Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

FineBadminton: A Multi-Level Dataset for Fine-Grained Badminton Video Understanding

T0 review · 3 major / 5 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 A new multi-level badminton video dataset and benchmark show that fine-tuned open-source multimodal models can outperform commercial ones on fine-grained sports understanding.

desk verdict FineBadminton is a genuinely useful new badminton dataset with a smart annotation hierarchy, but the benchmark numbers are suspect until the authors show a train/test split. read the letter →

arxiv 2508.07554 v1 pith:4ODDAQIJ submitted 2025-08-11 cs.MM

Xusheng He , Wei Liu , Shanshan Ma , Qian Liu , Chenghao Ma , Jianlong Wu This is my paper
classification cs.MM
keywords badmintonvideounderstandingmultimodallargelanguagemodelsquestionansweringbenchmarkhierarchicalannotationfine-grainedactionrecognitionkeyframeselectioncoordinate-guidedcondensationsportsintelligence
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 introduces FineBadminton, a large-scale badminton dataset with three levels of annotation—foundational actions, tactical semantics, and decision evaluation—and FBBench, a 2,563-question benchmark built on it. The authors also propose an optimized baseline that selects hit-centric keyframes and condenses visual information based on court coordinates, then fine-tune 7B open-source multimodal models. They report that the fine-tuned Qwen2.5VL-7B reaches 42.06% multiple-choice accuracy on FBBench, surpassing all commercial models they evaluated including Gemini 2.5 Pro. A sympathetic reader would care because this constitutes evidence that domain-specific hierarchical annotation and tailored video preprocessing can substantially improve multimodal LLMs on intricate sports reasoning, and because the dataset fills a gap in fine-grained sports video resources.

What carries the argument

The central object is the three-level annotation hierarchy: Foundational Actions (11 stroke types, 20 subtypes), Tactical Semantics (trajectory, intent, and outcome labels), and Decision Evaluation (per-shot quality scores plus rally-level narratives). The baseline's key mechanism is a two-stage video token transformation: Hit-Centric Keyframe Selection uses a VideoMAE-based hit detector to pick anchor frames around strokes, and Coordinate-Guided Condensation uses a ReSampler (Perceiver-style) with a coordinate bias tensor that adds a scalar to attention scores for tokens aligned with detected ball/player/court regions. This pipeline compresses a rally into a compact token sequence while pre

What would settle it

Request a disjoint split: train the fine-tuned model on QA pairs from one set of matches and evaluate on FBBench questions built exclusively from different matches, then compare the accuracy to the reported 42.06%. Alternatively, search the FBBench question text and its source rally clips for overlap with the 60,000 training QA pairs; any significant overlap would invalidate the claim that the fine-tuned model generalizes beyond memorization.

Watch

Extended reading notes

Core claim

The central claim is that FineBadminton is the first public badminton dataset with a multi-level hierarchical annotation scheme combining stroke-level detail, tactical semantics, and expert evaluation, and that this depth enables meaningful benchmarking and improvement of multimodal large language models. The paper documents that current commercial and open-source MLLMs perform poorly on FBBench—even the best commercial model, Gemini 2.5 Pro, only reaches 38.62% multiple-choice accuracy—yet a modestly sized open-source model fine-tuned on roughly 60,000 QA pairs derived from the dataset and fed a compact, hit-centric token sequence reaches 42.06%. The authors attribute this gain to their two

Load-bearing premise

The load-bearing premise is that the roughly 60,000 QA pairs used to fine-tune the open-source models do not overlap with the FBBench evaluation questions or rallies; the paper does not describe a split between training and benchmark rallies, so if the benchmark questions derive from the same rallies as the training pairs, the reported gains are inflated.

Editorial extensions

If this is right

  • Fine-grained sports video understanding can be formulated as a hierarchical annotation problem, enabling evaluation beyond simple action recognition to tactical reasoning and decision quality.
  • MLLM-generated proposals refined by human experts can scale the creation of dense, multi-level sports annotations that would be prohibitively expensive to produce manually.
  • Hit-centric and coordinate-guided visual condensation strategies can improve MLLM performance on high-speed, temporally dense video tasks, potentially extending to other sports with predictable court geometry.
  • Open-source 7B models fine-tuned on tailored, domain-specific data can surpass commercial API models on specialized benchmarks, suggesting that domain adaptation may matter more than raw model scale for such tasks.
  • The FineBadminton/FBBench ecosystem provides a reusable testbed for advancing spatio-temporal reasoning, action quality assessment, and strategic inference in sports AI.

Reading between the lines

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

  • The coordinate-bias attention mechanism could generalize to other court-based sports (tennis, squash, volleyball) where ball and player coordinates are detectable, but this application is not tested in the paper.
  • If the fine-tuning QA pairs and FBBench questions are drawn from the same rally clips, the reported superiority of fine-tuned models may reflect benchmark leakage; a clean held-out evaluation across disjoint matches would be a stronger test, and the paper does not report such a split.
  • The annotation pipeline's reliance on a VideoMAE hit detector and Fast R-CNN coordinate detector means that downstream model quality is bounded by the accuracy of these upstream components; errors there would propagate to the semantic descriptions and benchmark labels.
  • The open-ended scoring uses GPT-4.1 as a judge, so the reported 43.67% open-ended score for Qwen2.5VL could be influenced by judge preference for certain response styles, not purely by factual correctness.
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 / 5 minor

Summary. The paper introduces FineBadminton, a badminton video dataset with a three-level annotation hierarchy (Foundational Actions, Tactical Semantics, Decision Evaluation), built through an MLLM-proposal-plus-human-refinement pipeline. It also introduces FBBench, a benchmark with 2,563 QA pairs over four task families (Count, Action, Position, Cognition), and proposes an optimized baseline with Hit-Centric Keyframe Selection (S) and Coordinate-Guided Condensation (C). Experiments on FBBench show that fine-tuning Qwen2.5VL-7B and VideoLLaMA3-7B with S+C yields large gains, with Qwen2.5VL-7B reaching 42.06% multiple-choice accuracy, surpassing commercial models such as Gemini 2.5 Pro.

Significance. If the benchmark is clean, the dataset is a potentially valuable resource: it is, to my knowledge, the first badminton dataset with coupled stroke-level, tactical, and evaluative textual annotations, and the annotation pipeline is a practical approach to scaling such annotations. FBBench covers a broader set of reasoning tasks than prior badminton datasets, and the proposed strategies show consistent gains in the ablations. However, the central empirical claim—that fine-tuned 7B models surpass commercial zero-shot models—hinges on FBBench being a held-out evaluation set, and this is not established. The open-ended scoring methodology also needs calibration. These issues are fixable but currently limit the reliability of the headline results.

major comments (3)
  1. [§4 and §6.1 (Fine-Tuning)] No train/evaluation split is described between the ~60k QA pairs used for fine-tuning and the FBBench evaluation set. Section 6.1 says the training pairs are 'constructed from the FineBadminton' with question types based on FBBench categories, and Section 4 says FBBench is 'built upon' FineBadminton. The paper reports results on FBBench for models fine-tuned on these pairs, but it never states that the FBBench questions/rallies are disjoint from the training rallies or question templates. This is load-bearing for Table 1 and Table 2: without a match-level or rally-level split, the 42.06% accuracy of Qwen2.5VL-7B+S+C versus 38.62% for Gemini 2.5 Pro+S may reflect memorization of seen answers rather than generalization. Please specify the exact split (e.g., disjoint matches, disjoint rallies, and no overlapping questions) and report the number of training/evaluation examples per split.
  2. [§6.1 (Metrics)] Open-ended questions are scored by GPT-4.1 on a 0–10 scale against reference answers, but no human agreement or calibration is reported. Because the reference answers derive from the same MLLM-proposal-plus-human-refinement pipeline (Section 3.2), the open-ended scores in Table 1 (e.g., Qwen2.5VL-7B: 655) may reflect alignment with GPT-4.1-style phrasing rather than task competence. Please report human-evaluator agreement on a subset (e.g., 50 responses) or, at minimum, correlation between GPT-4.1 scores and human scores, and describe the evaluation rubric.
  3. [§3.2 and §4] FBBench is built on annotations that were initially proposed by MLLMs and then refined by human experts. The degree of human refinement is not quantified (e.g., percentage of text changed, number of corrections per rally, inter-annotator agreement). This matters because a benchmark derived from MLLM-generated text may inadvertently test an MLLM's ability to imitate the annotation style rather than general badminton understanding. I recommend reporting annotation statistics and, if possible, a human-baseline score on FBBench to calibrate the difficulty and to confirm that the benchmark reflects expert consensus.
minor comments (5)
  1. [Table 2 vs. §4] Table 2 uses the category name 'Localization' while Table 1 and Section 4 use 'Position'. Please align terminology.
  2. [§5 and §3.2] Key hyperparameters are not reported: the coordinate-bias scalar α in Eq. (1), the context window size w in §3.2, the number of ReSampler queries R, and the rule for choosing the variable number of query frames N_i in §5.1. Without these values, the method is not fully reproducible. Please provide them in the text or an appendix, and ideally a sensitivity analysis for α and R.
  3. [§6.1 (Fine-Tuning)] The fine-tuning section says the training set is ~60k QA pairs with 10% multiple-choice and 90% open-ended, but it does not specify the number of epochs, learning rate, LoRA rank, or data sampling strategy. These details are needed to reproduce the fine-tuned results.
  4. [§3.1 (Annotation Process)] The annotation process describes an initial manual phase and a later pipeline-assisted phase, but no statistics are given on the distribution of samples between the two phases or on annotation time/cost. A brief quantitative summary would strengthen the claim of scalability.
  5. [§6.1 (Models and Methodologies)] For commercial models, only the S strategy is evaluated, while open-source models receive both S and C plus fine-tuning. This is a reasonable experimental design, but the paper should state explicitly that C is not applicable to API-based models in its current form, to avoid confusion when comparing Table 1 rows.

Circularity Check

1 steps flagged · score 4.0 of 10

No held-out split is documented between FBBench and the ~60k fine-tuning QA pairs; reported fine-tuned gains are not shown to be independent of the dataset used for training.

  1. fitted input called prediction [Section 5 'Fine-Tuning' and Section 4 'FBBench' (no train/eval split specified)]
    "The training set consists of approximately 60,000 QA pairs. We construct these pairs from the FineBadminton, designing question types based on the categories outlined in the FBBench. ... Built upon the meticulously annotated FineBadminton dataset, FBBench is designed to rigorously assess a model's capabilities."

    The ~60k QA pairs used for fine-tuning are 'constructed from the FineBadminton,' and the benchmark on which the paper reports 'substantial performance gains' is 'Built upon the meticulously annotated FineBadminton dataset.' The paper gives no train/eval split between these two sets—no disjoint matches, rallies, or question sources are specified. Consequently, the fine-tuned models' FBBench scores (e.g., Qwen2.5VL-7B 42.06% multiple-choice vs Gemini 2.5 Pro+S 38.62%) are not demonstrated to be generalization; absent a held-out split, the 'prediction' on FBBench can be statistically forced by memorizing the same annotation pool used for fitting. The headline comparison against zero-shot commercial models is therefore not an independent evaluation of the S+C strategies.

full rationale

The dataset construction itself is not circular: the annotation pipeline uses MLLM proposals plus human refinement, an external check, and the multi-level hierarchy is a new resource. The proposed S and C strategies are self-contained (coordinate-bias attention and keyframe selection) and not a restatement of the benchmark. The main circularity risk is at the evaluation level: training QA pairs and the FBBench benchmark both come from FineBadminton, with no documented split, so the central empirical claim of fine-tuned gains may reduce to memorization rather than generalization. I do not count the reuse of the VideoMAE hit detector in S as a separate circular step because the paper does not explicitly state that detector was trained on the FBBench videos, though it remains an additional data-leakage risk. Overall, the central claim is partially suspect but the dataset contribution has independent content, giving a moderate score.

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

The paper creates a dataset and benchmark, not new physical or formal entities. The multi-level annotation hierarchy is a classification scheme, not an invented entity. The central empirical claims rest on the reliability of off-the-shelf video tools, the validity of the custom annotation taxonomy, the effectiveness of human refinement, and an undocumented train/test separation for the benchmark. The paper lists several learned hyperparameters (α, w, R, hit threshold) without disclosing values.

free parameters (4)
  • Coordinate-bias scalar α = not disclosed
    In Section 5.2, α>0 is added to attention scores when a token spatially aligns with a detected ball, player, or court keypoint; its value is chosen by hand and not reported, and it directly steers which visual tokens the ReSampler retains.
  • Context window size w = not disclosed
    In Section 3.2, the refining LLM uses a window of w stroke descriptions around the current stroke to produce T'_j; w controls how much rally context is merged and is not reported.
  • ReSampler query count R = not disclosed
    In Section 5.2, the condensation module uses R learnable queries in R^{R×D}; R sets the compression ratio of each inter-hit segment and its value is not reported.
  • Anchor frame selection threshold = not disclosed
    The hit-event detection module outputs an anchor probability P(hit|f_i,V) and frames above an implicit threshold become anchors; the threshold is not specified in the paper.
assumptions (5)
  • domain assumption Off-the-shelf video tools (TrackNetV3, FastRCNN, VideoMAE) provide sufficiently accurate ball tracking, court/player localization, and hit detection to support the annotation pipeline.
    Invoked in Section 3.2 Structural Parsing; if these tools fail on occlusion or fast motion, the structured data D used for all downstream annotations is corrupted.
  • domain assumption The custom annotation taxonomy (11 primary stroke types, 20 subtypes, 9 court regions, tactical and decision categories) is a complete and valid representation of badminton semantics.
    Defined in Section 3.1 with professional input; the benchmark tasks are built on this ontology, so any mismatch with real gameplay limits the dataset's validity.
  • domain assumption Human refinement of MLLM-generated proposals yields unbiased ground-truth annotations at scale.
    Section 3.1 and Section 3.2 rely on four trained enthusiasts to correct MLLM drafts; no inter-annotator agreement or audit results are reported, so annotation bias from the MLLM seed may persist.
  • ad hoc to paper FBBench evaluation QA pairs do not overlap the fine-tuning QA pairs.
    Section 4 builds FBBench from FineBadminton and Section 6.1 builds ~60k training pairs from the same source, but no split is documented; the reported fine-tuned gains assume this independence.
  • domain assumption GPT-4.1's open-ended scores correlate with human judgment.
    Section 6.1 uses GPT-4.1 as an automated evaluator with scale 0-10; no validation against human ratings is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FineBadminton: A Multi-Level Dataset for Fine-Grained Badminton Video Understanding." pith.science (2026). https://pith.science/paper/4ODDAQIJ

@misc{pith2026250807554,
  author       = {Pith},
  title        = {Pith review of: FineBadminton: A Multi-Level Dataset for Fine-Grained Badminton Video Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ODDAQIJ}},
  note         = {Machine review of arXiv:2508.07554}
}
read the original abstract

Fine-grained analysis of complex and high-speed sports like badminton presents a significant challenge for Multimodal Large Language Models (MLLMs), despite their notable advancements in general video understanding. This difficulty arises primarily from the scarcity of datasets with sufficiently rich and domain-specific annotations. To bridge this gap, we introduce FineBadminton, a novel and large-scale dataset featuring a unique multi-level semantic annotation hierarchy (Foundational Actions, Tactical Semantics, and Decision Evaluation) for comprehensive badminton understanding. The construction of FineBadminton is powered by an innovative annotation pipeline that synergistically combines MLLM-generated proposals with human refinement. We also present FBBench, a challenging benchmark derived from FineBadminton, to rigorously evaluate MLLMs on nuanced spatio-temporal reasoning and tactical comprehension. Together, FineBadminton and FBBench provide a crucial ecosystem to catalyze research in fine-grained video understanding and advance the development of MLLMs in sports intelligence. Furthermore, we propose an optimized baseline approach incorporating Hit-Centric Keyframe Selection to focus on pivotal moments and Coordinate-Guided Condensation to distill salient visual information. The results on FBBench reveal that while current MLLMs still face significant challenges in deep sports video analysis, our proposed strategies nonetheless achieve substantial performance gains. The project homepage is available at https://finebadminton.github.io/FineBadminton/.

Figures

Figures reproduced from arXiv: 2508.07554 by the authors.

Figure 1
Figure 1. An annotation example from FineBadminton il [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of our automated annotation pipeline. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Hierarchical task structure of FBBench, showcasing [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Architecture of the optimized baseline approach. It employs a two-stage strategy: Hit-Centric Keyframe Selection and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. TactiPlay: Multi-Granularity Tactical Parsing and Video-Anchored Match Review for Amateur Badminton Players

    cs.HC 2026-07 conditional novelty 5.5 of 10

    A taxonomy-guided, rally-level, video-anchored review system elicits more frequent, concrete, actionable, and appropriate tactical reflections from amateur badminton players than a report-and-statistics baseline.

Reference graph

Works this paper leans on

53 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    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

  2. [2]

    Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. 2024. Sharegpt4video: Im- proving video understanding and generation with better captions. In Proceedings of the Advances in Neural Information Processing Systems , Vol. 37. 19472–19495

    work page 2024

  3. [3]

    Yu-Jou Chen and Yu-Shuen Wang. 2023. Tracknetv3: Enhancing shuttlecock tracking with augmentations and trajectory rectification. In Proceedings of ACM International Conference on Multimedia in Asia . 1–7

    work page 2023

  4. [4]

    Muhe Ding, Yang Ma, Pengda Qin, Jianlong Wu, Yuhong Li, and Liqiang Nie. 2025. RA-BLIP: Multimodal Adaptive Retrieval-Augmented Bootstrapping Language- Image Pre-training. IEEE Transactions on Multimedia (2025)

    work page 2025

  5. [5]

    Bernard Ghanem Fabian Caba Heilbron, Victor Escorcia and Juan Carlos Niebles

    work page

  6. [6]

    Miquel Farré, Andi Marafioti, Lewis Tunstall, Leandro Von Werra, and Thomas Wolf. 2024. FineVideo. https://huggingface.co/datasets/HuggingFaceFV/ finevideo

    2024

  7. [7]

    Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. 2024. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075 (2024)

    arXiv 2024

  8. [8]

    Ross Girshick. 2015. Fast r-cnn. InProceedings of the IEEE International Conference on Computer Vision. IEEE, 1440–1448

    work page 2015

Show all 53 references
  1. [9]

    Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. 2024. ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools. arXiv preprint arXiv:2406.12793 (2024)

    2024 arXiv

  2. [10]

    something something

    Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. 2017. The" something something" video database for learning and evaluating visual common sense....

    2017

  3. [11]

    Hulingxiao He, Geng Li, Zijun Geng, Jinglin Xu, and Yuxin Peng. 2025. Analyzing and Boosting the Power of Fine-Grained Visual Recognition for Multi-modal Large Language Models. arXiv preprint arXiv:2501.15140 (2025)

    2025 arXiv

  4. [12]

    Yuchen He, Zeqing Yuan, Yihong Wu, Liqi Cheng, Dazhen Deng, and Yingcai Wu

  5. [13]

    Jan Held, Hani Itani, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. 2024. X-vars: Introducing explainability in football refereeing with multi-modal large language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recog...

    2024

  6. [14]

    Kairui Hu, Penghao Wu, Fanyi Pu, Wang Xiao, Yuanhan Zhang, Xiang Yue, Bo Li, and Ziwei Liu. 2025. Video-MMMU: Evaluating Knowledge Acquisition from Multi-Discipline Professional Videos. arXiv preprint arXiv:2501.13826 (2025)

    2025 arXiv

  7. [15]

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. 2017. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)

    2017 arXiv

  8. [16]

    Jie Lei, Tamara L Berg, and Mohit Bansal. 2021. Detecting moments and highlights in videos via natural language queries. In Proceedings of the Advances in Neural Information Processing Systems, Vol. 34. NeurIPS Foundation, 11846–11858

    2021

  9. [17]

    Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. 2018. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696 (2018)

    2018 arXiv

  10. [18]

    Haopeng Li, Andong Deng, Qiuhong Ke, Jun Liu, Hossein Rahmani, Yulan Guo, Bernt Schiele, and Chen Chen. 2024. Sports-qa: A large-scale video question answering benchmark for complex and professional sports. arXiv preprint arXiv:2401.01505 (2024)

    2024

  11. [19]

    Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. 2024. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . IEEE, 22195–22206

    2024

  12. [20]

    Qi Li, Tzu-Chen Chiu, Hsiang-Wei Huang, Min-Te Sun, and Wei-Shinn Ku. 2024. Videobadminton: a video dataset for badminton action recognition. In 2024 IEEE International Conference on Big Data . IEEE, 1387–1392

    2024

  13. [21]

    Xiaojie Li, Shaowei He, Jianlong Wu, Yue Yu, Liqiang Nie, and Min Zhang. 2023. Mask again: Masked knowledge distillation for masked video modeling. In Pro- ceedings of the ACM International Conference on Multimedia . ACM, 2221–2232

    2023

  14. [22]

    Xiaojie Li, Jianlong Wu, Shaowei He, Shuo Kang, Yue Yu, Liqiang Nie, and Min Zhang. 2023. Fine-grained key-value memory enhanced predictor for video representation learning. In Proceedings of the ACM International Conference on Multimedia. ACM, 2264–2274

    2023

  15. [23]

    Yixuan Li, Lei Chen, Runyu He, Zhenzhi Wang, Gangshan Wu, and Limin Wang

  16. [24]

    Zijing Liang, Yanjie Xu, Yifan Hong, Penghui Shang, Qi Wang, Qiang Fu, and Ke Liu. 2024. A Survey of Multimodel Large Language Models. In Proceedings of In- ternational Conference on Computer, Artificial Intelligence and Control Engineering. 405–409

    2024

  17. [25]

    Zhaoyu Liu, Kan Jiang, Murong Ma, Zhe Hou, Yun Lin, and Jin Song Dong. 2025. F3Set: Towards Analyzing Fast, Frequent, and Fine-grained Events from Videos. arXiv preprint arXiv:2504.08222 (2025)

    2025 arXiv

  18. [26]

    Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. 2023. Egoschema: A diagnostic benchmark for very long-form video language un- derstanding. In Proceedings of the Advances in Neural Information Processing Systems, Vol. 36. 46212–46244

    2023

  19. [27]

    OpenAI. 2025. GPT-4.1. https://openai.com/index/gpt-4-1/

    2025

  20. [28]

    Jiayuan Rao, Haoning Wu, Hao Jiang, Ya Zhang, Yanfeng Wang, and Weidi Xie

  21. [29]

    Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. 2024. MatchTime: Towards Automatic Soccer Game Commentary Generation. In Pro- ceedings of the Conference on Empirical Methods in Natural Language Processing

    2024

  22. [30]

    Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. 2020. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . IEEE, 2616–2625

    2020

  23. [31]

    Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. 2019. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . IEEE, 1207–1216

    2019

  24. [32]

    Yunlong Tang, Junjia Guo, Hang Hua, Susan Liang, Mingqian Feng, Xinyang Li, Rui Mao, Chao Huang, Jing Bi, Zeliang Zhang, et al. 2025. Vidcomposition: Can mllms analyze compositions in compiled videos?. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recogn...

    2025

  25. [33]

    ByteDance Seed Team. 2025. Doubao-1.5-pro. https://seed.bytedance.com/en/ special/doubao_1_5_pro

    2025

  26. [34]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

    2023 arXiv

  27. [35]

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. 2022. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. In Proceedings of the Advances in Neural Information Processing Systems . 10078– 10093

    2022

  28. [36]

    Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. 2023. VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 14549–14560

    2023

  29. [37]

    Wei-Yao Wang, Wei-Wei Du, Wen-Chih Peng, and Tsi-Ui Ik. 2024. Benchmark- ing Stroke Forecasting with Stroke-Level Badminton Dataset. In Proceedings of International Joint Conference on Artificial Intelligence . 8829–8832

    2024

  30. [38]

    Wei-Yao Wang, Yung-Chang Huang, Tsi-Ui Ik, and Wen-Chih Peng. 2023. Shut- tleset: A human-annotated stroke-level singles dataset for badminton tactical analysis. In Proceedings of ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5126–5136

    2023

  31. [39]

    Xiao Wang, Jingyun Hua, Weihong Lin, Yuanxing Zhang, Fuzheng Zhang, Jian- long Wu, Di Zhang, and Liqiang Nie. 2025. HAIC: Improving Human Action Understanding and Generation with Better Captions for Multi-modal Large Language Models. In Proceedings of the Annual Meeting of the...

    2025

  32. [40]

    Xiao Wang, Qingyi Si, Shiyu Zhu, Jianlong Wu, Li Cao, and Liqiang Nie. 2025. AdaReTaKe: Adaptive Redundancy Reduction to Perceive Longer for Video- language Understanding. In Proceedings of the Annual Meeting of the Association for Computational Linguistics. ACL, 5417–5432. MM...

    2025

  33. [41]

    Xiao Wang, Jianlong Wu, Zijia Lin, Fuzheng Zhang, Di Zhang, and Liqiang Nie

  34. [42]

    Tao Wu, Runyu He, Gangshan Wu, and Limin Wang. 2024. Sportshhi: A dataset for human-human interaction detection in sports videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . IEEE, 18537–18546

    2024

  35. [43]

    Haotian Xia, Zhengbang Yang, Junbo Zou, Rhys Tracy, Yuqing Wang, Chi Lu, Christopher Lai, Yanjun He, Xun Shao, Zhuoqing Xie, Yuan fang Wang, Weining Shen, and Hanjie Chen. 2025. SPORTU: A Comprehensive Sports Understand- ing Benchmark for Multimodal Large Language Models. In P...

    2025

  36. [44]

    Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. 2017. Video Question Answering via Gradually Refined At- tention over Appearance and Motion. In Proceedings of the ACM International Conference on Multimedia. ACM, 1645–1653

    2017

  37. [45]

    Jinglin Xu, Guohao Zhao, Sibo Yin, Wenhao Zhou, and Yuxin Peng. 2024. Fines- ports: A multi-person hierarchical sports video dataset for fine-grained action understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 21773–21782

    2024

  38. [46]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2025), 2912–2923

    Video DataFlywheel: Resolving the Impossible Data Trinity in Video- Language Understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025), 2912–2923

    2025

  39. [47]

    Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. 2025. VideoLLaMA 3: Frontier Multimodal Foundation Models for Image and Video Understanding. arXiv preprint arXiv:2501.13106 (2025)

    2025 arXiv

  40. [48]

    Shiyi Zhang, Wenxun Dai, Sujia Wang, Xiangwei Shen, Jiwen Lu, Jie Zhou, and Yansong Tang. 2023. Logo: A long-form video dataset for group action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2405–2414. A The FineBad...

    2023

  41. [51]

    Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. 2024. Vript: A video is worth thousands of words. In Proceedings of the Advances in Neural Information Processing Systems , Vol. 37. 57240–57261

    2024

  42. [2015]

    In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    ActivityNet: A Large-Scale Video Benchmark for Human Activity Under- standing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 961–970

  43. [2021]

    In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Multisports: A multi-person video dataset of spatio-temporally localized sports actions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 13536–13545

  44. [2024]

    In Proceedings of the AAAI Conference on Artificial Intelligence

    ViSTec: Video Modeling for Sports Technique Recognition and Tactical Analysis. In Proceedings of the AAAI Conference on Artificial Intelligence . 8490– 8498

  45. [2025]

    In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Towards Universal Soccer Video Understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . IEEE

Pith tools

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