Pith. sign in

REVIEW 6 major objections 5 minor 39 references

Enhancing Sports Strategy with Video Analytics and Data Mining: Assessing the effectiveness of Multimodal LLMs in tennis video analysis

T0 review · 6 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A video-language model can identify tennis rally sequences much better when supplied with coordinates from traditional detectors than from raw video, yet still lags specialized sequence models.

desk verdict Plausible and useful MLLM sports-analysis results, but the central edit-score gains rest on an unspecified output-parsing procedure. read the letter →

arxiv 2507.02904 v1 pith:77TQZIOD submitted 2025-06-24 cs.CV cs.AI

Charlton Teo This is my paper
classification cs.CVcs.AI
keywords sportsanalyticssequenceidentificationmultimodallargelanguagemodelstennisvideoanalysisCLIPfine-tuningeditscoreobjectdetectionpromptengineering
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 dissertation asks whether multimodal large language models (MLLMs) can do fine-grained tennis video analysis: classifying individual shots and identifying the full sequence of events in a rally. Using the FineTennis dataset and VideoLLaMA2, the author finds that the model on its own reaches only 0.41 single-event accuracy and a 39.7 rally edit score, with the biggest weakness being the vision backbone rather than language understanding. Feeding the model textual coordinates from player, ball, and court detectors every two frames raises the edit score to 76.0, close to the traditional F3EST baseline of 82.1 (evaluated at a different granularity). Separately fine-tuning the CLIP vision encoder and freezing it back into the model lifts single-event accuracy to 0.56 and sequence edit score to 54.6. The thesis is that current MLLMs are not yet reliable standalone sequence identifiers for sports, but hybrid pipelines and two-stage vision-encoder tuning are concrete routes forward.

What carries the argument

The load-bearing machinery is prompt-injected coordinate information. A court-detection model supplies the four court corners, a ball tracker supplies ball-center coordinates, and a person detector supplies player bounding boxes; these are serialized as text into the VideoLLaMA2 prompt, using every other frame to fit memory limits. The second mechanism is two-stage CLIP fine-tuning: the frozen vision encoder is fine-tuned on tennis frames, frozen again, and reinserted into VideoLLaMA2 before LoRA fine-tuning of the rest of the model. The measuring instrument is the normalized segmental edit score, the Levenshtein distance between the predicted and true event sequences divided by the longer label length. This score turns how close a predicted rally is to the real rally into a single number.

What would settle it

Re-run F3EST on the same 56-event FineTennis test split and compare edit scores directly; if F3EST does not clearly beat the hybrid MLLM's 76.0, the paper's conclusion that traditional models outperform MLLMs on tennis sequence identification would collapse.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that VideoLLaMA2, fine-tuned with LoRA on the FineTennis dataset, can learn the structure of a tennis rally but not its fine-grained event content from video alone. It reliably produces sequences that start with a serve, end with a last shot, and alternate between near and far players, pattern knowledge the author attributes to textual reasoning, while failing on shot type and direction. The bottleneck is the frozen CLIP vision encoder and the STC (spatial-temporal) connector; a simplified model using only STC connector features reaches 0.018 accuracy, near random. Two interventions break the bottleneck: converting player bounding boxes, court corners, and ball positions into prompt text, which raises rally edit score to 76.0, and separately fine-tuning the vision encoder on tennis frames before freezing it back into the model, which raises single-event accuracy to 0.56 and edit score to 54.6. The author concludes that MLLMs have capable textual processing but a vision backbone that cannot yet extract such information on its own.

Load-bearing premise

The comparison with the traditional F3EST model assumes that edit scores computed on different event taxonomies (38 or 111 event types for F3EST, 56 for FineTennis) are directly comparable; if that assumption fails, the paper's claim that VideoLLaMA2 cannot match traditional models is not supported.

Editorial extensions

If this is right

  • For fine-grained video tasks, current MLLMs are best used as text reasoners over externally supplied detections, not as standalone video understanders.
  • Two-stage vision-encoder fine-tuning — tune CLIP separately, freeze it back into the model, then LoRA-tune the rest — is a reusable recipe that the paper shows improves both single-event accuracy and sequence edit score.
  • Event-count guidance produces a large jump (edit score from 34.4 to 49.8), indicating that a major part of the MLLM's sequence error is its inability to count events, not just classify them.
  • Frame sampling is not the main bottleneck: raising sampling from 8 to 32 frames helps, but sampling every frame does not, pointing to internal model limits.
  • The best hybrid configuration (76.0) approaches the traditional F3EST baseline (82.1) despite the granularity mismatch, suggesting hybrid MLLM pipelines are a near-term path to practical sports annotation tools.

Reading between the lines

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

  • If the prompt-injection result transfers, sports-analytics systems could use MLLMs as a flexible decoder layer: swap in detector outputs for other sports and keep the same language model, which would make fine-grained event annotation cheaper to adapt.
  • The two-stage CLIP tuning result suggests a general fine-tuning recipe for domain-specific video-language models; testing it on other base models and tasks would show whether the effect is specific to VideoLLaMA2 or general.
  • Because the F3EST comparison uses different event taxonomies, an exact re-run of F3EST on the 56-event FineTennis split is the natural next experiment; until then, the size of the gap between MLLMs and traditional models is not fully settled.
  • A cheap diagnostic benchmark suggested by the paper's event-counting experiment: ask an MLLM to count actions in a video, then compare counting accuracy against sequence edit score to isolate counting failures from classification failures.
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

6 major / 5 minor

Summary. The paper presents an empirical evaluation of VideoLLaMA2 on the FineTennis tennis dataset, testing single-event classification and rally sequence identification. It explores modifications including frame sampling (8 vs 32 vs all frames), audio, event counting (oracle and self-predicted), injecting player bounding boxes, court and ball coordinates, pose keypoints, and separately fine-tuning the CLIP vision encoder. The main findings are that the default MLLM achieves a best edit score of 39.7 on sequence identification, that injecting textual detection information raises this to 76.0, that providing the event count as an oracle improves the score to 49.8, and that separately fine-tuning CLIP improves single-event accuracy from 0.41 to 0.56 and sequence edit score to 54.6. The paper concludes that MLLMs still lag traditional models on fine-grained sequence identification and that their visual backbone is a bottleneck.

Significance. If the reported results are reliable, the paper contributes a useful negative result for MLLMs in a structured sports-video task and a practically promising recipe (separate vision-encoder fine-tuning, textual injection of detections) that others can build on. The use of an external dataset (FineTennis) and a comparison with the F3EST model is appropriate in spirit, and the paper is transparent about the oracle nature of the event-count variation. However, the quantitative claims are currently not fully verifiable because the sequence-output parsing procedure is unspecified, the benchmark comparison uses different event taxonomies without justification, and no statistical confidence is reported. These gaps substantially weaken the evidence for the central conclusions.

major comments (6)
  1. [3.6] The edit score metric is defined only at the level of 'each individual event as a single word,' but the paper never specifies how the free-form model output is converted into an event string. Section 3.6 lacks a parsing rule, sample outputs, or error analysis, and Section 3.4 provides a parsing rule only for single-event classification (substring matching with exclusion of other labels). Since all sequence results in Tables 5 and 7 and the event-counting results in Table 6 depend on this conversion, and since different parser rules (e.g., accepting partial events, handling missing sub-labels, stripping frame-number annotations in Variation 5) can materially change Levenshtein distances, the reported numbers (e.g., 39.7 vs 76.0) are not independently verifiable.
  2. [3.2, 4.4] The comparison of VideoLLaMA2's edit scores with the F3EST benchmark scores (88.4 on 38 event types; 82.1 on 111 event types) is not valid without controlling for the event-taxonomy granularity of FineTennis, which has 56 event types. The statement in Section 3.2 that 'these scores still serve as a good beginning benchmark' is an assertion, not a demonstration; edit scores are sensitive to label granularity. Consequently, the conclusion in Section 4.4 that VideoLLaMA2 'cannot compare to the current benchmark scores' is not supported by the presented evidence.
  3. [4.1] The claim that 'All rallies correctly began with a serve for e3 and ended with a last for e5' and that 'All rallies also correctly alternated between the near and far player' is presented as an empirical result, but the paper provides no measurement procedure, no error rates, and no supporting table or figure. This claim is used to argue for the model's textual reasoning strength, so it needs a concrete evaluation with numbers.
  4. [3.10.2] The full-CLIP experiment is ambiguous about how video is handled. CLIP is an image-text model, and the paper does not state whether each frame is classified independently, how the per-frame logits are aggregated (e.g., majority voting), or how temporal order is used. Without this, the 0.43 accuracy in Table 8 cannot be interpreted or compared with the VideoLLaMA2 single-event accuracy of 0.41.
  5. [3.4–3.10] No error bars, confidence intervals, or significance tests are reported for any experiment, and most results appear to come from a single run. Several load-bearing comparisons hinge on small differences (e.g., 0.56 vs 0.55 for 6 vs 10 epochs in Section 3.10.3; 34.4 vs 33.1 for 10 vs 20 epochs in Table 5). The paper should report multiple seeds and statistical testing, or at least acknowledge the noise level, before claiming improvements or lack thereof.
  6. [Appendix A] The GitHub repository URL is given as 'https://github.com/bigcrushes/videollama2 tennis', which contains a space and is not a valid URL. The code and evaluation scripts are therefore inaccessible, which is a concrete blocker for verifying the parsing and preprocessing steps that the edit-score results depend on.
minor comments (5)
  1. [2.2] The word 'abve' should be 'above'.
  2. [3.2] The reference for F3EST is listed as unpublished in the bibliography, but the same work appears as (Liu, Jiang, et al., 2025) elsewhere; please reconcile the citation.
  3. [3.5, Table 6] The 'Accuracy 0.36' in Table 6 is undefined; specify whether it is exact-count accuracy, per-rally accuracy, or another measure.
  4. [3.9.3, Table 7] Table 7 labels the pose experiments as 'in both video and prompt,' but the text says only the prompt line was replaced; clarify whether the videos were also annotated and, if so, how.
  5. [3.6] The edit score formula is informal; use the standard normalized segmental edit score notation and define the denominator precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: results are measured against an external dataset and benchmark, with the main caveat being a reproducibility issue, not circularity.

full rationale

This paper is an empirical evaluation, not a derivation. All central quantities (edit scores, accuracies, event-count differences) are measured from VideoLLaMA2 outputs against the FineTennis ground-truth labels, and the benchmark comparison uses F3EST as a separate model with its own reported scores. Nothing in the paper defines a predicted quantity in terms of the fitted values or equates an output to an input by construction. The 'providing event count' variation is explicitly an oracle condition and the paper states it will not be available on new data, so it is not a disguised prediction. The 'providing self-predicted event count' variation is a two-stage pipeline rather than a circular definition. The FineTennis dataset and F3EST benchmark come from the advisor's research group, which is a self-citation-adjacent concern, but the paper's measurements do not reduce to accepting those citations: the accuracy and edit scores are produced by the author's own runs. The main weakness identified by the skeptical reading is that the edit-score computation for free-form sequence outputs is not fully specified, so the reported numbers may be hard to reproduce; that is a reproducibility and correctness risk, not a circularity, because the evaluation procedure is not defined in terms of the target conclusion. No load-bearing step reduces to the paper's own assumptions or fitted parameters, so the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central empirical results depend mainly on hyperparameter choices (epochs, frame count, coordinate interval) and on benchmark and annotation validity assumptions. No new theoretical entities are introduced.

free parameters (3)
  • epochs = 10 (with 8, 20, 50 also tried)
    Chosen as the 'optimal balance' between accuracy and time (Section 3.6); results vary with epoch count, so the central numbers depend on this choice.
  • frame_sampling_count = 32 (with 8 and all-frames also tried)
    Selected because 32 frames improved edit score and all rallies have fewer than 32 events (Section 3.5, variation 4; Section 3.8). The main improvements are reported under 32-frame sampling.
  • coordinate_frame_interval = every 2 frames (also every 5 and every 20 frames)
    Chosen in the best-performing prompt-injection experiment due to RAM constraints (Section 3.9.2). Performance changed when the interval changed, so the reported 76.0 edit score depends on this interval.
assumptions (5)
  • domain assumption FineTennis event annotations (5 sub-classes, frame timings) are accurate and complete ground truth.
    All edit scores and accuracies are computed against these labels (Section 3.1).
  • standard math The normalized edit score is an appropriate measure of sequence identification quality.
    The metric is standard in action segmentation (Lea et al., 2016) and is used to compare methods (Section 3.6).
  • domain assumption Edit scores are comparable across different event taxonomies (F3EST's 38/111 event types versus FineTennis's 56 event types).
    Section 3.2 acknowledges the granularity differs but treats F3EST's scores as a benchmark; the conclusion that MLLMs underperform traditional models depends on this comparability.
  • ad hoc to paper The simplified STC connector probe (one hidden layer plus output layer) can meaningfully evaluate the STC connector's spatial-temporal features.
    Section 3.7 retrains the connector from scratch and adds a simple head; the paper itself notes that without another connector to compare to, the failure cannot be attributed to the STC connector.
  • domain assumption Fine-tuning the full CLIP model (including text embeddings) isolates the effect of the vision encoder when the resulting visual encoder is re-inserted into VideoLLaMA2.
    Section 3.10.2 uses the full CLIP model with its in-built output layer; the text-embedding component is assumed not to distort the vision encoder improvement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Sports Strategy with Video Analytics and Data Mining: Assessing the effectiveness of Multimodal LLMs in tennis video analysis." pith.science (2026). https://pith.science/paper/77TQZIOD

@misc{pith2026250702904,
  author       = {Pith},
  title        = {Pith review of: Enhancing Sports Strategy with Video Analytics and Data Mining: Assessing the effectiveness of Multimodal LLMs in tennis video analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/77TQZIOD}},
  note         = {Machine review of arXiv:2507.02904}
}
read the original abstract

The use of Large Language Models (LLMs) in recent years has also given rise to the development of Multimodal LLMs (MLLMs). These new MLLMs allow us to process images, videos and even audio alongside textual inputs. In this project, we aim to assess the effectiveness of MLLMs in analysing sports videos, focusing mainly on tennis videos. Despite research done on tennis analysis, there remains a gap in models that are able to understand and identify the sequence of events in a tennis rally, which would be useful in other fields of sports analytics. As such, we will mainly assess the MLLMs on their ability to fill this gap - to classify tennis actions, as well as their ability to identify these actions in a sequence of tennis actions in a rally. We further looked into ways we can improve the MLLMs' performance, including different training methods and even using them together with other traditional models.

Figures

Figures reproduced from arXiv: 2507.02904 by the authors.

Figure 1
Figure 1. Sample shot in the US Open 2022 Finals [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Sample shot in the Australian Open 2021 Semi-Finals [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Simplified STC Connector model Similar to previous experiments, we trained this model on 10 epochs. In the end, this simplified STC connector model performed poorly, attaining an accuracy of only 0.018, which was comparable to randomly guessing the class. This posed the question if the current STC connector was truly capable of cap￾turing the spatial temporal information it was supposed to. Without another connector… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Sample player bounding boxes 1 [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: Sample player bounding boxes 2 After extracting these player bounding boxes, we tried three different meth￾16 [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: Sample court detection result 1 18 [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]
Figure 7
Figure 7. Figure 7: Sample court detection result 2 We took it one step further and also found a ball detection model for tennis. In a similar fashion, we used the open source model to detect the ball in each frame of the video. Since the tennis balls are small in each frame, we only took…
Figure 8
Figure 8. Figure 8: Sample ball detection result 1 19 [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]
Figure 9
Figure 9. Figure 9: Sample ball detection result 2 Same as the player bounding boxes, if the court or ball is not detected, all coordinates will be defaulted to -1 for the model differentiate. To pass this information to the model, we again modify our prompt to the fol￾lowing: Given this …
Figure 10
Figure 10. Figure 10: Sample pose detection result 1 21 [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]
Figure 11
Figure 11. Figure 11: Sample pose detection result 2 Our initial visualisations show that our pose estimation model also does have decent results. First, we tried feeding all the keypoints into the model. With some basic searching, we found that the minimum interval is 20 frames. We then r…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 26 canonical work pages

  1. [1]

    Analyzing fast, frequent, and fine-grained event sequences from videos

    unpublishedkeyA Analyzing Fast, Frequent, and Fine-grained Event Sequences from Videos. Analyzing fast, frequent, and fine-grained event sequences from videos. 2024 . unpublished

    work page 2024

  2. [2]

    \ M \"u ller, O

    caron2023tacticalgpt APACrefauthors Caron, M. \ M \"u ller, O. APACrefauthors \ 2023 . TacticalGPT: uncovering the potential of LLMs for predicting tactical decisions in professional football Tacticalgpt: uncovering the potential of llms for predicting tactical decisions in professional football . StatsBomb Conference Statsbomb conference \ ( \ 1--11)

    work page 2023

  3. [3]

    \ Dolan, W B

    chen2011collecting APACrefauthors Chen, D. \ Dolan, W B. APACrefauthors \ 2011 . Collecting highly parallel data for paraphrase evaluation Collecting highly parallel data for paraphrase evaluation . Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies Proceedings of the 49th annual meeting of...

    work page 2011

  4. [4]

    , Leng, S

    cheng2024videollama APACrefauthors Cheng, Z. , Leng, S. , Zhang, H. , Xin, Y. , Li, X. , Chen, G. others APACrefauthors \ 2024 . VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms . arXiv preprint arXiv:2406.07476

    arXiv 2024

  5. [5]

    \ Karaku s , O

    cook2024llm APACrefauthors Cook, A. \ Karaku s , O. APACrefauthors \ 2024 . LLM-Commentator: Novel fine-tuning strategies of large language models for automatic commentary generation using football event data Llm-commentator: Novel fine-tuning strategies of large language models for automatic commentary generation using football event data . Knowledge-Bas...

    work page 2024

  6. [6]

    , Jiang, K

    dong2023sports APACrefauthors Dong, J S. , Jiang, K. , Liu, Z. , Dong, C. , Hou, Z. , Hundal, R S. Lin, Y. APACrefauthors \ 2023 . Sports analytics using probabilistic model checking and deep learning Sports analytics using probabilistic model checking and deep learning . 2023 27th International Conference on Engineering of Complex Computer Systems (ICECC...

    work page 2023

  7. [7]

    , Dai, Y

    fu2024video APACrefauthors Fu, C. , Dai, Y. , Luo, Y. , Li, L. , Ren, S. , Zhang, R. others APACrefauthors \ 2024 . Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis . arXiv preprint arXiv:2405.21075

    arXiv 2024

  8. [8]

    APACrefauthors \ 2004

    hawkins2004problem APACrefauthors Hawkins, D M. APACrefauthors \ 2004 . The problem of overfitting The problem of overfitting . Journal of chemical information and computer sciences 44 1 1--12

    work page 2004

Show all 39 references
  1. [9]

    , Shen, Y

    hu2021lora APACrefauthors Hu, E J. , Shen, Y. , Wallis, P. , Allen-Zhu, Z. , Li, Y. , Wang, S. Chen, W. APACrefauthors \ 2021 . Lora: Low-rank adaptation of large language models Lora: Low-rank adaptation of large language models . arXiv preprint arXiv:2106.09685

    2021 arXiv

  2. [10]

    , Liu, Z

    hundal2024soccer APACrefauthors Hundal, R S. , Liu, Z. , Wadhwa, B. , Hou, Z. , Jiang, K. \ Dong, J S. APACrefauthors \ 2024 . Soccer Strategy Analytics Using Probabilistic Model Checkers Soccer strategy analytics using probabilistic model checkers . International Sports Analy...

    2024

  3. [11]

    , Izadi, M

    jiang2020deep APACrefauthors Jiang, K. , Izadi, M. , Liu, Z. \ Dong, J S. APACrefauthors \ 2020 . Deep learning application in broadcast tennis video annotation Deep learning application in broadcast tennis video annotation . 2020 25th International Conference on Engineering o...

    2020

  4. [12]

    jiang2023court APACrefauthors Jiang, K. , Li, J. , Liu, Z. \ Dong, C. APACrefauthors \ 2023 . Court detection using masked perspective fields network Court detection using masked perspective fields network . 2023 IEEE 28th Pacific Rim International Symposium on Dependable Comp...

    2023

  5. [13]

    , Liu, Z

    jiang2024tracking APACrefauthors Jiang, K. , Liu, Z. , Wu, Q. , Ma, M. \ Dong, J S. APACrefauthors \ 2024 . Tracking Small and Fast Moving Ball in Broadcast Videos Using Transfer Learning and the Enhanced Interactive Multi-motion Model Tracking small and fast moving ball in br...

    2024

  6. [14]

    , Vidal, R

    lea2016learning APACrefauthors Lea, C. , Vidal, R. \ Hager, G D. APACrefauthors \ 2016 . Learning convolutional action primitives for fine-grained action recognition Learning convolutional action primitives for fine-grained action recognition . 2016 IEEE international conferen...

    2016

  7. [15]

    , Liu, Y

    li2024aria APACrefauthors Li, D. , Liu, Y. , Wu, H. , Wang, Y. , Shen, Z. , Qu, B. Li, J. APACrefauthors \ 2024 . Aria: An Open Multimodal Native Mixture-of-Experts Model Aria: An open multimodal native mixture-of-experts model . arXiv preprint arXiv:2410.05993

    2024 arXiv

  8. [16]

    , Wang, Y

    li2024mvbench APACrefauthors Li, K. , Wang, Y. , He, Y. , Li, Y. , Wang, Y. , Liu, Y. others APACrefauthors \ 2024 . Mvbench: A comprehensive multi-modal video understanding benchmark Mvbench: A comprehensive multi-modal video understanding benchmark . Proceedings of the IEEE/...

    2024

  9. [17]

    , Chen, K

    li2024surveying APACrefauthors Li, M. , Chen, K. , Bi, Z. , Liu, M. , Peng, B. , Niu, Q. others APACrefauthors \ 2024 . Surveying the mllm landscape: A meta-review of current surveys Surveying the mllm landscape: A meta-review of current surveys . arXiv preprint arXiv:2409.18991

    2024

  10. [18]

    , Zhu, B

    lin2023video APACrefauthors Lin, B. , Zhu, B. , Ye, Y. , Ning, M. , Jin, P. \ Yuan, L. APACrefauthors \ 2023 . Video-llava: Learning united visual representation by alignment before projection Video-llava: Learning united visual representation by alignment before projection . ...

    2023 arXiv

  11. [19]

    , Dong, C

    liu2025analyzing APACrefauthors Liu, Z. , Dong, C. , Chen, J W. , Jiang, A M J. , Chen, G. , Shaikh, A F. Dong, J S. APACrefauthors \ 2025 . Analyzing the Formation Strategy in Tennis Doubles Game Analyzing the formation strategy in tennis doubles game . SN Computer Science 6 2 100

    2025

  12. [20]

    , Dong, C

    liu2024exploring APACrefauthors Liu, Z. , Dong, C. , Wang, C. , Dong, T Y. \ Jiang, K. APACrefauthors \ 2024 . Exploring team strategy dynamics in tennis doubles matches Exploring team strategy dynamics in tennis doubles matches . International Sports Analytics Conference and ...

    2024

  13. [21]

    , Dong, Y

    liu2024oryx APACrefauthors Liu, Z. , Dong, Y. , Liu, Z. , Hu, W. , Lu, J. \ Rao, Y. APACrefauthors \ 2024 . Oryx MLLM: On-Demand Spatial-Temporal Understanding at Arbitrary Resolution Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution . arXiv preprint ...

    2024 arXiv

  14. [22]

    , Durrani, M

    liu2024strategy APACrefauthors Liu, Z. , Durrani, M. , Xuan, L Y. , Simon, J F. \ Deon, T Y F. APACrefauthors \ 2024 . Strategy Analysis in NFL Using Probabilistic Reasoning Strategy analysis in nfl using probabilistic reasoning . International Sports Analytics Conference and ...

    2024

  15. [23]

    , Guo, J

    liu2023recognizing APACrefauthors Liu, Z. , Guo, J. , Wang, M. , Wang, R. , Jiang, K. \ Dong, J S. APACrefauthors \ 2023 . Recognizing a sequence of events from tennis video clips: addressing timestep identification and subtle class differences Recognizing a sequence of events...

    2023

  16. [24]

    , Jiang, K

    liu2023sports APACrefauthors Liu, Z. , Jiang, K. \ Dong, J S. APACrefauthors \ 2023 . Sports injury prediction in professional tennis Sports injury prediction in professional tennis . 2023 IEEE 28th Pacific Rim International Symposium on Dependable Computing (PRDC) 2023 ieee 2...

    2023

  17. [25]

    , Jiang, K

    liu2023insight APACrefauthors Liu, Z. , Jiang, K. , Hou, Z. , Lin, Y. \ Dong, J S. APACrefauthors \ 2023 . Insight analysis for tennis strategy and tactics Insight analysis for tennis strategy and tactics . 2023 IEEE International Conference on Data Mining (ICDM) 2023 ieee int...

    2023

  18. [26]

    , Jiang, K

    liu2025f APACrefauthors Liu, Z. , Jiang, K. , Ma, M. , Hou, Z. , Lin, Y. \ Dong, J S. APACrefauthors \ 2025 . F ^3 S et: Towards Analyzing Fast, Frequent, and Fine-grained Events from Videos F ^3 S et: Towards analyzing fast, frequent, and fine-grained events from videos . arX...

    2025 arXiv

  19. [27]

    liu2024pcsp APACrefauthors Liu, Z. , Ma, M. , Jiang, K. , Hou, Z. , Shi, L. \ Dong, J S. APACrefauthors \ 2024 . Pcsp\# denotational semantics with an application in sports analytics Pcsp\# denotational semantics with an application in sports analytics . The Application of For...

    2024

  20. [28]

    , Ghosh, S

    madan2024mip APACrefauthors Madan, S. , Ghosh, S. , Sookha, L R. , Ganaie, M. , Subramanian, R. , Dhall, A. \ Gedeon, T. APACrefauthors \ 2024 . MIP-GAF: A MLLM-annotated Benchmark for Most Important Person Localization and Group Context Understanding Mip-gaf: A mllm-annotated...

    2024 arXiv

  21. [29]

    , Akshulakov, R

    mangalam2023egoschema APACrefauthors Mangalam, K. , Akshulakov, R. \ Malik, J. APACrefauthors \ 2023 . Egoschema: A diagnostic benchmark for very long-form video language understanding Egoschema: A diagnostic benchmark for very long-form video language understanding . Advances...

    2023

  22. [30]

    , Aljundi, R

    panos2024imperfect APACrefauthors Panos, A. , Aljundi, R. , Reino, D O. \ Turner, R E. APACrefauthors \ 2024 . Imperfect Vision Encoders: Efficient and Robust Tuning for Vision-Language Models Imperfect vision encoders: Efficient and robust tuning for vision-language models . ...

    2024 arXiv

  23. [31]

    , Smaira, L

    patraucean2024perception APACrefauthors Patraucean, V. , Smaira, L. , Gupta, A. , Recasens, A. , Markeeva, L. , Banarse, D. others APACrefauthors \ 2024 . Perception test: A diagnostic benchmark for multimodal video models Perception test: A diagnostic benchmark for multimodal...

    2024

  24. [32]

    APACrefauthors \ 2024

    qiu2024impact APACrefauthors Qiu, Y. APACrefauthors \ 2024 . The impact of llm hallucinations on motor skill learning: A case study in badminton The impact of llm hallucinations on motor skill learning: A case study in badminton . IEEE Access

    2024

  25. [33]

    , Kim, J W

    radford2021learning APACrefauthors Radford, A. , Kim, J W. , Hallacy, C. , Ramesh, A. , Goh, G. , Agarwal, S. others APACrefauthors \ 2021 . Learning transferable visual models from natural language supervision Learning transferable visual models from natural language supervis...

    2021

  26. [34]

    , Chen, S

    ren2023testa APACrefauthors Ren, S. , Chen, S. , Li, S. , Sun, X. \ Hou, L. APACrefauthors \ 2023 . TESTA: Temporal-spatial token aggregation for long-form video-language understanding Testa: Temporal-spatial token aggregation for long-form video-language understanding . arXiv...

    2023 arXiv

  27. [35]

    sun2023comparative APACrefauthors Sun, X. , Ji, Y. , Ma, B. \ Li, X. APACrefauthors \ 2023 . A comparative study between full-parameter and lora-based fine-tuning on chinese instruction data for instruction following large language model A comparative study between full-parame...

    2023 arXiv

  28. [36]

    wu2024longvideobench APACrefauthors Wu, H. , Li, D. , Chen, B. \ Li, J. APACrefauthors \ 2024 . Longvideobench: A benchmark for long-context interleaved video-language understanding Longvideobench: A benchmark for long-context interleaved video-language understanding . arXiv p...

    2024 arXiv

  29. [37]

    yu2019activitynet APACrefauthors Yu, Z. , Xu, D. , Yu, J. , Yu, T. , Zhao, Z. , Zhuang, Y. \ Tao, D. APACrefauthors \ 2019 . Activitynet-qa: A dataset for understanding complex web videos via question answering Activitynet-qa: A dataset for understanding complex web videos via...

    2019

  30. [38]

    , Tong, S

    zhai2024investigating APACrefauthors Zhai, Y. , Tong, S. , Li, X. , Cai, M. , Qu, Q. , Lee, Y J. \ Ma, Y. APACrefauthors \ 2024 . Investigating the catastrophic forgetting in multimodal large language model fine-tuning Investigating the catastrophic forgetting in multimodal la...

    2024

  31. [39]

    , Zang, Q

    zhu2024lime APACrefauthors Zhu, K. , Zang, Q. , Jia, S. , Wu, S. , Fang, F. , Li, Y. others APACrefauthors \ 2024 . LIME: Less Is More for MLLM Evaluation Lime: Less is more for mllm evaluation . arXiv preprint arXiv:2409.06851

    2024 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.