is large, publicly available, reuses SoccerNet's standard splits, and is built from professionally edited official summaries. The paper ships several concrete assets: the dataset and code, a quantitative analysis of the automatic alignment pipeline (Tables 3 and 4), and a simple baseline with ablations over backbone, chunk size, heads, and augmentation. The proposed F1@T metric is a reasonable attempt to decouple summary content selection from length constraints. However, the benchmark's gold standard rests on an unmeasured manual refinement step, and the baseline numbers are single-run point estimates. These issues must be addressed before the dataset can serve as a stable reference benchmark.

major comments (3)
  1. [Section 3.2; Tables 3 and 4] The gold-standard labels are produced by the authors' manual refinement step, but the paper reports no inter-annotator agreement, no number of annotators, and no reliability analysis. The text states that scenes are aligned by 'human judgment' and that out-of-context shots are 'discarded,' which are subjective editorial decisions. This is load-bearing because every downstream metric (pipeline F1 in Table 3, shot-level recall in Table 4, baseline F1 in Tables 5-8, and F1@T in Table 9) is measured against this target. Table 3's high DINOv2 F1 of 0.8407 only shows that the automatic alignment approximates the authors' own final labels; it does not establish that the labels are reproducible or unbiased. Please add a double-annotation study on a representative subset (e.g., 20-30 matches), report agreement in terms of interval-level IoU or kappa, specify annotation instructions and the number
  2. [Section 4.2; Tables 5-9] All baseline results are single-run point estimates with no standard deviations, seeds, or significance tests. For example, the VideoMAEv2-small vs. giant difference (0.3428 vs. 0.3956 in Table 5), the chunk-size peak at 60 s in Table 6, and the MixUp gain of +0.0301 in Table 8 could easily be within run-to-run variability. Since the paper presents this as a benchmark reference, the lack of variance reporting undermines confidence in the recommended configuration and in comparisons by future methods. Please report mean +/- std over at least 3-5 training seeds and, where relevant, paired differences.
  3. [Eq. (11); Section 4.1/4.2] The regression loss weight lambda is never specified. The implementation details list learning rate, beta1/beta2, weight decay, batch size, patience, and MixUp parameters, but not the value of lambda in Eq. (11). This is not a cosmetic omission: Table 7 shows that the regression head plus NMS changes F1 from 0.3493 to 0.3956, so the contribution of L_reg depends directly on lambda. Please state the value used, and ideally include an ablation over lambda (e.g., 0, 0.1, 1, 10) to show sensitivity.
minor comments (4)
  1. [Eq. (9)-(10)] The definition of M is inconsistent with the surrounding text. Eq. (10) sums the indicator over c=1,2, but for every frame inside a ground-truth summary segment both Y_start and Y_end are non-negative, so M is twice the number of positive frames. If M is intended to be the number of positive frames, the sum over c should be replaced by an indicator that a frame has at least one valid offset, or the text should say 'number of positive frame-channel pairs.' This does not change the training direction but affects the loss scale and should be corrected.
  2. [Section 5] The F1@T metric is underspecified. Please clarify what happens when the ranked predicted proposals have a total duration shorter than the ground-truth length T, how ties in the ranking are broken, and whether the selected proposals are allowed to overlap or are first passed through NMS. Also state whether the duration constraint is applied per game or globally and how the F1@T is computed at the shot level (the current text says 'per game and macro-averaged' but does not define the unit of matching).
  3. [Section 4.2] The text says 'increasing the duration of the chunk from 10 to 60 seconds,' but Table 6 starts at 15 seconds. Please correct the lower bound to 15 seconds or add the missing 10-second result.
  4. [Table 2] The dataset is heavily imbalanced across leagues and seasons: for example, France contributes only 1 training game in 2014/15 and 2 in 2015/16, while Spain 2016/17 contributes 42. This is acknowledged in the text, but it would be useful to state explicitly in the dataset documentation whether the provided splits are stratified by league and season, since the paper motivates the three-league design by editorial-style diversity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dataset, baseline, and metric are self-contained and do not reduce to fitted inputs or self-citation chains.

full rationale

The paper's central claims are a new dataset, a baseline model, and an evaluation metric. The baseline is trained on the training split and evaluated on the held-out test split, which is standard supervised benchmarking, not circular. The proposed F1@T metric uses the ground-truth summary length as an evaluation constraint, but this is a protocol choice: predicted shots are ranked by model confidence and truncated to the reference length, with no parameter fitted to the test labels. The annotation pipeline is assessed by comparing the automatic pre-refinement alignment to the final manually refined correspondences; this is an external human annotation stage, so the evaluation measures the utility of the automatic proposals rather than defining them as ground truth. No equation in the paper defines a predicted quantity in terms of its own target. The only self-citations ([5,6], same group) are used to motivate a chunk-based architecture, not as load-bearing evidence for the paper's conclusions. The concern raised about the manual refinement step (no inter-annotator agreement) is a validity/reliability limitation, not a circularity: the ground truth is asserted as human-corrected, and downstream numbers inherit any noise in those labels, but this does not make the derivation equivalent to its inputs. Therefore no specific circular step can be identified, and the paper receives a score of 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities. Its free parameters are standard training and annotation hyperparameters, several of which are underexplored or unspecified. The domain assumptions are the typical transfer-learning and ground-truth assumptions of benchmark papers, with the manual refinement assumption being the most fragile.

free parameters (7)
  • kNN k = 10
    Number of nearest neighbors in the custom shot segmentation; hand-chosen and affects the oversegmentation of summary shots (Section 3.1).
  • TransNetv2 decision thresholds = 0.5 and 0.05
    Two thresholds tested for shot boundary detection; the 0.05 value is chosen to induce oversegmentation (Section 3.1, Table 3).
  • NMS confidence threshold = 0.5
    Frames with classification score above 0.5 are kept before Non-Maximum Suppression during inference (Section 4.1).
  • Chunk size = 60 seconds
    Best chunk length selected by ablation on validation; directly controls the context window for the baseline (Section 4.2, Table 6).
  • Regression loss weight lambda = not reported
    Weight of the regression loss in Eq. (11); its value is never given, yet it changes the balance between classification and boundary regression.
  • MixUp alpha and beta = 0.3 and 0.3
    Parameters of the beta distribution used for MixUp augmentation on the training set (Section 4.1).
  • Learning rate and weight decay = 5e-5 and 0.05
    AdamW optimizer settings; hand-selected and standard for small Transformer training runs (Section 4.1).
assumptions (5)
  • domain assumption Official league summaries are a valid ground truth for video summarization.
    The dataset treats curated editorial summaries as the target output, but summaries encode subjective editorial choices rather than an objective gold standard (Section 3.2).
  • ad hoc to paper Manual refinement by the authors produces accurate alignments without measurable inter-annotator agreement.
    The final ground truth is produced by the authors' manual refinement, and no reliability statistics are reported (Sections 3.2 and 3.3).
  • domain assumption Temporally averaged visual features are sufficient to align summary shots to broadcast segments.
    Eq. (3) uses Euclidean distance between averaged features; no learned verification is applied beyond manual correction (Section 3.1).
  • domain assumption Frozen pretrained backbones generalize to soccer broadcast video without domain adaptation.
    DINOv2 and VideoMAEv2 features are extracted with frozen weights and assumed to carry soccer-relevant semantics (Sections 3.1 and 4.2).
  • domain assumption SoccerNet broadcast videos and retrieved official summaries correspond to the same match.
    The pairing assumes exact match identity between the SoccerNet video and the online official summary (Section 3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of SoccerHigh: A Benchmark Dataset for Automatic Soccer Video Summarization." pith.science (2026). https://pith.science/paper/NMNEKASF

@misc{pith2026250901439,
  author       = {Pith},
  title        = {Pith review of: SoccerHigh: A Benchmark Dataset for Automatic Soccer Video Summarization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NMNEKASF}},
  note         = {Machine review of arXiv:2509.01439}
}
read the original abstract

Video summarization aims to extract key shots from longer videos to produce concise and informative summaries. One of its most common applications is in sports, where highlight reels capture the most important moments of a game, along with notable reactions and specific contextual events. Automatic summary generation can support video editors in the sports media industry by reducing the time and effort required to identify key segments. However, the lack of publicly available datasets poses a challenge in developing robust models for sports highlight generation. In this paper, we address this gap by introducing a curated dataset for soccer video summarization, designed to serve as a benchmark for the task. The dataset includes shot boundaries for 237 matches from the Spanish, French, and Italian leagues, using broadcast footage sourced from the SoccerNet dataset. Alongside the dataset, we propose a baseline model specifically designed for this task, which achieves an F1 score of 0.3956 in the test set. Furthermore, we propose a new metric constrained by the length of each target summary, enabling a more objective evaluation of the generated content. The dataset and code are available at https://ipcv.github.io/SoccerHigh/.

Figures

Figures reproduced from arXiv: 2509.01439 by the authors.

Figure 1
Figure 1. Baseline model overview. Chunks of 𝑁 frames are extracted from the full-match broadcast (a) and are first passed through a feature extraction stage (b), where each frame is represented by a feature vector and projected into a shared embedding space. Sinusoidal positional encoding (p.e.) is then added before feeding the sequence into a Transformer encoder (c). The confidence score 𝑐 and the boundary offsets 𝛿start an… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 43 canonical work pages

  1. [1]

    Rockson Agyeman, Rafiq Muhammad, and Gyu Sang Choi. 2019. Soccer video summarization using deep learning. In 2019 IEEE Conference on multimedia infor- mation processing and retrieval (MIPR) . IEEE, 270–273

    work page 2019

  2. [2]

    Evlampios Apostolidis, Eleni Adamantidou, Alexandros I Metsai, Vasileios Mezaris, and Ioannis Patras. 2020. AC-SUM-GAN: Connecting actor-critic and generative adversarial networks for unsupervised video summarization. IEEE Transactions on Circuits and Systems for Video Technology 31, 8 (2020), 3278–3292

    work page 2020

  3. [3]

    Evlampios Apostolidis, Georgios Balaouras, Vasileios Mezaris, and Ioannis Patras

    work page

  4. [4]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision. 9650–9660

    2021

  5. [5]

    Alejandro Cartas, Coloma Ballester, and Gloria Haro. 2022. A graph-based method for soccer action spotting using unsupervised player classification. InProceedings of the 5th International ACM Workshop on Multimedia Content Analysis in Sports . 93–102

    work page 2022

  6. [6]

    Alejandro Cartas, Coloma Ballester, and Gloria Haro. 2024. Two Weakly Super- vised Approaches for Role Classification of Soccer Players. In Proceedings of the 7th ACM International Workshop on Multimedia Content Analysis in Sports. 81–89

    work page 2024

  7. [7]

    Yaosen Chen, Bing Guo, Yan Shen, Renshuang Zhou, Weichen Lu, Wei Wang, Xuming Wen, and Xinhua Suo. 2022. Video summarization with u-shaped trans- former. Applied Intelligence 52, 15 (2022), 17864–17880

    work page 2022

  8. [8]

    Anthony Cioppa, Adrien Deliège, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. 2022. Scaling up SoccerNet with multi-view spatial localiza- tion and re-identification. Scientific data 9, 1 (2022), 355

    work page 2022

Show all 54 references
  1. [9]

    Adrien Deliege, Anthony Cioppa, Silvio Giancola, Meisam J Seikavandi, Jacob V Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B Moeslund, and Marc Van Droogenbroeck. 2021. Soccernet-v2: A dataset and benchmarks for holis- tic understanding of broadcast soccer videos. In Proc...

    2021

  2. [10]

    Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. 2023. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing...

    2023

  3. [11]

    Xin Gao, Xusheng Liu, Taotao Yang, Guilin Deng, Hao Peng, Qiaosong Zhang, Hai Li, and Junhui Liu. 2020. Automatic key moment extraction and highlights generation based on comprehensive soccer video understanding. In 2020 IEEE international conference on multimedia & expo works...

    2020

  4. [12]

    Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. 2018. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops . 1711– 1721

    2018

  5. [13]

    Yaowei Guo, Jiazheng Xing, Xiaojun Hou, Shuo Xin, Juntao Jiang, Demetri Ter- zopoulos, Chenfanfu Jiang, and Yong Liu. 2025. CFSum: A Transformer-Based Multi-Modal Video Summarization Framework With Coarse-Fine Fusion. In ICASSP 2025-2025 IEEE International Conference on Acoust...

    2025

  6. [14]

    Michael Gygli, Helmut Grabner, Hayko Riemenschneider, and Luc Van Gool

  7. [15]

    Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. 2023. Align and attend: Multimodal summarization with dual contrastive losses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 14867–14878

    2023

  8. [16]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

    2016

  9. [17]

    Zhong Ji, Kailin Xiong, Yanwei Pang, and Xuelong Li. 2019. Video summarization with attention-based encoder–decoder networks. IEEE Transactions on Circuits and Systems for Video Technology 30, 6 (2019), 1709–1717

    2019

  10. [18]

    Yudong Jiang, Kaixu Cui, Leilei Chen, Canjin Wang, and Changliang Xu. 2020. SoccerDB: A large-scale database for comprehensive video understanding. In Proceedings of the 3rd International Workshop on Multimedia Content Analysis in Sports. 1–8

    2020

  11. [19]

    Yinhui Jiang, Sihui Luo, Lijun Guo, and Rong Zhang. 2024. MCT-VHD: Multi- modal contrastive transformer for video highlight detection. Journal of Visual Communication and Image Representation 101 (2024), 104162

    2024

  12. [20]

    Yunjae Jung, Donghyeon Cho, Sanghyun Woo, and In So Kweon. 2020. Global- and-local relative position embedding for unsupervised video summarization. In European conference on computer vision . Springer, 167–183

    2020

  13. [21]

    J Lei, TL Berg, and M Bansal. 2021. Qvhighlights: Detecting moments and highlights in videos via natural language queries.(2021). URL https://arxiv. org/abs/2107.09609 (2021)

    2021 arXiv

  14. [22]

    Ping Li, Qinghao Ye, Luming Zhang, Li Yuan, Xianghua Xu, and Ling Shao

  15. [23]

    Jingxu Lin, Sheng-hua Zhong, and Ahmed Fares. 2022. Deep hierarchical LSTM networks with attention for video summarization. Computers & Electrical Engi- neering 97 (2022), 107618

    2022

  16. [24]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

    2017 arXiv

  17. [25]

    Pattern Recognition 111 (2021), 107677

    Exploring global diverse attention via pairwise temporal relation for video summarization. Pattern Recognition 111 (2021), 107677

    2021

  18. [26]

    Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. 2019. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision ....

    2019

  19. [27]

    Vinod Nair and Geoffrey E Hinton. 2010. Rectified linear units improve re- stricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10) . 807–814

    2010

  20. [28]

    George B Mathews. 1896. On the partition of numbers. Proceedings of the London Mathematical Society 1, 1 (1896), 486–490

  21. [29]

    Nielsen Sports. 2022. Fans Are Changing the Game. https://nielsensports.com/wp- content/uploads/2022/02/Nielsen-Sports-Fans-are-changing-the-game-1.pdf

    2022

  22. [30]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, et al. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    2023 arXiv

  23. [31]

    Medhini Narasimhan, Anna Rohrbach, and Trevor Darrell. 2021. Clip-it! language- guided video summarization. Advances in neural information processing systems 34 (2021), 13988–14000

    2021

  24. [32]

    Jiayuan Rao, Haoning Wu, Hao Jiang, Ya Zhang, Yanfeng Wang, and Weidi Xie. 2024. Towards universal Soccer video understanding. arXiv preprint arXiv:2412.01820 (2024)

    2024 arXiv

  25. [33]

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. 2019. Do imagenet classifiers generalize to imagenet?. In International conference on machine learning. PMLR, 5389–5400

    2019

  26. [34]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  27. [35]

    Melissa Sanabria, Frédéric Precioso, and Thomas Menguy. 2020. Profiling actions for sport video summarization: An attention signal analysis. In 2020 IEEE 22nd International Workshop on Multimedia Signal Processing (MMSP) . IEEE, 1–6

    2020

  28. [36]

    Melissa Sanabria, Frédéric Precioso, and Thomas Menguy. 2021. Hierarchical multimodal attention for deep video summarization. In 2020 25th International conference on pattern recognition (ICPR) . IEEE, 7977–7984

    2021

  29. [37]

    Melissa Sanabria, Frédéric Precioso, Pierre-Alexandre Mattei, and Thomas Men- guy. 2022. A Multi-stage deep architecture for summary generation of soccer videos. arXiv preprint arXiv:2205.00694 (2022)

    2022 arXiv

  30. [38]

    Mehdi Houshmand Sarkhoosh, Sushant Gautam, Cise Midoglu, Saeed Shafiee Sabet, and Pål Halvorsen. 2024. Multimodal AI-based summarization and sto- rytelling for soccer on social media. In Proceedings of the 15th ACM Multimedia Systems Conference. 485–491

    2024

  31. [39]

    Mehdi Houshmand Sarkhoosh, Sushant Gautam, Cise Midoglu, Saeed Shafiee Sabet, Thomas Torjusen, and Pål Halvorsen. 2024. The SoccerSum Dataset for Automated Detection, Segmentation, and Tracking of Objects on the Soccer Pitch. In Proceedings of the 15th ACM Multimedia Systems C...

    2024

  32. [40]

    Ramon Sanabria, Ozan Caglayan, Shruti Palaskar, Desmond Elliott, Loïc Barrault, Lucia Specia, and Florian Metze. 2018. How2: a large-scale dataset for multimodal language understanding. arXiv preprint arXiv:1811.00347 (2018)

    2018 arXiv

  33. [41]

    Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. 2015. Tvsum: Summarizing web videos using titles. In Proceedings of the IEEE conference on computer vision and pattern recognition . 5179–5187

    2015

  34. [42]

    Ivan Sosnovik, Artem Moskalev, Cees Kaandorp, and Arnold Smeulders

  35. [43]

    Vincenzo Scotti, Licia Sbattella, Roberto Tedesco, et al. 2019. Sferanet: automatic generation of football highlights. In CS & IT Conference Proceedings , Vol. 9. CS & IT Conference Proceedings

    2019

  36. [44]

    Min Sun, Ali Farhadi, and Steve Seitz. 2014. Ranking domain-specific highlights by analyzing edited videos. InComputer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. Springer, 787–802

    2014

  37. [45]

    Tobias Weyand, Andre Araujo, Bingyi Cao, and Jack Sim. 2020. Google landmarks dataset v2-a large-scale benchmark for instance-level recognition and retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 2575–2584

    2020

  38. [46]

    Yifang Xu, Yunzhuo Sun, Benxiang Zhai, Youyao Jia, and Sidan Du. 2024. Mh-detr: Video moment and highlight detection with cross-modal transformer. In 2024 International Joint Conference on Neural Networks (IJCNN) . IEEE, 1–8

    2024

  39. [47]

    Tomáš Souček and Jakub Lokoč. 2020. TransNet V2: An effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838 (2020)

    2020 arXiv

  40. [48]

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. 2017. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)

    2017 arXiv

  41. [49]

    Wencheng Zhu, Jiwen Lu, Jiahao Li, and Jie Zhou. 2021. DSNet: A Flexible Detect- to-Summarize Network for Video Summarization. IEEE Transactions on Image Processing 30 (2021), 948–962. doi:10.1109/TIP.2020.3039886

    2021

  42. [50]

    Yubo Zhu, Wentian Zhao, Rui Hua, and Xinxiao Wu. 2023. Topic-aware video summarization using multimodal transformer. Pattern Recognition 140 (2023), 109578

    2023

  43. [51]

    Kuo-Hao Zeng, Tseng-Hung Chen, Juan Carlos Niebles, and Min Sun. 2016. Title generation for user generated videos. In Computer Vision–ECCV 2016: 14th Euro- pean Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14. Springer, 609–625

    2016

  44. [2014]

    In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13

    Creating summaries from user videos. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13. Springer, 505–520

    2014

  45. [2021]

    In 2021 IEEE international symposium on multimedia (ISM)

    Combining global and local attention with positional encoding for video summarization. In 2021 IEEE international symposium on multimedia (ISM) . IEEE, 226–234

    2021

  46. [2023]

    arXiv preprint MMSports ’25, October 27–28, 2025, Dublin, Ireland Artur Díaz-Juan, Coloma Ballester & Gloria Haro arXiv:2301.05213 (2023)

    Learning to Summarize Videos by Contrasting Clips. arXiv preprint MMSports ’25, October 27–28, 2025, Dublin, Ireland Artur Díaz-Juan, Coloma Ballester & Gloria Haro arXiv:2301.05213 (2023)

    2025 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.