Pith. sign in

REVIEW 3 major objections 6 minor 53 references

TimeSoccer: An End-to-End Multimodal Large Language Model for Soccer Commentary Generation

T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read TimeSoccer generates soccer commentary in one pass over full matches.

desk verdict A promising end-to-end soccer captioning system undermined by inconsistent ablations and a possibly confounded SoTA comparison; worth a second look after major fixes. read the letter →

arxiv 2504.17365 v3 pith:SL2PEJWW submitted 2025-04-24 cs.CV cs.CL

Ling You , Wenxuan Huang , Xinni Xie , Xiangyi Wei , Bangyan Li , Shaohui Lin , Yang Li , Changbo Wang This is my paper · ORCID
classification cs.CVcs.CL
keywords soccercommentarygenerationdensevideocaptioningmultimodallargelanguagemodeltemporallocalizationlongunderstandingmotion-awareframeselectionend-to-end
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

TimeSoccer aims to establish that soccer commentary can be generated end-to-end from a full 45-minute match video, without any pre-detected event timestamps or a separate timestamping stage. The paper proposes a multimodal large language model that predicts when events happen and writes the commentary in a single forward pass, while compressing the long video with a motion-aware frame-selection module. It argues that this single-pass design preserves global match context, so commentary can refer back to earlier events, and reports state-of-the-art performance on the Single-anchor Dense Video Captioning task. A sympathetic reader should care because current soccer captioning systems either need ground-truth timestamps or split the match into short clips and lose the global picture.

What carries the argument

The load-bearing component is MoFA-Select, a training-free, motion-aware frame compression module with a coarse-to-fine strategy. It first clusters frames with a time-constrained variant of K-Means, then allocates the fixed frame budget to clusters according to the variance of their features, and finally merges adjacent redundant frames while protecting high-motion segments. This reduces a 45-minute match to a fixed-length representation that can pass through the sliding video Q-former and the language model. Two complementary training mechanisms—progressive lengthening of training videos and position-embedding extrapolation by periodic replication—let the pretrained model generalize beyond its original temporal range.

What would settle it

Run a blinded human evaluation on the test matches, with expert soccer viewers comparing TimeSoccer's single-pass 45-minute commentary with the two-step baseline's commentary for temporal accuracy and informativeness; if the baseline is rated equal or better, the paper's full-match quality claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that the two-step paradigm for soccer dense video captioning—first localize events, then caption each short clip—can be replaced by one end-to-end model that jointly outputs timestamps and captions for full 45-minute matches. On the SoccerNet-Caption benchmark, TimeSoccer is reported to exceed the previous best method in temporal localization precision, for example Precision@0.3 of 17.0 versus 12.5 in the 45-minute setting, and to score higher when a large vision-language model judges match-level semantic alignment and commentary quality. The authors attribute the gains to jointly modeling the whole match, so a caption can use context from earlier moments, and to training procedures that extend the model to match length.

Load-bearing premise

The strongest claim rests on the assumption that the automated language-model judge gives valid, unbiased quality scores, because on the standard CIDEr metric TimeSoccer's full-match output trails the two-step baseline it challenges.

Editorial extensions

If this is right

  • Full-match soccer commentary can be produced in a single pass, removing the need for a separate event-spotting stage or ground-truth timestamps at inference.
  • Commentary can use global context, so a caption can explain a substitution by referring to an earlier injury or continue an attacking narrative after a prior corner.
  • Motion-aware frame compression is training-free, so it can be dropped into other long-video multimodal models without retraining the compression module.
  • On strict temporal overlap thresholds such as IoU 0.7 and 0.9, the reported gains are larger, suggesting the model localizes key moments more sharply rather than only coarsely.
  • The same single-pass objective could make full-length sports commentary systems practical for broadcast settings, where events are sparse and long-range context matters.

Reading between the lines

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

  • If the single-pass design transfers, the same joint timestamp-and-caption objective could be applied to other long sports broadcasts, where events are sparse and global context matters.
  • Because the headline quality gains come from a vision-language model judge rather than n-gram overlap, a direct human comparison of TimeSoccer and SoccerNet-Caption outputs on the same test matches would be a natural next check.
  • The motion-aware allocation rule is simple enough to be reused as a preprocessing front-end for any long-form video multimodal model that has a frame-token budget, independent of the soccer task.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes TimeSoccer, an end-to-end multimodal large language model for Single-anchor Dense Video Captioning (SDVC) over full 45-minute soccer matches. The model jointly predicts timestamps and generates commentary in a single pass, building on TimeChat and introducing a training-free motion-aware frame compression module (MoFA-Select) together with progressive training and position-embedding extrapolation. Experiments on the SoccerNet-Caption dataset with MatchTime-refined temporal labels report temporal grounding metrics, CIDEr/METEOR/SODA_c, and LLM-judge scores (M-S and C-S), and the paper claims state-of-the-art performance in the end-to-end setting.

Significance. If the experimental claims are validated, the end-to-end formulation for long-form soccer commentary is a useful contribution: it removes the two-stage dependence on a separate spotting module, enables global-context modeling across a full half, and the proposed frame-selection module is training-free and architecture-agnostic. The paper also provides a broad comparison set and explicit ablations of its components. However, the central performance claim currently rests on a potentially confounded baseline comparison and on unvalidated LLM-judge metrics, since on standard caption metrics the 45-minute model does not beat SoccerNet-Caption. The strengths are real but the evidence is not yet sufficient to support the stated SoTA conclusion.

major comments (3)
  1. [Section 4.2 / Table 1] The evaluation-protocol mismatch with SoccerNet-Caption is load-bearing for the SoTA claim. The text states that the authors 'replace the original timestamp annotations with the refined temporal labels provided by MatchTime' and use the 422/49 split from MatchTime, yet Table 1 reports SoccerNet-Caption's P@0.3/0.5/0.7/0.9 = 12.5/7.2/3.0/1.1, F1 = 8.9, and CIDEr = 11.0 without stating that SoccerNet-Caption was re-run on the same refined labels and the same 49-video split. Because CIDEr and SODA_c are computed against the same ground-truth timestamps and captions, a comparison across differently labeled test sets is not apples-to-apples. The claimed temporal gains (+4.5, +3.8, +3.0, +2.3) and the caption-quality comparison may therefore be artifacts of label refinement or split differences. Please specify the exact evaluation protocol for every baseline, or re-run SoccerNet-Caption under the identical protocol used for TimeSoccer.
  2. [Section 4.2 / Table 1] The 45-minute superiority claim relies on the unvalidated LLM-judge metrics M-S and C-S. On the standard caption metrics, TimeSoccer (45-minute) scores CIDEr 8.3 vs. 11.0, F1 8.8 vs. 8.9, and SODA_c 2.7 vs. 3.9 compared with SoccerNet-Caption; only the Qwen2.5-VL-72B-Instruct scores (M-S 5.03 vs. 3.41, C-S 5.14 vs. 4.30) favor TimeSoccer. No human evaluation, judge-versus-human correlation, prompt template, or scoring rubric is reported for these 1–10 ratings. Without such validation, the headline claim that TimeSoccer generates higher-quality 45-minute commentary is not supported. Please add a human study or provide evidence that the LLM judge agrees with human judgments.
  3. [Section 4.4 / Tables 2 and 3] The ablation statements are contradicted by the reported numbers. Table 2 shows that removing the Motion-Aware component raises SODA_c from 2.7 (full) to 4.1 while lowering METEOR from 6.2 to 2.9, so the text's claim that 'removing any of them leads to a performance drop' is false for SODA_c. In Table 3, the full model (Progressive + MoFA-Select) has CIDEr 8.3, which is lower than the progressive-training-with-repeated-PosEnc row at 9.2, suggesting that MoFA-Select may hurt CIDEr in that configuration. These inconsistencies must be explained, or the component-necessity claims should be revised.
minor comments (6)
  1. [Abstract and Section 1] The term 'Single-anchor Dense Video Captioning' is used in the abstract and introduction but is never formally defined; please define it and explain how it differs from the standard dense video captioning setting.
  2. [Throughout] The module name is spelled inconsistently as both 'MoFA-Select' and 'MoFa-Select'; please use one spelling consistently.
  3. [Section 1, Contributions] The first contribution bullet contains the typo 'We propos'; please correct it.
  4. [Section 4.1] Several hyperparameters needed for reproducibility are missing: the number of sampled frames N_o, the compressed length N_p, the sliding-window length M_w and stride S, and the details of the progressive training schedule (number of epochs or steps at each video length).
  5. [Section 4.2] The prompts and scoring instructions used with Qwen2.5-VL-72B-Instruct for the M-S and C-S metrics are not reported; please include them in the appendix.
  6. [Section 4.3] The paper uses the 3-minute via 15x-inference setting for the 'surpassing the previous SoTA by +2.8 CIDEr' claim, while also stating that this setting is not fully end-to-end; the framing should clearly separate the end-to-end 45-minute results from the non-end-to-end 3-minute results so that the SoTA claim is not ambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; TimeSoccer is an empirical system paper whose claims rest on external benchmarks and external LLM judges rather than on self-citation or definitional reduction.

full rationale

The paper contains no derivation chain in which a predicted quantity is defined in terms of itself, a fitted parameter is renamed as a prediction, or a load-bearing premise is imported solely from the authors' own prior work. TimeSoccer is an end-to-end MLLM built on the external TimeChat baseline with a newly proposed MoFA-Select frame-selection module, and its headline claims are evaluated on the external SoccerNet-Caption dataset against external baselines under standard metrics (P@IoU, F1, CIDEr, METEOR, SODA_c). The Qwen2.5-VL-72B-Instruct judge used for M-S and C-S scores is an external model, not a self-citation, and the scores are reported as additional evidence rather than as a fitted input called a prediction. The self-citations in the Related Works section (e.g., [15,16,20]) are not load-bearing for the central claim. The most serious concern in the paper is experimental rather than circular: the authors state they 'replace the original timestamp annotations with the refined temporal labels provided by MatchTime' and evaluate their model on that version, while Table 1 reports SN-Caption's numbers without explicitly stating that the baseline was re-run on the same refined labels and the same 49-video split. This is a potential apples-to-oranges comparison and a validity risk, but it does not make any result equivalent to its own inputs by construction, so it falls outside the circularity patterns defined here. Because no circular step can be quoted and exhibited as a specific reduction, the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper's central claims depend on several modeling choices and external evaluation assumptions. The most consequential is the reliance on LLM-judged metrics for the 45-minute superiority claim, plus the unstated value of the compressed frame length N_p. No new physical or conceptual entities are introduced.

free parameters (5)
  • delta (motion penalty threshold) = 0.3
    Set empirically based on validation performance (Sec. 3.3).
  • number of clusters U = 6
    Hyperparameter for coarse-grained temporal clustering (Sec. 4.1).
  • LoRA rank = 32
    Rank used for parameter-efficient fine-tuning of the LLM (Sec. 4.1).
  • compressed length N_p
    Fixed-length frame budget after MoFA-Select; value not explicitly stated (Sec. 3.3).
  • progressive training schedule = 3 -> 15 -> 45 minutes
    Training video lengths used in the progressive training strategy (Sec. 3.4).
assumptions (4)
  • domain assumption TimeChat provides a valid time-aware visual feature extractor and sliding video Q-former for long video understanding
    The entire architecture builds on TimeChat as the baseline; Sec. 3.2 states this dependency.
  • domain assumption Soccernet-Caption with MatchTime refined labels is an appropriate benchmark for SDVC
    Evaluation uses this dataset, and the refined temporal labels may change the ground truth; Sec. 4.2 describes the data.
  • domain assumption Qwen2.5-VL-72B-Instruct provides reliable and unbiased scoring of commentary quality
    The M-S and C-S scores are used to claim superiority in the 45-minute setting where standard CIDEr is not superior; Sec. 4.2 introduces these metrics.
  • domain assumption Motion variance in feature space is a sufficient proxy for semantic importance of frames
    MoFA-Select allocates frames based on feature variance (Sec. 3.3), assuming that dynamic moments are the most important for commentary.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TimeSoccer: An End-to-End Multimodal Large Language Model for Soccer Commentary Generation." pith.science (2026). https://pith.science/paper/SL2PEJWW

@misc{pith2026250417365,
  author       = {Pith},
  title        = {Pith review of: TimeSoccer: An End-to-End Multimodal Large Language Model for Soccer Commentary Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SL2PEJWW}},
  note         = {Machine review of arXiv:2504.17365}
}
read the original abstract

Soccer is a globally popular sporting event, typically characterized by long matches and distinctive highlight moments. Recent advances in Multimodal Large Language Models (MLLMs) offer promising capabilities in temporal grounding and video understanding, soccer commentary generation often requires precise temporal localization and semantically rich descriptions over long-form video. However, existing soccer MLLMs often rely on the temporal a priori for caption generation, so they cannot process the soccer video end-to-end. While some traditional approaches follow a two-step paradigm that is complex and fails to capture the global context to achieve suboptimal performance. To solve the above issues, we present TimeSoccer, the first end-to-end soccer MLLM for Single-anchor Dense Video Captioning (SDVC) in full-match soccer videos. TimeSoccer jointly predicts timestamps and generates captions in a single pass, enabling global context modeling across 45-minute matches. To support long video understanding of soccer matches, we introduce MoFA-Select, a training-free, motion-aware frame compression module that adaptively selects representative frames via a coarse-to-fine strategy, and incorporates complementary training paradigms to strengthen the model's ability to handle long temporal sequences. Extensive experiments demonstrate that our TimeSoccer achieves State-of-The-Art (SoTA) performance on the SDVC task in an end-to-end form, generating high-quality commentary with accurate temporal alignment and strong semantic relevance.

Figures

Figures reproduced from arXiv: 2504.17365 by the authors.

Figure 1
Figure 1. Comparison of soccer commentary performance between SoccerNet-Caption [ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An illustration highlighting the global contextual understanding of TimeSoccer compared to SN-Caption [ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of TimeSoccer. Given a full 45-minute soccer video, frame features are extracted using an Image Encoder [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Quality comparison results on different methods. TimeSoccer demonstrates its advantages from multiple perspectives: [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 26 canonical work pages

  1. [1]

    Mohammad Abu Tami, Huthaifa I Ashqar, Mohammed Elhenawy, Sebastien Glaser, and Andry Rakotonirainy. 2024. Using multimodal large language models (MLLMs) for automated detection of traffic safety-critical events. Vehicles 6, 3 (2024), 1571–1590

    work page 2024

  2. [2]

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. 2022. Flamingo: a visual language model for few-shot learning, Vol. 35. 23716–23736

    work page 2022

  3. [3]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al . 2025. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

    arXiv 2025

  4. [4]

    Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization. 65–72

    work page 2005

  5. [5]

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feicht- enhofer, and Judy Hoffman. 2023. Token Merging: Your ViT But Faster. In ICLR

    work page 2023

  6. [6]

    Ho Kei Cheng and Alexander G Schwing. 2022. XMem: Long-Term Video Object Segmentation with an Atkinson-Shiffrin Memory Model. In Computer Vision– ECCV 2022: 17th European Conference, Tel A viv, Israel, October 23–27, 2022, Pro- ceedings, Part XXVIII . Springer, 640–658

    work page 2022

  7. [7]

    Anthony Cioppa, Adrien Deliege, Noor Ul Huda, Rikke Gade, Marc Van Droogen- broeck, and Thomas B Moeslund. 2020. Multimodal and multiview distillation for real-time player detection on a football field. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops . 880–881

    work page 2020

  8. [8]

    Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. 2022. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. 3491–3502

    work page 2022

Show all 53 references
  1. [9]

    Anthony Cioppa, Silvio Giancola, Vladimir Somers, Victor Joos, Floriane Magera, Jan Held, Seyed Abolfazl Ghasemzadeh, Xin Zhou, Karolina Seweryn, Mateusz Kowalczyk, Zuzanna Mróz, Szymon Łukasik, Michał Hałoń, Hassan Mkhallati, Adrien Deliège, Carlos Hinojosa, Karen Sanchez, Am...

    2024 arXiv

  2. [10]

    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

  3. [11]

    Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. 2024. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision . Springer, 75–92

    2024

  4. [12]

    Soichiro Fujita, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. 2020. SODA: Story Oriented Dense Video Captioning Eval- uation Framework. In European Conference on Computer Vision . https://api. semanticscholar.org/CorpusID:226291804

    2020

  5. [13]

    Huijun Gao, Hao An, Weiyang Lin, Xinghu Yu, and Jianbin Qiu. 2023. Trajec- tory tracking of variable centroid objects based on fusion of vision and force perception. IEEE Transactions on Cybernetics 53, 12 (2023), 7957–7965

    2023

  6. [14]

    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

  7. [15]

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Yao Hu, and Shaohui Lin. 2025. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749 (2025)

    2025 arXiv

  8. [16]

    Wenxuan Huang, Zijie Zhai, Yunhang Shen, Shaosheng Cao, Fei Zhao, Xiangfeng Xu, Zheyu Ye, and Shaohui Lin. 2025. Dynamic-llava: Efficient multimodal large language models via dynamic vision-language context sparsification.ICLR (2025)

    2025

  9. [17]

    Yutao Jiang, Qiong Wu, Wenhao Lin, Wei Yu, and Yiyi Zhou. 2025. What Kind of Visual Tokens Do We Need? Training-free Visual Token Pruning for Multi- modal Large Language Models from the Perspective of Graph. arXiv preprint arXiv:2501.02268 (2025)

    2025 arXiv

  10. [18]

    Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. 2024. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 13700–13710

    2024

  11. [19]

    Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles

  12. [20]

    Bangyan Li, Wenxuan Huang, Yunhang Shen, Yeqiang Wang, Shaohui Lin, Jingzhong Lin, Ling You, Yinqi Zhang, Ke Li, Xing Sun, et al . 2025. LLaVA- RadZ: Can Multimodal Large Language Models Effectively Tackle Zero-shot Radiology Recognition? arXiv preprint arXiv:2503.07487 (2025)

    2025

  13. [21]

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. 2024. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326 (2024)

    2024 arXiv

  14. [22]

    Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. 2023. Llava-med: Train- ing a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems 36 (...

    2023

  15. [23]

    Haoxuan Li, Zhengmao Yang, Yunshan Ma, Yi Bin, Yang Yang, and Tat-Seng Chua

  16. [24]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. 19730–19742

    2023

  17. [25]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning . PMLR, 12888–12900

    2022

  18. [26]

    Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. 2023. VideoChat: Chat-Centric Video Understanding. CoRR (2023)

    2023

  19. [27]

    Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. 2024. MVBench: A Comprehensive Multi-modal Video Understanding Benchmark. 22195–22206

    2024

  20. [28]

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Khan. 2024. Video-ChatGPT: Towards Detailed Video Understanding via Large Vision and Language Models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . 12585–12602

    2024

  21. [29]

    Hassan Mkhallati, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. 2023. SoccerNet-caption: Dense video captioning for soccer broadcasts commentaries. In Proceedings of the IEEE/CVF Conference on Computer Conference acronym ’XX, June 03–05, 2018, Wo...

    2023

  22. [30]

    Ji Qi, Jifan Yu, Teng Tu, Kunyu Gao, Yifan Xu, Xinyu Guan, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li, et al. 2023. GOAL: A challenging knowledge-grounded video captioning benchmark for real-time soccer commentary generation. 5391–5395

    2023

  23. [31]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al

  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]

    Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. 2024. Matchtime: Towards automatic soccer game commentary generation. arXiv preprint arXiv:2406.18530 (2024)

    2024 arXiv

  26. [34]

    Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. 2024. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14313–14323

    2024

  27. [35]

    Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. 2024. Moviechat: From dense token to sparse memory for long video understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and...

    2024

  28. [36]

    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

  29. [37]

    Quan Sun, Yuxin Fang, Ledell Yu Wu, Xinlong Wang, and Yue Cao. 2023. EVA- CLIP: Improved Training Techniques for CLIP at Scale. ArXiv abs/2303.15389 (2023). https://api.semanticscholar.org/CorpusID:257766387

    2023 arXiv

  30. [38]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023)

    2023 arXiv

  31. [39]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

    2023 arXiv

  32. [40]

    Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. 4566–4575

    2015

  33. [41]

    Ye Wang, Boshen Xu, Zihao Yue, Zihan Xiao, Ziheng Wang, Liang Zhang, Dingyi Yang, Wenxuan Wang, and Qin Jin. 2025. TimeZero: Temporal Video Grounding with Reasoning-Guided LVLM. arXiv preprint arXiv:2503.13377 (2025)

    2025 arXiv

  34. [42]

    Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang

  35. [43]

    Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont- Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. 2023. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. 10714–10726

    2023

  36. [44]

    Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. 2024. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985 (2024)

    2024 arXiv

  37. [45]

    Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. 2024. Timesuite: Improving mllms for long video understanding via grounded tuning. arXiv preprint arXiv:2410.19702 (2024)

    2024 arXiv

  38. [46]

    In European Conference on Computer Vision

    Longvlm: Efficient long video understanding via large language models. In European Conference on Computer Vision . Springer, 453–470

  39. [47]

    Hang Zhang, Xin Li, and Lidong Bing. 2023. Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding. In EMNLP (Demos)

    2023

  40. [48]

    Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chun- yuan Li. 2024. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713 (2024)

    2024 arXiv

  41. [49]

    Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. 2024. Streaming Dense Video Captioning

    2024

  42. [50]

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. 2023. Sig- moid loss for language image pre-training. 11975–11986

    2023

  43. [2017]

    In Proceedings of the IEEE international conference on computer vision

    Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision . 706–715

  44. [2021]

    Learning transferable visual models from natural language supervision

  45. [2024]

    In Proceedings of the 32nd ACM International Conference on Multimedia

    MM-Forecast: A Multimodal Approach to Temporal Event Forecasting with Large Language Models. In Proceedings of the 32nd ACM International Conference on Multimedia. 2776–2785

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.