Pith. sign in

REVIEW 4 major objections 6 minor 31 references

Grid-LOGAT: Grid Based Local and Global Area Transcription for Video Question Answering

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Grid-based visual prompting lets a zero-shot open-source VLM+LLM pipeline outperform comparable models on two VideoQA benchmarks, with a 24-point gain on localization questions.

desk verdict Solid incremental engineering with a plausible core idea, but the headline 24-point localization claim rests on an unreleased 120-question dataset and test-split layout selection. read the letter →

arxiv 2505.24371 v3 pith:EYSTVOLC submitted 2025-05-30 cs.CV cs.AI

Md Intisar Chowdhury , Kittinun Aukkapinyo , Hiroshi Fujimura , Joo Ann Woo , Wasu Wasusatein , Fadoua Ghourabi This is my paper
classification cs.CVcs.AI
keywords videoquestionansweringgrid-basedvisualpromptinggroundingzero-shotvision-languagemodellargelanguageprivacy-preservingtranscription
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

The paper proposes Grid-LoGAT, a two-stage video question answering system that first converts each video frame into text captions and then lets a large language model answer questions from those captions alone. Its central idea is to overlay a grid on each frame and ask the vision-language model to describe every grid cell separately, producing local captions in addition to a global caption. The authors claim this grid-based visual prompting recovers fine-grained spatial details that ordinary captioning misses, yielding a 24-point accuracy gain on localization-style questions they created from NExT-QA. On standard benchmarks, the system reaches 65.9% on NExT-QA and 50.11% on STAR-QA in a zero-shot setting, outperforming comparable open-source systems that use the same underlying models.

What carries the argument

The load-bearing mechanism is grid-based visual prompting: an n×m (chosen as 2×3) grid is overlaid on each video frame, and the VLM is prompted with a system message that names the six cells and asks for a caption of each cell plus an overall description. This forces the VLM to produce structured local captions T_l alongside the global caption T_g, which are concatenated element-wise per frame with timestamps to form the transcript T. The LLM then answers the question using only this transcript, not the raw video.

What would settle it

Construct a localization question set from NExT-QA videos independently of the authors, run the grid and no-grid variants on it, and compute the accuracy difference with confidence intervals. If the grid's advantage drops below the reported 24 points or fails to appear, the paper's attribution of the gain to the visual grid is falsified.

Watch

Extended reading notes

Core claim

The central claim is that local area transcription via grid-based visual prompting materially improves video question answering without any fine-tuning, and that the full local+global transcript ensemble beats prior multi-stage systems. On the paper's own NExT-QA-Loc evaluation, adding the rendered grid (with a grid-aware prompt) raises accuracy from 53.3% to 66.6%, and to 77.5% when both the grid and the grid prompt are used. The authors attribute this to the VLM being forced to attend to each cell, capturing object locations and relationships that a single dense caption would gloss over. On the official NExT-QA and STAR-QA test sets, the combined transcript reaches 65.9% and 50.11% accuracy respectively, surpassing IG-VLM, SeViLA, and another zero-shot pipeline using a 34B model.

Load-bearing premise

The whole localization improvement rests on 120 questions the authors wrote themselves, with no release or cross-validation; if those questions are phrased in a way that rewards cell-by-cell description, the 24-point gain could be an artifact of the test, not the grid.

Editorial extensions

If this is right

  • VideoQA can be performed without ever sending raw video frames to the cloud; only the text transcript crosses the network.
  • The system is modular: the VLM can be swapped for a smaller or larger open-source model at deployment time without changing the QA phase.
  • Grid prompting improves localization-style question accuracy by 24 points on the paper's custom NExT-QA-Loc set, with the rendered grid contributing most of the gain.
  • The full local+global transcript outperforms the global-only transcript by about 1.5 points on NExT-QA and 1.7 points on STAR-QA.

Reading between the lines

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

  • This suggests grid-based visual prompting could be applied beyond VideoQA, e.g., to any captioning or VQA pipeline that needs spatial grounding, as a zero-training drop-in improvement.
  • An immediate test is to see if the grid helps on small-object or attribute questions in standard image VQA benchmarks; the paper's mechanism implies it would, but that is our inference.
  • Because the QA phase receives only text transcripts, the system's privacy guarantee depends on the richness of the captions; a transcript that describes faces or license plates in detail could still leak sensitive information, so deployment should treat transcripts as sensitive too.
  • The authors do not report variance or multiple runs, so the 24-point gain lacks error bars; an external re-implementation on a released question set would establish whether the gain is robust.
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

4 major / 6 minor

Summary. The paper proposes Grid-LoGAT, a two-stage zero-shot VideoQA pipeline in which a frame-level VLM (LLaVA-1.6-7B) produces both a global caption for each frame and a set of local captions for cells of an overlaid n×m grid, and an open-source LLM (Llama-3.1-8B) answers multiple-choice questions from the concatenated text transcript. The authors report 65.9% accuracy on NExT-QA and 50.11% on STAR-QA, framing the system as a privacy-preserving solution because only text, not video, is sent to the cloud. The central claim is that grid-based visual prompting improves localization, supported by a 24-point gain over a non-grid baseline on an author-created 120-question set called NExT-QA-Loc. The paper also provides ablations separating local-only, global-only, and local+global transcript configurations.

Significance. If the reported numbers are robust, the contribution is a practical, modular, zero-shot recipe that uses only open-source models and keeps video data on the edge, which is a meaningful direction for privacy-sensitive VideoQA. The paper gives the prompt templates verbatim and evaluates on standard benchmarks, which supports reproducibility in principle, and the ablation design does separate the ensemble effect from the individual local and global captions. The main significance, however, is conditional on the NExT-QA-Loc evidence and on the statistical reliability of the small benchmark margins; the 24-point localization claim is the most striking result, but it is currently supported only by an unreleased, non-standard dataset.

major comments (4)
  1. [Section 4.3.2 / Table 4] The central localization claim, a gain of 53.3% to 77.5% (24.2 points), is measured only on 120 self-created NExT-QA-Loc questions that are not released, with no annotation protocol, no inter-annotator agreement, and no confidence intervals. Since these questions determine the headline claim that the grid improves grounding, the authors must either release the dataset with a clear construction protocol or replace this evidence with a public localization benchmark; without this, the reader cannot rule out that the questions were phrased to reward cell-label answers such as 'left/lower'.
  2. [Table 4 / Section 4.3.2] The same table shows that the grid-text prompt alone, without a rendered grid, already raises accuracy from 53.3% to 66.6%. Thus 13.3 of the 24.2 points come from the wording change, and only 10.9 points are attributable to the visual grid overlay. The paper's language in Section 4.3.2 ('It clearly shows the effectiveness of a visual grid') overstates the evidence; the prompt-wording effect should be explicitly separated and discussed as a limitation.
  3. [Table 2 / Section 4.3.1] The 2×3 grid layout is selected using the same test splits that later produce the final results, which is a form of test-set selection. On STAR-QA, the 2×2 variant yields 46.6% average accuracy, below the IG-VLM 7B baseline's 49.6%, and the 2×3 row duplicates the 'with grid (local)' row exactly, so the layout comparison does not test the final local+global configuration. The grid size should be chosen on a validation split or a pre-registered criterion, and the relation between the layout rows and the final ensembled system should be stated explicitly.
  4. [Table 1 / Section 4.3.1] The headline benchmark differences are small (65.9% vs. 63.1% on NExT-QA; 50.11% vs. 49.6% on STAR-QA relative to IG-VLM-7B), yet no repeated-run variance, confidence intervals, or significance tests are provided. With a VLM temperature of 1.0 and an LLM temperature of 0.1, single-run accuracy differences of this size may be within run-to-run noise; the paper should report the variance across at least a few independent runs or provide a statistical test.
minor comments (6)
  1. [Table 4] The 'Visual Prompt' column is ambiguous: it presumably indicates whether the rendered grid overlay was used, but the scenario names also contain 'grid prompt,' making it unclear what is changed between rows. Define the column explicitly in the caption.
  2. [Section 4.1] The text says 'greedy sampling with a temperature of 1.0'; greedy decoding normally implies temperature 0 or no sampling. Clarify whether the VLM uses greedy decoding or multinomial sampling.
  3. [Table 1] The 'Zero-Shot Inference' column uses 'O' and 'X' without a definition in the caption; please add 'O = zero-shot, X = fine-tuned' or similar.
  4. [Section 3.1] The concatenation symbol in Eq. (3), T = T_g ⊕ T_l, is not defined; the text should specify that this is an element-wise, per-frame concatenation with timestamps and frame numbers appended.
  5. [Abstract / throughout] The model name is rendered as 'LLaV A-1.6' in the abstract and in several places in the text; keep the spelling 'LLaVA-1.6' for consistency.
  6. [Figure 3] The caption says 'Ground-truth answers are depicted in red font,' but a black-and-white print or low-resolution version would not convey this; consider marking the correct answers with an explicit label or a symbol in the figure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark results are externally measured, and the self-authored localization set is a validation weakness, not a circular derivation.

full rationale

Grid-LoGAT is evaluated zero-shot on the official NExT-QA and STAR-QA test splits, with the headline accuracies (65.9% and 50.11%) obtained by exact-match against benchmark ground truth. The method's equations (T_g = F(v, P_g), T_l = F(v_g, P_l), T = T_g ⊕ T_l) define a processing pipeline; no fitted parameter or model output is fed back into the claim as if it were an independent prediction. The grid layout (2×3) is selected on the same test splits used for final comparison, which is a test-set-selection / overfitting concern, not a circular reduction: the reported accuracy is still a measured value rather than a quantity forced by construction. The 24-point localization gain is supported only by the authors' unreleased 120-question NExT-QA-Loc set, weakening external validation, but this is an auxiliary measurement set rather than a self-definitional or fitted-input relationship; the grid prompt's cell labels could in principle align with localization questions, yet the paper does not provide enough question text to exhibit such a reduction. No load-bearing self-citations, imported uniqueness theorems, or ansatz-smuggling-by-citation are present. The derivation chain is therefore self-contained against external benchmarks, and the auxiliary-dataset limitation belongs to correctness risk rather than circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The system's success rests on a set of unverified engineering assumptions: caption quality, frame-rate adequacy, LLM reasoning, and privacy-by-text. The only explicit tuned choice is the grid size, chosen on test data, plus hand-set sampling and token limits.

free parameters (3)
  • Grid layout (n x m) = 2x3
    Selected from ablations on NExT-QA and STAR-QA test sets (Table 2), making the reported final numbers the best of three tested layouts; the choice is made on the evaluation data itself.
  • Frame sampling rate = 1 FPS
    Frames extracted at one per second, set by design; temporal recall depends on it, and it is not ablated.
  • VLM token output limit = min 100, max 1000 tokens
    Chosen for all experiments; affects caption detail and is not ablated.
assumptions (4)
  • domain assumption Zero-shot dense captions from LLaVA-1.6-7B are accurate enough to support downstream QA.
    The entire pipeline relies on VLM caption quality; Section 3.1 uses F(v, P) without any verification of caption correctness.
  • domain assumption One frame per second captures the events needed to answer NExT-QA and STAR-QA questions.
    Section 3.1 sets |v| equal to video duration in seconds; no ablation of frame rate is reported.
  • domain assumption Language models can answer multiple-choice video questions from concatenated frame-level captions alone.
    The QA phase in Section 3.2 feeds only text transcript T to the LLM; temporal and causal reasoning is assumed to survive transcription.
  • domain assumption Text transcripts do not reveal visual identity, so edge-cloud processing preserves privacy.
    Stated in Sections 1 and 3; the paper acknowledges residual risk but uses it as a design motivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Grid-LOGAT: Grid Based Local and Global Area Transcription for Video Question Answering." pith.science (2026). https://pith.science/paper/EYSTVOLC

@misc{pith2026250524371,
  author       = {Pith},
  title        = {Pith review of: Grid-LOGAT: Grid Based Local and Global Area Transcription for Video Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EYSTVOLC}},
  note         = {Machine review of arXiv:2505.24371}
}
read the original abstract

In this paper, we propose a Grid-based Local and Global Area Transcription (Grid-LoGAT) system for Video Question Answering (VideoQA). The system operates in two phases. First, extracting text transcripts from video frames using a Vision-Language Model (VLM). Next, processing questions using these transcripts to generate answers through a Large Language Model (LLM). This design ensures image privacy by deploying the VLM on edge devices and the LLM in the cloud. To improve transcript quality, we propose grid-based visual prompting, which extracts intricate local details from each grid cell and integrates them with global information. Evaluation results show that Grid-LoGAT, using the open-source VLM (LLaVA-1.6-7B) and LLM (Llama-3.1-8B), outperforms state-of-the-art methods with similar baseline models on NExT-QA and STAR-QA datasets with an accuracy of 65.9% and 50.11% respectively. Additionally, our method surpasses the non-grid version by 24 points on localization-based questions we created using NExT-QA. (This paper is accepted by IEEE ICIP 2025.)

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 22 canonical work pages

  1. [1]

    INTRODUCTION In recent years, VideoQA task has gained significant attention for its potential applications in various real-world scenarios. For instance, by applying Vision Language Models (VLMs) to camera footage, it is possible to accurately comprehend the local environment and pro- vide appropriate responses for tasks such as robotic operations [1], as...

    work page Pith review arXiv 2025

  2. [2]

    Further research [5] extends the image VLM to support multi-frame and video data modalities

    LITERATURE REVIEW Video Question Answering(VideoQA) through LLaV A [10] is pro- posed to handle single-frame. Further research [5] extends the image VLM to support multi-frame and video data modalities. While these methods excel at video analysis tasks like summa- rization and question answering, they often face limitations in ex- tracting fine-grained in...

    work page

  3. [3]

    The first phase is the tran- scription generation phase

    METHOD The system operates in two key phases. The first phase is the tran- scription generation phase. For any given video input, we utilize a VLM to extract both local and global transcriptions at the frame level. In this research, a transcription is defined as a structured video caption, which is composed of a timestamp and its corresponding dense capti...

    work page

  4. [4]

    Baseline models We leveraged the LLaV A-1.6-7B VLM [23] to generate local and global transcriptions and the Llama-3.1-8B LLM [24] for the VideoQA task

    EXPERIMENTS 4.1. Baseline models We leveraged the LLaV A-1.6-7B VLM [23] to generate local and global transcriptions and the Llama-3.1-8B LLM [24] for the VideoQA task. Both models were chosen due to their open-source availability, reasonable parameter count, and fair comparison with methods using similar baseline models. For the VLM, we employed greedy s...

    work page

  5. [5]

    The system is inherently modular and built on open-source VLM and LLM

    CONCLUSION In this research, we propose a multi-model, multi-stage VideoQA system coupled with grid-based visual prompts for the VideoQA task, which we term Grid-LoGAT. The system is inherently modular and built on open-source VLM and LLM. It first executes the VLM, prompted with a visual grid prompt, to extract both local and global information. This inf...

    work page

  6. [6]

    MOKA: Open- vocabulary robotic manipulation through mark-based visual prompting,

    F. Liu, K. Fang, P. Abbeel, and S. Levine, “MOKA: Open- vocabulary robotic manipulation through mark-based visual prompting,” inFirst Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024, 2024

    work page 2024

  7. [7]

    VLAAD: Vision and language assistant for au- tonomous driving,

    S. Park, M. Lee, J. Kang, H. Choi, Y . Park, J. Cho, A. Lee, and D. Kim, “VLAAD: Vision and language assistant for au- tonomous driving,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 980–987

    work page 2024

  8. [8]

    Smart customer service in unmanned retail store enhanced by large language model,

    W. Wang, P. Zhang, C. Sun, and D. Feng, “Smart customer service in unmanned retail store enhanced by large language model,”Scientific Reports, vol. 14, no. 1, p. 19838, 2024

    work page 2024

Show all 31 references
  1. [9]

    Qwen2-VL: Enhancing vision- language model’s perception of the world at any resolution,

    P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Geet al., “Qwen2-VL: Enhancing vision- language model’s perception of the world at any resolution,” arXiv preprint arXiv:2409.12191, 2024

    2024 arXiv

  2. [10]

    LLaV A-OneVision: Easy visual task transfer,

    B. Li, Y . Zhang, D. Guo, R. Zhang, F. Li, H. Zhang, K. Zhang, P. Zhang, Y . Li, Z. Liu, and C. Li, “LLaV A-OneVision: Easy visual task transfer,”Transactions on Machine Learning Re- search, 2025

    2025

  3. [11]

    VideoChat: Chat-centric video under- standing,

    K. Li, Y . He, Y . Wang, Y . Li, W. Wang, P. Luo, Y . Wang, L. Wang, and Y . Qiao, “VideoChat: Chat-centric video under- standing,”arXiv preprint arXiv:2305.06355, 2023

    2023 arXiv

  4. [12]

    A simple LLM framework for long- range video question-answering,

    C. Zhang, T. Lu, M. M. Islam, Z. Wang, S. Yu, M. Bansal, and G. Bertasius, “A simple LLM framework for long- range video question-answering,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, E...

    2024

  5. [13]

    GPT-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “GPT-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  6. [14]

    Gem- ini: a family of highly capable multimodal models,

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millicanet al., “Gem- ini: a family of highly capable multimodal models,”arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  7. [15]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” Advances in neural information processing systems, vol. 36, 2024

    2024

  8. [16]

    An image grid can be worth a video: Zero-shot video question answering using a vlm,

    W. Kim, C. Choi, W. Lee, and W. Rhee, “An image grid can be worth a video: Zero-shot video question answering using a vlm,”IEEE Access, vol. 12, pp. 193 057–193 075, 2024

    2024

  9. [17]

    Self-chained image- language model for video localization and question answer- ing,

    S. Yu, J. Cho, P. Yadav, and M. Bansal, “Self-chained image- language model for video localization and question answer- ing,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  10. [18]

    Set-of- mark prompting unleashes extraordinary visual grounding in gpt-4v,

    J. Yang, H. Zhang, F. Li, X. Zou, C. Li, and J. Gao, “Set-of- mark prompting unleashes extraordinary visual grounding in gpt-4v,”arXiv preprint arXiv:2310.11441, 2023

    2023 arXiv

  11. [19]

    GeoChat: Grounded large vision-language model for remote sensing,

    K. Kuckreja, M. S. Danish, M. Naseer, A. Das, S. Khan, and F. S. Khan, “GeoChat: Grounded large vision-language model for remote sensing,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2024, pp. 27 831–27 840

    2024

  12. [20]

    Earth- Marker: A visual prompt learning framework for region-level and point-level remote sensing imagery comprehension,

    W. Zhang, M. Cai, T. Zhang, Y . Zhuang, and X. Mao, “Earth- Marker: A visual prompt learning framework for region-level and point-level remote sensing imagery comprehension,”arXiv preprint arXiv:2407.13596, 2024

    2024 arXiv

  13. [21]

    PIVOT: iterative visual prompting elicits actionable knowl- edge for vlms,

    S. Nasiriany, F. Xia, W. Yu, T. Xiao, J. Liang, I. Dasgupta, A. Xie, D. Driess, A. Wahid, Z. Xu, Q. Vuong, T. Zhang, T.- W. E. Lee, K.-H. Lee, P. Xu, S. Kirmani, Y . Zhu, A. Zeng, K. Hausman, N. Heess, C. Finn, S. Levine, and B. Ichter, “PIVOT: iterative visual prompting elici...

    2024

  14. [22]

    Open-vocabulary temporal action localization us- ing vlms,

    N. Wake, A. Kanehira, K. Sasabuchi, J. Takamatsu, and K. Ikeuchi, “Open-vocabulary temporal action localization us- ing vlms,”arXiv preprint arXiv:2408.17422, 2024

    2024 arXiv

  15. [23]

    Video graph trans- former for video question answering,

    J. Xiao, P. Zhou, T.-S. Chua, and S. Yan, “Video graph trans- former for video question answering,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 39–58

    2022

  16. [24]

    InternVideo: General video foundation models via generative and discriminative learning,

    Y . Wang, K. Li, Y . Li, Y . He, B. Huang, Z. Zhao, H. Zhang, J. Xu, Y . Liu, Z. Wanget al., “InternVideo: General video foundation models via generative and discriminative learning,” arXiv preprint arXiv:2212.03191, 2022

    2022 arXiv

  17. [25]

    MVBench: A comprehensive multi- modal video understanding benchmark,

    K. Li, Y . Wang, Y . He, Y . Li, Y . Wang, Y . Liu, Z. Wang, J. Xu, G. Chen, P. Luoet al., “MVBench: A comprehensive multi- modal video understanding benchmark,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2024, pp. 22 195–22 206

    2024

  18. [26]

    VISTA- LLAMA: Reducing hallucination in video language models via equal distance to visual tokens,

    F. Ma, X. Jin, H. Wang, Y . Xian, J. Feng, and Y . Yang, “VISTA- LLAMA: Reducing hallucination in video language models via equal distance to visual tokens,” in2024 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 13 151–13 160

    2024

  19. [27]

    CogAgent: A visual lan- guage model for gui agents,

    W. Hong, W. Wang, Q. Lv, J. Xu, W. Yu, J. Ji, Y . Wang, Z. Wang, Y . Dong, M. Dinget al., “CogAgent: A visual lan- guage model for gui agents,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 14 281–14 290

    2024

  20. [28]

    LLaV A-NeXT: Improved reasoning, OCR, and world knowledge,

    H. Liu, C. Li, Y . Li, B. Li, Y . Zhang, S. Shen, and Y . J. Lee, “LLaV A-NeXT: Improved reasoning, OCR, and world knowledge,” January 2024. [Online]. Available: https://llava-vl.github.io/blog/2024-01-30-llava-next/

    2024

  21. [29]

    The llama 3 herd of models,

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Yang, A. Fan et al., “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  22. [30]

    NExT-QA: Next phase of question-answering to explaining temporal actions,

    J. Xiao, X. Shang, A. Yao, and T.-S. Chua, “NExT-QA: Next phase of question-answering to explaining temporal actions,” inProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, 2021, pp. 9777–9786

    2021

  23. [31]

    STAR: A benchmark for situated reasoning in real-world videos,

    B. Wu, S. Yu, Z. Chen, J. B. Tenenbaum, and C. Gan, “STAR: A benchmark for situated reasoning in real-world videos,” in Thirty-fifth Conference on Neural Information Processing Sys- tems (NeurIPS), 2021

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.