Pith. sign in

REVIEW 4 major objections 5 minor 35 references

EgoSafe-Bench shows vision-language models can describe scenes well but fail to infer blind-spot threats, despite humans solving the task.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

A new egocentric safety benchmark shows current video-language models can describe scenes well but fail at multi-step causal reasoning about blind spots and covert actions.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection EgoSafe-Bench fills a real gap in first-person safety reasoning evaluation, but the central 'deterministically solvable from context' claim rests on Gemini-generated ground truth that the paper never validates. the 4 major comments →

arxiv 2607.26518 v2 pith:QMYAQVIL submitted 2026-07-29 cs.CV

EgoSafe: A First-Person Mobile-Captured Benchmark for Visual Safety Understanding

classification cs.CV
keywords egocentric video understandingvisual safety benchmarkhierarchical reasoning evaluationperception-reasoning decouplingblind-spot reasoninglarge vision-language modelsoverconfidence calibrationvideo question answering
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

EgoSafe-Bench tries to establish that current large vision-language models lack genuine evidence-based reasoning in first-person safety scenarios, even when their descriptive performance looks strong. The authors built 3,000 self-captured egocentric video clips paired with 12,000 question-answer items organized into a four-tier Hierarchical Reasoning Evaluation (HRE) that forces a reasoning chain from visible anchoring to blind-spot deduction to intent inference. Their central finding is a perception-reasoning decoupling: models such as Qwen3-VL reach high semantic-similarity scores (70.43) but drop sharply on reasoning (40.25), with the steepest decline at the blind-spot and covert-action stages, while humans reach 89.93 on reasoning. A second finding is that models are often overconfident when hallucinating under partial visibility, which matters for real-world safety where a cautious 'I don't know' is safer than a confident wrong answer.

Core claim

On its own terms, the central discovery is the perception-reasoning decoupling: state-of-the-art LVLMs show high semantic and factual alignment on egocentric safety clips (e.g., Qwen3-VL SemSim 70.43, InternVL 3.5 KFC 66.29) yet their reasoning scores collapse when they must construct a hidden blind spot, deduce a covert action occurring there, and trace it to a causal outcome (Qwen3-VL Reasoning 40.25, InternVL 3.5 52.92). The steepest performance drop occurs consistently at Levels 2 and 3 of the HRE chain, and some models partially rebound at Level 4 by guessing the outcome without the intermediate reasoning steps, evidence that the chain design exposes shortcut predictions. The human base

What carries the argument

The Hierarchical Reasoning Evaluation (HRE) protocol: a four-turn QA chain (Initial Feature Anchoring, Dynamic Blind-Spot Construction, Covert Action Deduction, and Causal Result Qualitative Analysis) that forces a model to reason about invisible events in sequence, penalizing shortcut answers. The associated Reasoning Score, which weights causal consistency between Levels 2 and 4, and the Expected Calibration Error (ECE) metric for overconfidence are the instruments that expose the perception-reasoning gap and the models' tendency to issue confident hallucinations under partial observability.

Load-bearing premise

The solvability claim rides on the assumption that the machine-generated, human-verified reasoning chains are the uniquely correct inferences for each clip, not artifacts of how the questions or distractors were built.

What would settle it

Have independent annotators re-derive the ground-truth answers for 300 randomly chosen clips directly from the raw videos, without seeing the published QA chains; if their answers disagree with the published chains on more than a small fraction of samples, the claim that the task is deterministically solvable from context is undermined.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Current LVLMs are not reliable for proactive egocentric harm prevention: they can describe a scene but will often hallucinate what happens in blind spots, frequently with high confidence.
  • Descriptive benchmarks (SemSim, KFC) overstate safety competence; reasoning-grounded evaluation like HRE is necessary to reveal whether a model truly understands an event sequence.
  • The HRE ablation shows that removing any reasoning level degrades performance, with the largest drop when Level 2 (blind-spot construction) is removed, identifying this step as the pivotal cognitive bottleneck.
  • Egocentric noise degrades temporal reasoning before spatial recognition: even the most robust evaluated model, Qwen3-VL, loses 31.2% of its Reasoning Score under the controlled degradations.
  • The high ECE values (up to 92.04) imply that safety-critical deployments of current LVLMs need uncertainty-aware mechanisms that can say 'I don't know' instead of asserting a confident guess.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the perception-reasoning decoupling is a general phenomenon, it likely appears in other partially observable domains, such as medical imaging under occlusion or autonomous driving with obstructed views, where high perception accuracy masks weak causal inference.
  • A testable extension: explicitly prompting a model to list what is occluded and then hypothesize events in the blind spot might close the reasoning gap without improving perception; observing this would confirm the deficit is at the reasoning level rather than feature extraction.
  • The human baseline's near-perfect agreement suggests the benchmark could double as a training source for reasoning-consistent instruction data, but only if the ground-truth chains are validated against independent human annotation rather than machine-generated consensus.
  • The sharp drop at Level 2 offers a cheap diagnostic: a model's blind-spot-construction score alone could serve as a proxy for overall safety reasoning capability in egocentric settings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces EgoSafe-Bench, a first-person (egocentric) video safety benchmark containing 3,000 video clips derived from 250 staged source videos and 11 controlled degradations, paired with 12,000 QA samples organized by a four-level Hierarchical Reasoning Evaluation (HRE) protocol: Initial Feature Anchoring, Dynamic Blind Spot Construction, Covert Action Deduction, and Causal Result Qualitative Analysis. The authors evaluate several state-of-the-art LVLMs (Qwen3-VL, VideoLLaMA 3, InternVL 3.5, MiniCPM-V 4.5, LLaVA-Video, GPT-5) in a zero-shot setting using semantic similarity, key-fact coverage, a composite Reasoning Score, and expected calibration error. They report a sharp "perception-reasoning decoupling": high descriptive scores but much lower reasoning scores, with the largest drop at the blind-spot/covert-action stages. A human baseline (Reasoning 89.93) is used to argue that the HRE task is deterministically solvable from context, implying that models' low reasoning scores reflect genuine cognitive deficits rather than data ambiguity. Robustness experiments show large performance drops under lighting, motion, and resolution degradations, and calibration analysis indicates severe overconfidence in most models.

Significance. If the central result holds, EgoSafe-Bench would fill a real gap: existing safety benchmarks rely mostly on third-person surveillance and recognition-style labels, while real-world mobile safety reasoning requires inferring hidden interactions from partial egocentric evidence. The benchmark design is thoughtful in several respects: it includes realistic first-person noise, a systematic 11-way degradation protocol, privacy-preserving de-identification, a human baseline, and a reasoning-chain structure that penalizes shortcut predictions. The machine-checked aspects—the dataset release intent, the human-in-the-loop verification pipeline, and the explicit degradation protocol—are strengths. However, the paper's most load-bearing claim—that the task is deterministically solvable and that low model reasoning scores reveal a reasoning deficit—rests on the correctness and uniqueness of the Gemini-3-Pro-generated ground-truth QA chains. The current manuscript does not provide enough evidence to establish this, and the evaluation design raises concerns about judge bias. These issues are fixable with additional validation and analysis, but they are central rather than cosmetic.

major comments (4)
  1. [§3.2, HRE Annotation] The ground-truth QA chains are generated by Gemini-3-Pro and reviewed by "independent verifiers," but the only reported agreement is on event-type classification. There is no inter-annotator agreement on the actual QA answers, no count of rejected/rewritten chains, and no description of how distractors were constructed or checked for leakage. This matters because the human baseline (Reasoning 89.93) is used to claim the task is "deterministically solvable from context." If verifier consensus simply ratifies Gemini's preferred inferences, the human baseline may measure question artifacts or scripted priors rather than visual blind-spot reasoning. Please provide per-stage agreement statistics, a breakdown of rejections, and a distractor-leakage analysis (e.g., whether the correct answer is identifiable without the video).
  2. [§4.1, Automated Evaluation Framework (GPT-5 Judge)] GPT-5 is used as the gold-standard judge for SemSim, KFC, and Reasoning Score, while also being one of the evaluated models. This creates a conflict: GPT-5's high Reasoning Score (74.60) may partly reflect judge self-preference rather than superior reasoning. The paper should either use an independent judge not in the evaluated set, report agreement between GPT-5 and human judgments on a sample, or otherwise demonstrate that the judge does not favor GPT-5's output style. Without this, the comparative ranking between GPT-5 and open-source models is not fully trustworthy.
  3. [§4.2, Human Performance Ceiling] The human baseline protocol is not directly comparable to the model evaluation. Humans answered single-choice questions, while models generated free-text responses scored by GPT-5. The paper does not explain how human choices were converted into SemSim, KFC, and Reasoning scores, nor whether human confidence was collected in a way compatible with the ECE computation. To support the claim that humans show zero perception-reasoning decoupling, the same metric pipeline should be applied to human answers, or the metric definitions should be adjusted to apply to multiple-choice responses.
  4. [§4.4 and Table 4] Table 4 reports Reasoning Scores for only 8 of the 11 degradation conditions described in §3.1 and Table 5; three illumination conditions are omitted. The paper states no scores are inferred, which is transparent, but the robustness conclusion that "Lighting and Motion degradations induce the most significant failures" is based on partial data. The authors should provide the complete 11-condition table, or explicitly limit the robustness claims to the reported subset and explain why the three lighting conditions are missing.
minor comments (5)
  1. [§4.1, Eq. (1)] The ECE formula has a typographical issue: "conf(B m)" should be "conf(B_m)" to match the notation for the m-th bin. Also, the equation is unnumbered in the text; please number it for ease of reference.
  2. [§3.3, Figure 3] The text refers to a "sunburst chart in Figure 3," but the figure as described in the caption appears to contain two subfigures (a) and (b) without a visible sunburst. Please ensure the figure and caption match the text.
  3. [§4.5] The phrase "Next-QA" should be "NExT-QA" for consistency with the reference list and standard naming.
  4. [Appendix C] The HRE ablation is conducted only on Qwen3-VL. The paper acknowledges this, but the conclusion that "removing any single level causes an average drop of 11.62 points" should be clearly framed as a single-model observation, not a general property of HRE.
  5. [§3.2] The four HRE levels are described but the exact weighting of the Reasoning Score (e.g., how Level 2 and Level 4 are "heavily weighted") is not specified. Please provide the precise scoring formula so the metric is reproducible.

Circularity Check

0 steps flagged

No significant circularity; benchmark scores are empirical measurements against externally generated ground truth, with no fitted parameter or self-citation chain.

full rationale

EgoSafe-Bench's central claims (perception-reasoning decoupling, human solvability) rest on zero-shot evaluation of external LVLMs against QA chains produced by Gemini-3-Pro and reviewed by independent verifiers. The evaluated models (Qwen3-VL, InternVL 3.5, VideoLLaMA 3, MiniCPM-V, LLaVA-Video, GPT-5) are not trained or fitted on EgoSafe; no benchmark parameter is estimated from model outputs and then renamed a prediction. The HRE four-level reasoning drop is measured per-level accuracy, not imposed by the metric definition. The human baseline (89.93 Reasoning) is an independent empirical check that the questions are answerable, not a definitional guarantee. The paper's own caveats (Table 4 reporting 8 of 11 degradation conditions; Appendix C ablating only Qwen3-VL with 'no results are inferred for the remaining models') are transparency about incomplete coverage, not circular reasoning. The use of GPT-5 as both a judge and an evaluated model could raise evaluation-bias concerns, but the ground-truth chains are fixed and external, so scores do not reduce to the judge's own outputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central results rest on the staged-data pipeline, LLM-generated ground truth, LLM-as-judge scoring, and a hand-designed reasoning hierarchy. The paper does not report inter-annotator agreement on QA chains, judge validation against humans, or a precise Reasoning Score formula. The counts are: 4 unspecified human/AI design choices (level definitions, weighting, judge rubric, distractor mapping), 5 domain assumptions, and 2 new artifacts (benchmark + protocol) that are staged and/or LLM-generated but publicly released as falsifiable handles. Most of the experimental quantities are not actually fitted to data in a statistical sense, but the protocol itself is a parameterized evaluation design whose parameters are tuned by the authors.

free parameters (4)
  • HRE four-stage level definitions = Anchoring, Blind Spot, Covert Action, Causal Outcome
    The four-stage chain is a hand-designed protocol; the paper gives no evidence that these four stages are the unique or optimal decomposition for safety reasoning, and the Reasoning Score weights "heavily" Levels 2 and 4 without a specified formula.
  • Reasoning Score weighting scheme = Not specified numerically
    Section 4.1 states the composite Reasoning Score "heavily weights the causal consistency between Blind Spot Construction (Level 2) and Causal Result (Level 4)" but no equation is given; different weightings would change every reported reasoning score.
  • GPT-5 judge prompt and criteria = Not released
    All SemSim, KFC, and Reasoning scores depend on the GPT-5 scoring prompt and rubric; without these, the numbers are not reproducible and threshold choices can shift model rankings.
  • Human distractor error mapping = Each distractor mapped to one evaluation dimension
    The human SemSim/KFC/Reasoning scores are derived from single-choice distractor selections; the mapping from distractor to metric is described qualitatively and could inflate or deflate human performance relative to the open-ended model evaluation.
axioms (5)
  • domain assumption The 11 staged degradations preserve the same safety event semantics as the original clip.
    Each of the 11 variants is treated as a distinct clip with a valid QA chain; if extreme low light, blur, or flip makes the event genuinely indistinguishable, then the benchmark measures guessability, not reasoning.
  • domain assumption Gemini-3-Pro-generated QA chains, after unnamed human verifier review, are factually correct and uniquely answerable.
    Section 3.2: the chains are instantiated by Gemini-3-Pro and filtered by verifiers who agree on event-type classification; no inter-annotator agreement statistics are reported. If the QA chains encode ambiguous or template answers, the reasoning scores are not a clean measure of reasoning.
  • domain assumption GPT-5 is a reliable gold-standard automatic judge for reasoning quality and hallucination detection.
    Section 4.1 introduces GPT-5 as "gold-standard judge"; the paper does not measure judge agreement with human annotators on a validation subset, so judge bias directly contaminates every reported reasoning score.
  • domain assumption Human single-choice performance on the same 12,000 samples is comparable to open-ended zero-shot model generations.
    Humans answered single-choice assessments after one viewing, whereas models generate free text evaluated by GPT-5. Different response formats can inflate/deflate the human-model gap independent of reasoning ability.
  • domain assumption The 250 staged source videos are representative of real-world first-person safety encounters.
    All scenarios are staged by the team using pre-written scripts in corridors, pathways, and entrances. Staged acting may not reproduce the distribution of real emergencies, and stated coverage of 45% indoor / 40% outdoor / 15% semi-open is a design choice, not an empirical fact.
invented entities (2)
  • EgoSafe-Bench dataset independent evidence
    purpose: Benchmark of 3,000 first-person video clips and 12,000 HRE QA samples for evaluating egocentric safety reasoning.
    The dataset itself provides falsifiable handles: it is released for public academic use, and any group can run models on it and check whether the decoupling reproduces. The independent evidence is the benchmark's public availability and the concrete quantitative predictions (model ranking pattern across HRE levels).
  • Hierarchical Reasoning Evaluation (HRE) four-tier QA chain independent evidence
    purpose: Evaluation protocol forcing models to show anchoring, blind-spot, covert-action, and causal-outcome reasoning.
    It yields falsifiable predictions (the ablation table: removing any level drops the Reasoning Score by ~11.62 on Qwen3-VL). Independent groups can re-run the HRE chain on other models, but the claimed evidence is currently limited to a single model.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of EgoSafe: A First-Person Mobile-Captured Benchmark for Visual Safety Understanding." pith.science (2026). https://pith.science/paper/QMYAQVIL

@misc{pith2026260726518,
  author       = {Pith},
  title        = {Pith review of: EgoSafe: A First-Person Mobile-Captured Benchmark for Visual Safety Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QMYAQVIL}},
  note         = {Machine review of arXiv:2607.26518}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reliable visual safety understanding in real-world scenarios demands more than just object recognition; it requires causal reasoning under epistemic uncertainty. While Large Vision-Language Models (LVLMs) demonstrate impressive semantic alignment on standard benchmarks, they often struggle to distinguish between superficial correlation and genuine forensic logic when grounded in the dynamic, partially observable nature of first-person experiences. Existing evaluations, dominated by third-person surveillance footage and binary classification metrics, fail to expose this cognitive gap. To address this, we introduce EgoSafe-Bench, a benchmark specifically designed to probe forensic reasoning in egocentric safety scenarios. It comprises 12,000 unique evaluation samples, generated by pairing each of the 3,000 video clips with a QA chain governed by our proposed Hierarchical Reasoning Evaluation (HRE) protocol. Unlike standard benchmarks, HRE mandates a rigorous reasoning trajectory from initial feature anchoring to blind-spot deduction and intent inference, thereby enforcing logical consistency and penalizing shortcut-based predictions. Extensive evaluations of state-of-the-art LVLMs (e.g., Qwen3-VL, Gemini, VideoLLaMA 3) reveal a significant perception-reasoning decoupling: models often achieve high descriptive scores but exhibit notable fragility in causal reasoning and logical closure. Our work provides both a challenging dataset and a systematic evaluation framework to foster the development of logically robust video understanding systems.

Figures

Figures reproduced from arXiv: 2607.26518 by Cen Chen, Hao Peng, Huiping Zhuang, Tianao Li, Tianquan Feng, Yuyun Chen, Ziqian Zeng.

Figure 1
Figure 1. Figure 1: From Recognition to Evidence-Based Reasoning. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the data generation and verification pipeline for EgoSafe [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comprehensive analysis of scene diversity. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: In-depth comparison of dataset capabilities and reasoning dimensions. (a) The radar chart evaluates benchmarks across five key metrics, illustrating that EgoSafe￾Bench provides a superior balance of image quality, robustness, and reasoning depth compared to traditional violence detection datasets. (b) Distribution of rea￾soning categories across violence detection benchmarks. While existing datasets pre￾do… view at source ↗
Figure 5
Figure 5. Figure 5: Performance degradation across reasoning levels. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Analysis of Perception-Reasoning Decoupling. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 14 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

  2. [2]

    arXiv preprint arXiv:2308.12966 (2023)

    Bai, J., Bai, S., Yang, S., Wang, S., Tan, S., Wang, P., Lin, J., Zhou, C., Zhou, J.: Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966 (2023)

  3. [3]

    arXiv preprint arXiv:2502.13923 (2025)

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923 (2025)

  4. [4]

    arXiv preprint arXiv:2501.13106 (2025)

    Zhang, B., Li, K., Cheng, Z., Hu, Z., Yuan, Y., Chen, G., Leng, S., Jiang, Y., Zhang, H., Li, X., Jin, P., Zhang, W., Wang, F., Bing, L., Zhao, D.: Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106 (2025)

  5. [5]

    In: 2020 25th International Conference on Pattern Recognition (ICPR)

    Cheng, M., Cai, K., Li, M.: Rwf-2000: An open large scale video database for violence detection. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 4183–4190 (2021)

  6. [6]

    arXiv preprint arXiv:2406.07476 (2024)

    Cheng, Z., Leng, S., Zhang, H., Xin, Y., Li, X., Chen, G., Zhu, Y., Zhang, W., Luo, Z., Zhao, D., Bing, L.: Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476 (2024)

  7. [7]

    IEEE Transactions on Pattern Anal- ysis and Machine Intelligence43(11), 4125–4141 (2021)

    Damen, D., Doughty, H., Farinella, G.M., Fidler, S., Furnari, A., Kazakos, E., Moltisanti, D., Munro, J., Perrett, T., Price, W., Wray, M.: The epic-kitchens dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Anal- ysis and Machine Intelligence43(11), 4125–4141 (2021)

  8. [8]

    In: Proceedings of the First Workshop on Comparative Performance Evaluation: From Rules to Language Models

    Ditchfield-Ogle, T., Mitkov, R.: A comparative study of vision transformers and multimodal language models for violence detection in videos. In: Proceedings of the First Workshop on Comparative Performance Evaluation: From Rules to Language Models. pp. 10–20 (2025)

  9. [9]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Grauman, K., Westbury, A., Byrne, E., Chavis, Z., Furnari, A., Girdhar, R., Ham- burger, J., Jiang, H., Liu, M., Liu, X., Martin, M., Nagarajan, T., Radosavovic, I., Ramakrishnan, S.K., Ryan, F., Sharma, J., Wray, M., Xu, M., Xu, E.Z., Zhao, C., Bansal, S., Batra, D., Cartillier, V., Crane, S., Do, T., Doulaty, M., Erapalli, A., 18 EgoSafe Feichtenhofer, ...

  10. [10]

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

    Grauman, K., Westbury, A., Torresani, L., Kitani, K., Malik, J., Afouras, T., Ashutosh, K., Baiyya, V., Bansal, S., Boote, B., et al.: Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19383–19400 (2024)

  11. [11]

    In: 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019

    Hendrycks, D., Dietterich, T.G.: Benchmarking neural network robustness to com- mon corruptions and perturbations. In: 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenRe- view.net (2019)

  12. [12]

    arXiv preprint arXiv:2506.05372 (2025)

    Kollias, D., Senadeera, D.C., Zheng, J., Yadav, K.K., Slabaugh, G., Awais, M., Yang, X.: Dvd: A comprehensive dataset for advancing violence detection in real- world scenarios. arXiv preprint arXiv:2506.05372 (2025)

  13. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)

    Li, J., Niu, L., Zhang, L.: From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2022)

  14. [14]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Liu, Y., Wang, Z., Xu, J., Chen, G., Luo, P., Wang, L., Qiao, Y.: Mvbench: A comprehensive multi-modal video under- standing benchmark. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22195–22206 (June 2024)

  15. [15]

    arXiv preprint arXiv:2311.10122 (2023)

    Lin, B., Zhu, B., Ye, Y., Ning, M., Jin, P., Yuan, L.: Video-llava: Learn- ing united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122 (2023)

  16. [16]

    In: Ku, L.W., Martins, A., Srikumar, V

    Maaz, M., Rasheed, H., Khan, S., Khan, F.: Video-ChatGPT: Towards detailed video understanding via large vision and language models. In: Ku, L.W., Martins, A., Srikumar, V. (eds.) Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Compu- tational Linguistics, Bangkok, Thailand (...

  17. [17]

    In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics

    Papineni, K., Roukos, S., Ward, T., Zhu, W.J.: Bleu: a method for automatic evaluation of machine translation. In: Proceedings of the 40th annual meeting of the Association for Computational Linguistics. pp. 311–318 (2002)

  18. [18]

    Soft Computing24(15), 11007–11017 (2020)

    Pujol, F.A., Mora, H., Pertegal, M.L.: A soft computing approach to violence detection in social media for smart cities. Soft Computing24(15), 11007–11017 (2020)

  19. [19]

    Multimedia Tools and Applica- tions83(22), 62107–62158 (2024)

    Sabha, A., Selwal, A.: Towards machine vision-based video analysis in smart cities: a survey, framework, applications and open issues. Multimedia Tools and Applica- tions83(22), 62107–62158 (2024)

  20. [20]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) EgoSafe 19

    Sultani, W., Chen, C., Shah, M.: Real-world anomaly detection in surveillance videos. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) EgoSafe 19

  21. [21]

    arXiv preprint arXiv:2312.11805 (2023)

    Team, G., Anil, R., Borgeaud, S., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., Millican, K., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023)

  22. [22]

    arXiv preprint arXiv:2403.05530 (2024)

    Team, G., Georgiev, P., Lei, V.I., Burnell, R., Bai, L., Gulati, A., Tanzer, G., Vin- cent,D.,Pan,Z.,Wang,S.,etal.:Gemini1.5:Unlockingmultimodalunderstanding across millions of tokens of context. arXiv preprint arXiv:2403.05530 (2024)

  23. [23]

    Team, Q.: Qwen3 technical report (2025)

  24. [24]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Vedantam, R., Lawrence Zitnick, C., Parikh, D.: Cider: Consensus-based image description evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4566–4575 (2015)

  25. [25]

    arXiv preprint arXiv:2409.12191 (2024)

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., Fan, Y., Dang, K., Du, M., Ren, X., Men, R., Liu, D., Zhou, C., Zhou, J., Lin, J.: Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191 (2024)

  26. [26]

    arXiv preprint arXiv:2508.18265 (2025)

    Wang, W., Gao, Z., Gu, L., Pu, H., Cui, L., Wei, X., Liu, Z., Jing, L., Ye, S., Shao, J., et al.: Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265 (2025)

  27. [27]

    arXiv preprint arXiv:2411.08380 (2024)

    Wang,X.,Zhao,K.,Liu,F.,Wang,J.,Zhao,G.,Bao,X.,Zhu,Z.,Zhang,Y.,Wang, X.: Egovid-5m: A large-scale video-action dataset for egocentric video generation. arXiv preprint arXiv:2411.08380 (2024)

  28. [28]

    In: European Conference on Computer Vision (ECCV) (2020)

    Wu, P., Liu, j., Shi, Y., Sun, Y., Shao, F., Wu, Z., Yang, Z.: Not only look, but also listen: Learning multimodal violence detection under weak supervision. In: European Conference on Computer Vision (ECCV) (2020)

  29. [29]

    IEEE Transactions on Multimedia26, 8557–8568 (2024)

    Xiang, T., Pan, H., Nan, Z.: Video violence rating: A large-scale public database and a multimodal rating model. IEEE Transactions on Multimedia26, 8557–8568 (2024)

  30. [30]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Xiao,J.,Shang,X.,Yao,A.,Chua,T.S.:Next-qa:Nextphaseofquestion-answering to explaining temporal actions. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9777–9786 (June 2021)

  31. [31]

    CoRRabs/2110.06513(2021)

    Yi,C.,Yang,S.,Li,H.,Tan,Y.,Kot,A.C.:Benchmarkingtherobustnessofspatial- temporal models against corruptions. CoRRabs/2110.06513(2021)

  32. [32]

    Yu, T., Wang, Z., Wang, C., Huang, F., Ma, W., He, Z., Cai, T., Chen, W., Huang, Y., Zhao, Y., Xu, B., Cui, J., Xu, Y., Ruan, L., Zhang, L., Liu, H., Tang, J., Liu, H., Guo, Q., Hu, W., He, B., Zhou, J., Cai, J., Qi, J., Guo, Z., Chen, C., Zeng, G., Li, Y., Cui, G., Ding, N., Han, X., Yao, Y., Liu, Z., Sun, M.: Minicpm-v 4.5: Cooking efficient mllms via a...

  33. [33]

    arXiv preprint arXiv:2306.02858 (2023)

    Zhang, H., Li, X., Bing, L.: Video-llama: An instruction-tuned audio-visual lan- guage model for video understanding. arXiv preprint arXiv:2306.02858 (2023)

  34. [34]

    Zhang, Y., Li, B., Liu, h., Lee, Y.j., Gui, L., Fu, D., Feng, J., Liu, Z., Li, C.: Llava-next: A strong zero-shot video understanding model (April 2024)

  35. [35]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)

    Zhu, C., Xiao, F., Alvarado, A., Babaei, Y., Hu, J., El-Mohri, H., Chang, S., Sumbaly, R., Yan, Z.: Egoobjects: A large-scale egocentric dataset for fine-grained object understanding. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023)

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.