Pith. sign in

REVIEW 4 major objections 5 minor 63 references

Flashback: Memory-Driven Zero-shot, Real-time Video Anomaly Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Flashback claims that moving the LLM offline and scoring video segments by retrieval against a pseudo-scene memory of captions makes video anomaly detection simultaneously zero-shot, real-time, and explainable, with state-of-the-art…

desk verdict Genuinely new memory-driven architecture for zero-shot VAD, but the zero-shot claim is softened by test-set hyperparameter tuning and an unmeasured memory coverage assumption. read the letter →

arxiv 2505.15205 v2 pith:5WJ44NFH submitted 2025-05-21 cs.CV

classification cs.CV
keywords videoanomalydetectionzero-shotlearningreal-timeinferencepseudo-scenememoryretrieval-basedscoringrepulsivepromptingscaledpenalizationexplainableAI
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

Flashback claims that video anomaly detection can be recast as a retrieval task: instead of running a large language model on every video segment, the system builds a one-time pseudo-scene memory of normal and anomalous captions offline, then at inference simply matches each segment's embedding against that memory. The paper argues this design achieves all three properties that earlier systems traded off—zero-shot deployment, genuine real-time speed (under one second per one-second segment), and human-readable text explanations—and reports state-of-the-art results on UCF-Crime (87.3 AUC) and XD-Violence (75.1 AP), beating prior zero-shot methods by large margins. Two small tricks, repulsive prompting and scaled anomaly penalization, correct the text encoder's bias toward anomalous captions. If right, Flashback is the first VAD that simultaneously offers all three capabilities.

What carries the argument

The pseudo-scene memory: a set of up to one million normal and anomalous captions generated by GPT-4o with no video input, stored with their text-encoder embeddings. Two bias-correction mechanisms carry the argument: repulsive prompting wraps normal and anomalous captions in distinct templates containing the keyword 'Normal' or 'Anomalous', widening the centroid angle between the two groups in embedding space (from 8.12 degrees to 33.29 degrees); scaled anomaly penalization multiplies the anomalous caption vectors by alpha = 0.95 before dot-product retrieval, dampening the encoder's skew toward anomalous text. The retrieval step itself—top-K softmax-weighted labels—is the online mechanism that replaces all LLM calls.

What would settle it

Build a test set with an anomaly category absent from the memory—for example, a rare surveillance event the LLM never described—and measure whether Flashback flags it. A concrete check is to compare the memory's category list against the ground-truth anomaly types of UCF-Crime and XD-Violence and compute per-category recall; any category with near-zero recall would invalidate the coverage claim.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the expensive, autoregressive language-model inference used by explainable zero-shot VAD can be moved entirely offline without losing accuracy. An off-the-shelf LLM generates a broad set of short captions of normal and anomalous scenes; those captions are embedded once and stored. At test time a frozen cross-modal encoder embeds each video segment, and the segment's anomaly score is the softmax-weighted average of the anomaly flags of its top-K retrieved captions, with the retrieved sentences serving as explanations. On UCF-Crime and XD-Violence this retrieval pipeline exceeds prior zero-shot methods and reaches 42 frames per second on a single consumer GPU, while a direct VLM-plus-LLM baseline runs at about 1.3 frames per second.

Load-bearing premise

The one million LLM-generated captions, produced without ever seeing video, already contain the normal and anomalous action categories that actually occur in the deployment videos; if a category is missing or mislabeled, retrieval will silently score that anomaly as normal.

Editorial extensions

If this is right

  • Flashback is claimed to be the first VAD that is simultaneously zero-shot, real-time, and explainable.
  • Accuracy scales with memory size and does not appear to saturate at one million captions, suggesting larger memories would help while throughput stays roughly constant.
  • Retrieval-based scores produce sharper anomaly curves than LAVAD's flattened scores, making threshold selection easier for downstream users.
  • The two debiasing tricks are training-free and portable to any frozen video-text encoder, so they can be reused outside this specific pipeline.
  • A one-second segment is processed in 0.713 seconds on an RTX 3090, meeting the paper's real-time criterion and sustaining about 42 frames per second.

Reading between the lines

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

  • The paper does not directly validate category-level recall: the memory's coverage of the actual anomaly types in UCF-Crime and XD-Violence is only indirectly tested by scaling memory size. A per-category recall analysis would be a stronger evaluation.
  • A fixed-label memory cannot handle context-dependent anomalies; the pickpocketing example in Figure 3(c) shows a case where the label conflicts with the benchmark. A dynamic or context-aware label layer would be a natural extension.
  • Because throughput barely changes with memory size, one could push to tens of millions of captions and likely gain further accuracy, though the cost of generating and storing those captions would need to be weighed.
  • The retrieval-based design invites a multi-modal memory: adding audio or motion descriptions to the captions could improve recall of subtle events, a direction the paper itself flags for future work.
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 / 5 minor

Summary. The paper proposes Flashback, a two-stage video anomaly detection (VAD) system. In an offline 'Recall' stage, a frozen LLM (GPT-4o) is prompted to generate one million normal-anomalous caption pairs without any video input; these captions are wrapped in class-specific templates ('Normal'/'Anomalous') and encoded by a frozen video-text encoder (ImageBind or PerceptionEncoder) into a pseudo-scene memory. At test time, the 'Respond' stage embeds each 1-second video segment, retrieves the top-K captions by dot-product similarity, applies a scaled anomaly penalization factor α=0.95, and averages the retrieved anomaly flags to produce frame-level scores, with retrieved captions serving as explanations. Flashback reports UCF-Crime AUC 87.29 (+7.0 over LAVAD) and XD-Violence AP 75.13 (+13.1), at 42 fps on an RTX 3090, and claims to be the first VAD system that is simultaneously zero-shot, real-time, and explainable.

Significance. If the claims are supported, Flashback is a practically valuable contribution: it removes all online LLM inference, enabling explainable VAD at real-time throughput on a consumer GPU, and its component ablations (repulsive prompting, top-K, memory size, segment parameters) are informative. The paper is also candid about limitations, including the inability to adapt labels to context. However, the central zero-shot SOTA claim currently rests on two unverified assumptions: that the LLM-generated memory covers the benchmark anomaly categories with labels aligned to each dataset's annotation scheme, and that the hyperparameters chosen on the test benchmarks do not compromise the zero-shot characterization. The significance is therefore contingent on additional evaluation.

major comments (4)
  1. [§3.2, §4.5(d), Fig. 3(c)] The pseudo-scene memory is built without video, and the paper states in §3.2 that coverage is sufficient, but the only supporting evidence is the memory-size scaling experiment in §4.5(d). That experiment shows accuracy increasing up to the 1M-budget cutoff and explicitly says 'performance growth does not saturate', so 1M is a budget limit, not a demonstrated sufficiency point. More importantly, there is no category-level recall or label-alignment measurement against the 13 UCF-Crime anomaly types or the 6 XD-Violence categories. Figure 3(c) gives a specific failure: 'Pickpocketing' is stored as anomalous, while XD-Violence treats it as normal, producing a systematic false positive. Because Flashback is a fixed lookup table, omitted categories cause silent misses and mislabeled categories cause systematic false alarms that retrieval cannot recover from. Please add a category-level analysis: per-type retrieval recall against ground-truth anomaly categories, a confusion matrix between memory labels and benchmark labels, and per-category AUC/AP.
  2. [§4.5(c), §4.5(e), Fig. 5] Key hyperparameters — α, K, segment length, overlap, and sampling rate — are selected by sweeping on the UCF-Crime and XD-Violence test sets. Specifically, §4.5(c) fixes K=10 based on Table 3(c) accuracy on both datasets; Table 3(e) selects segment parameters using both datasets; and the α sweep in Fig. 5 is performed on the merged UCF+XD test pool. The final numbers are then reported on the same test sets. This is test-set-driven model selection, which weakens the zero-shot claim. The phrase 'without per-dataset tuning' in §4.5 is true only in the narrow sense that one configuration is used for both datasets; the configuration itself was chosen using both test sets. Please either introduce a validation split or a leave-one-dataset-out selection protocol, or demonstrate insensitivity of the reported accuracy across a plausible parameter range.
  3. [§4.2, Table 1] The main results are single-run values without variance estimates. Table 3(a) quantifies only memory-subset variability (UCF AUC 84.38±0.60), not the variability of the full pipeline, which includes stochastic LLM caption generation and retrieval. Since the reported advantages over some baselines are modest (e.g., UCF AUC 87.29 vs. VERA's 86.55), single-run numbers make it difficult to judge whether the improvements are significant. Please report means and standard deviations over multiple runs of the complete memory-generation and inference pipeline, or at least bootstrapped confidence intervals for the main configurations.
  4. [§2, Table 1] Several recent zero-shot VAD methods cited in Related Work — AnyAnomaly [1], EventVAD [38], MCANet [13] — are absent from the comparison table. The paper's central claim is that Flashback 'outperforms prior zero-shot VAD methods by large margins' (§1), but the zero-shot comparison in Table 1 includes only LAVAD. Without comparisons to these recent methods, the SOTA claim is not fully established. Please add these methods to Table 1 or explain why a direct comparison is infeasible (e.g., missing implementations or different evaluation protocols).
minor comments (5)
  1. [§6] The section heading 'Limiations' is a typo for 'Limitations'.
  2. [§3.4] The sentence 'L is large (e.g., dozens of N)' is unclear; N denotes the number of stored captions, and it does not make sense for L to be 'dozens of N'. Please clarify the intended magnitude (e.g., number of tokens or frames).
  3. [§4.5, Table 3] Table 3(e) uses 'Tstride' while §3.3 and §4.3 define the overlap as Toverlap; please define the relationship between Tstride and Toverlap and use consistent notation.
  4. [Figure 4] The subplot labels in Figure 4 ('(a) w/ RP', '(b) w/o RP') appear inconsistent with the caption text, which describes (a) before and (b) after applying repulsive prompting. Please align the labels with the description.
  5. [§3.3] The symbol y*s used in the soft-label average is not defined; please make explicit that the retrieved anomaly flags are taken from the vector Y defined in Eq. (1).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the anomaly score is a weighted average of offline LLM-assigned caption labels, benchmark accuracy is measured externally, and the few self-citations are non-load-bearing baselines.

full rationale

Flashback's derivation chain is a retrieval system rather than a fitted prediction. Offline, an LLM generates normal/anomalous captions with fixed binary flags Y (Eq. 1); online, the segment score A_s is the softmax-weighted average of the top-K retrieved flags (Sec. 3.3). No equation defines its inputs in terms of its outputs, and the pseudo-scene memory is constructed without video input or benchmark labels, so the UCF-Crime/XD-Violence scores are external checks of the LLM prior rather than forced identities. The paper's only self-citations (RareAnom [42], DyAnNet [43]) appear as unsupervised baselines in Table 1 and do not carry the zero-shot/real-time/explainable claim. Two adjacent concerns are correctness or validity issues, not circularity: (i) Sec. 4.5 selects K=10 and alpha=0.95 by sweeping on the merged test benchmarks, which is test-set hyperparameter tuning rather than a derivation; (ii) Sec. 4.5(d) states 'performance growth does not saturate' while Sec. 3.1 asserts coverage is sufficient, and Fig. 3(c) plus Sec. 6 acknowledge fixed-label/context mismatches. These undermine the zero-shot generalization claim but do not make the reported results equivalent to the method's inputs by construction.

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

The central claim rests on cross-modal embedding quality, LLM caption coverage, fixed per-action anomaly labels, and a scale factor chosen on test data. No new physical entities are introduced.

free parameters (4)
  • alpha (scaled anomaly penalization) = 0.95
    Rescales all anomalous caption embeddings before dot-product retrieval; chosen by sweeping alpha on the combined UCF-Crime and XD-Violence test pool (Fig. 5).
  • K (number of retrieved captions) = 10
    Number of top-matching captions used in the softmax-weighted sum; selected by a sweep over K in {1,5,10,20,40} on the test sets (Table 3c).
  • Tsegment / Toverlap / Tsample = 1.0s / 0.0s / 16 frames
    Segment length, stride, and sampling frames selected by test-set experiments (Table 3e); the gray row is the default because it gives the highest UCF AUC while meeting the real-time constraint.
  • Gaussian smoothing kernel width and sigma = width 100 frames, sigma 0.5
    Post-processing parameters used for frame-level score refinement; set without reported sensitivity analysis.
assumptions (4)
  • domain assumption Embeddings from ImageBind or PerceptionEncoder place video segments and short text captions in a shared space where dot-product similarity reflects semantic relatedness.
    The entire retrieval pipeline relies on this cross-modal alignment; invoked in Sec. 3.2 and 3.3 without verification on unseen domains.
  • domain assumption GPT-4o can generate diverse, accurate captions of normal and anomalous surveillance scenes without seeing video, and the resulting memory covers the anomaly categories in UCF-Crime and XD-Violence.
    Offline recall stage, Sec. 3.2; the paper argues coverage is sufficient in Sec. 4.5 but only via memory-size scaling, not category-level recall.
  • domain assumption Assigning each action category a fixed binary anomaly label is adequate for frame-level evaluation.
    The Limitations section admits anomalies are context-dependent and the design cannot adapt labels on the fly.
  • ad hoc to paper Scaled anomaly penalization with alpha = 0.95 generalizes without per-dataset tuning.
    The value is chosen on the combined test pool (Fig. 5), so it is an ad hoc setting for this benchmark setup rather than a derived constant.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Flashback: Memory-Driven Zero-shot, Real-time Video Anomaly Detection." pith.science (2026). https://pith.science/paper/5WJ44NFH

@misc{pith2026250515205,
  author       = {Pith},
  title        = {Pith review of: Flashback: Memory-Driven Zero-shot, Real-time Video Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5WJ44NFH}},
  note         = {Machine review of arXiv:2505.15205}
}
read the original abstract

Video Anomaly Detection (VAD) automatically identifies anomalous events from video, mitigating the need for human operators in large-scale surveillance deployments. However, two fundamental obstacles hinder real-world adoption: domain dependency and real-time constraints -- requiring near-instantaneous processing of incoming video. To this end, we propose Flashback, a zero-shot and real-time video anomaly detection paradigm. Inspired by the human cognitive mechanism of instantly judging anomalies and reasoning in current scenes based on past experience, Flashback operates in two stages: Recall and Respond. In the offline recall stage, an off-the-shelf LLM builds a pseudo-scene memory of both normal and anomalous captions without any reliance on real anomaly data. In the online respond stage, incoming video segments are embedded and matched against this memory via similarity search. By eliminating all LLM calls at inference time, Flashback delivers real-time VAD even on a consumer-grade GPU. On two large datasets from real-world surveillance scenarios, UCF-Crime and XD-Violence, we achieve 87.3 AUC (+7.0 pp) and 75.1 AP (+13.1 pp), respectively, outperforming prior zero-shot VAD methods by large margins.

Figures

Figures reproduced from arXiv: 2505.15205 by the authors.

Figure 1
Figure 1. Bridging speed and reasoning. (a) Real-time VAD keeps a light video encoder online but cannot work zero-shot or explain its decisions. (b) Explainable VAD adds a large VLM + LLM in the loop; reasoning is possible, yet speed drops and zero-shot ability is partial. (c) Flashback moves the LLM offline, builds a pseudo-scene memory once, and uses a frozen cross-modal encoder at test time, so it is simultaneously real-ti… view at source ↗
Figure 2
Figure 2. Overview of Flashback. Flashback operates in two disjoint stages. Offline Recall: a frozen LLM [33] generates a diverse set of normal and anomalous scene sentences using context and format prompts PC, PF, which are embedded by a frozen video-text encoder and stored in a million-entry Pseudo-Scene Memory CN, CA. Repulsive Prompting widens the separation between normal and anomalous embeddings, countering the encoder’… view at source ↗
Figure 3
Figure 3. Qualitative examples. The plots show frame-wise anomaly curves. Red boxes on both the video strip and the plot mark ground-truth anomalous intervals . For selected frames we list the retrieved category-caption pairs (κ, c) and their anomaly flags y. Black text denotes a correct description, gray text an incorrect one. (a) & (b) The top captions describes the event precisely. (c) Flashback flags “Pickpocketing” as ab… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: T-SNE embeddings of cap￾tion features. We subsample 5,000 normal-anomalous caption pairs and visualize (a) before and (b) after ap￾plying repulsive prompting (RP). RP clearly separates the two groups. (b) Repulsive prompting (RP). Removing RP lowers UCF AUC from 87.29 …
Figure 5
Figure 5. Figure 5: AUC vs. scale factor α. A mild reduction (α≈0.95) yields fa￾vorable AUC, confirming that scaled anomaly penalization is effective with￾out fine-tuning. (e) Segment length and sampling rate. Using 16 frames in a 1s segment reaches the highest scores—87.29 UCF AUC and 75…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 52 canonical work pages

  1. [1]

    S. Ahn, Y. Jo, K. Lee, S. Kwon, I. Hong, and S. Park. Anyanomaly: Zero-shot customizable video anomaly detection with lvlm. arXiv preprint arXiv:2503.04504, 2025

  2. [38]

    Eventvad: Training-free event-aware video anomaly detection.arXiv preprint arXiv:2504.13092, 2025

    Yihua Shao, Haojin He, Sijie Li, Siyu Chen, Xinwei Long, Fanhu Zeng, Yuxuan Fan, Muyang Zhang, Ziyang Yan, Ao Ma, et al. Eventvad: Training-free event-aware video anomaly detection.arXiv preprint arXiv:2504.13092, 2025

  3. [13]

    Mcanet: Multimodal caption aware training-free video anomaly detection via large language model

    Prabhu Prasad Dev, Raju Hazari, and Pranesh Das. Mcanet: Multimodal caption aware training-free video anomaly detection via large language model. InICPR, 2024

  4. [2]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. InNeurIPS, 2022

  5. [3]

    Lawrence Zitnick

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, Devi Parikh, and C. Lawrence Zitnick. Vqa: Visual question answering. InICCV, 2015

  6. [4]

    The proactive brain: Using analogies and associations to generate predictions.Trends in Cognitive Sciences, 2007

    Moshe Bar. The proactive brain: Using analogies and associations to generate predictions.Trends in Cognitive Sciences, 2007

  7. [5]

    Marius Zöllner

    Daniel Bogdoll, Maximilian Nitsche, and J. Marius Zöllner. Anomaly detection in autonomous driving: A survey. InIEEE Conf. Comput. Vis. Pattern Recog. Worksh., 2022

  8. [6]

    Perceptionencoder: Thebestvisualembeddings are not at the output of the network.arXiv:2504.13181, 2025

    Daniel Bolya, Po-Yao Huang, Peize Sun, Jang Hyun Cho, Andrea Madotto, Chen Wei, Tengyu Ma, Jiale Zhi,JathushanRajasegaran,HanoonaRasheed,JunkeWang,MarcoMonteiro,HuXu,ShiyuDong,Nikhila Ravi,DanielLi,PiotrDollár,andChristophFeichtenhofer. Perceptionencoder: Thebestvisualembeddings are not at the output of the network.arXiv:2504.13181, 2025

Show all 63 references
  1. [7]

    Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, et al

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, et al. Language models are few-shot learners. InNeurIPS, 2020

  2. [8]

    BusinessWire. Global Surveillance Camera Market Poised to Reach US$33.11 Billion by 2023; 278.6 Million Units Forecast.https://www.businesswire.com/news/home/20231025926921/en/ Global-Surveillance-Camera-Market-Poised-to-Reach-US33.11-Billion-by-2023 , 2023. Accessed: May 2025

  3. [9]

    Mgfn: magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection

    Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. InAAAI, 2023

  4. [10]

    Towards multi-domain learning for generalizable video anomaly detection

    MyeongAh Cho, Taeoh Kim, Minho Shim, Dongyoon Wee, and Sangyoun Lee. Towards multi-domain learning for generalizable video anomaly detection. InNeurIPS, 2024

  5. [11]

    Think you have solved the ai2 reasoning challenge? reconsidering the arc dataset

    Peter Clark and Oren Etzioni. Think you have solved the ai2 reasoning challenge? reconsidering the arc dataset. InACL, 2018

  6. [12]

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony M. H. Tiong, et al. Instructblip: Towards general-purpose vision-language models with instruction tuning. InNeurIPS, 2023

  7. [14]

    A theory of cortical responses.Philosophical Transactions of the Royal Society B, 2005

    Karl Friston. A theory of cortical responses.Philosophical Transactions of the Royal Society B, 2005

  8. [15]

    Imagebind: One embedding space to bind them all

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. InCVPR, 2023

  9. [16]

    Gong andet al.Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection

    H. Gong andet al.Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. In ICLR, 2024

  10. [17]

    Making the V in VQA matter: Elevating the role of image understanding in visual question answering

    Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. InCVPR, 2017

  11. [18]

    Video Surveillance Market Size, Share & Trends Analy- sis Report By Component, By Deployment Mode, By Application, By Region – Global Forecast to 2030

    Grand View Research. Video Surveillance Market Size, Share & Trends Analy- sis Report By Component, By Deployment Mode, By Application, By Region – Global Forecast to 2030. https://www.grandviewresearch.com/industry-analysis/ video-surveillance-market-report, 2024. Accessed: May 2025

  12. [19]

    Roy-Chowdhury, and Larry S

    Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal regularity in video sequences. InCVPR, 2016

  13. [20]

    Hudson and Christopher D

    Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. InCVPR, 2019. 10

  14. [21]

    InICIP, 2023

    HyekangKevinJoo,KhoaVo,KashuYamazaki,andNganLe.Clip-tsa: Clip-assistedtemporalself-attention for weakly-supervised video anomaly detection. InICIP, 2023

  15. [22]

    Triviaqa: Alargescaledistantlysupervised challenge dataset for reading comprehension

    MandarJoshi,EunsolChoi,DanielWeld,andLukeZettlemoyer. Triviaqa: Alargescaledistantlysupervised challenge dataset for reading comprehension. InACL, 2017

  16. [23]

    Karim, V

    H. Karim, V. Pande, and N. Ahuja. Reward: Real-time weakly supervised video anomaly detection. In WACV, 2024

  17. [24]

    Naturalquestions: abenchmarkforquestionansweringresearch

    TomKwiatkowskiandAlexisetal.Palmer. Naturalquestions: abenchmarkforquestionansweringresearch. TACL, 2019

  18. [25]

    Blip-2: Bootstrappinglanguage-imagepre-training with frozen image encoders and large language models

    JunnanLi,DongxuLi,SilvioSavarese,andStevenHoi. Blip-2: Bootstrappinglanguage-imagepre-training with frozen image encoders and large language models. InInternational conference on machine learning, pages 19730–19742. PMLR, 2023

  19. [26]

    Self-training multi-sequence learning with transformer for weakly supervised video anomaly detection

    Shuo Li, Fang Liu, and Licheng Jiao. Self-training multi-sequence learning with transformer for weakly supervised video anomaly detection. InAAAI, 2022

  20. [27]

    Multi-modal preference alignment remedies degradation of visual instruction tuning on language models.arXiv preprint arXiv:2402.10884, 2024

    Shengzhi Li, Rongyu Lin, and Shichao Pei. Multi-modal preference alignment remedies degradation of visual instruction tuning on language models.arXiv preprint arXiv:2402.10884, 2024

  21. [28]

    Improved baselines with visual instruction tuning, 2023

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023

  22. [29]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. InNeurIPS, 2023. Oral

  23. [30]

    Abnormal event detection at 150 fps in matlab

    Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 fps in matlab. InICCV, 2013

  24. [31]

    OK-VQA: A visual question answering benchmark requiring external knowledge

    Kenneth Marino, Zhou Yu, Yuchen Zhang, Junjie Luo, Mohit Bansal, Stefan Lee, and Dhruv Batra. OK-VQA: A visual question answering benchmark requiring external knowledge. InCVPR, 2019

  25. [32]

    Micorek, M

    M. Micorek, M. Rudzinski, and L. Zhang. Mulde: Multi-scale log-density estimation for video anomaly detection. InCVPR, 2024

  26. [33]

    Gpt-4o: Openai’s omnimodal model.https://openai.com/index/gpt-4o, 2024

    OpenAI. Gpt-4o: Openai’s omnimodal model.https://openai.com/index/gpt-4o, 2024

  27. [34]

    Wainwright, Pamela Mishkin, Chong Zhang, SandhiniAgarwal,etal

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, SandhiniAgarwal,etal. Traininglanguagemodelstofollowinstructionswithhumanfeedback. In NeurIPS, 2022

  28. [35]

    The lambada dataset: Word prediction requiring a broad discourse context

    Denis Paperno, Germán Kruszewski, Angeliki Lazaridou, and Marco Baroni. The lambada dataset: Word prediction requiring a broad discourse context. InACL, 2016

  29. [36]

    Grounding multimodal large language models to the world

    Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. InICLR, 2024

  30. [37]

    Towards total recall in industrial anomaly detection

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. InCVPR, 2022

  31. [39]

    Real-world anomaly detection in surveillance videos

    Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018

  32. [40]

    de Lange

    Christopher Summerfield and Floris P. de Lange. Expectation in perceptual decision making: Neural and computational mechanisms.Nature Reviews Neuroscience, 2014

  33. [41]

    Commonsenseqa: A question answering challenge targeting commonsense knowledge

    Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. Commonsenseqa: A question answering challenge targeting commonsense knowledge. InNACCL, 2019

  34. [42]

    Rareanom: A benchmark video dataset for rare type anomalies.PR, 2023

    Kamalakar Vijay Thakare, Debi Prosad Dogra, Heeseung Choi, Haksub Kim, and Ig-Jae Kim. Rareanom: A benchmark video dataset for rare type anomalies.PR, 2023

  35. [43]

    Dyannet: A scene dynamicity guided self-trained video anomaly detection network

    Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. InWACV, 2023. 11

  36. [44]

    Verjans, and Gustavo Carneiro

    Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W. Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. InICCV, 2021

  37. [45]

    LLaMA 2: Open foundation and fine-tuned language models

    Hugo Touvron, Louis Martin, Kevin Stone, Petr Barulina, Kevin Borlaug, Faisal Azhar, Gideon Dror, Armand Joulin, Edouard Grave, and Alexis Conneau. LLaMA 2: Open foundation and fine-tuned language models. arXiv preprint arXiv:2307.09288, 2023

  38. [46]

    Unsupervised video anomaly detection with diffusion models conditioned on compact motion representations

    Anil Osman Tur, Nicola Dall’Asen, Cigdem Beyan, and Elisa Ricci. Unsupervised video anomaly detection with diffusion models conditioned on compact motion representations. InICIAP, 2023

  39. [47]

    Exploring diffusion models for unsupervised video anomaly detection

    Anil Osman Tur, Nicola Dall’Asen, Cigdem Beyan, and Elisa Ricci. Exploring diffusion models for unsupervised video anomaly detection. InICIP, 2023

  40. [48]

    Gods: Generalizedone-classdiscriminativesubspacesforanomalydetection

    JueWangandAnoopCherian. Gods: Generalizedone-classdiscriminativesubspacesforanomalydetection. In ICCV, 2019

  41. [49]

    Chi, Quoc V

    Yizhong Wang, Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed H. Chi, Quoc V. Le, Denny Zhou, et al. Self-instruct: Aligning language models with self generated instructions. InACL, 2023

  42. [50]

    Le, and Ed H

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Quoc V. Le, and Ed H. Chi. Chain-of-thought prompting elicits reasoning in large language models. InNeurIPS, 2022

  43. [51]

    Le, and Ed H

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Quoc V. Le, and Ed H. Chi. Finetuned language models are zero-shot learners. InICLR, 2022

  44. [52]

    Self-supervised sparse representation for video anomaly detection

    Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. InECCV, 2022

  45. [53]

    Learning causal temporal relation and feature discrimination for anomaly detection

    Peng Wu and Jing Liu. Learning causal temporal relation and feature discrimination for anomaly detection. TIP, 2021

  46. [54]

    Not only look, but also listen: Learning multimodal violence detection under weak supervision

    Peng Wu, jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. InECCV, 2020

  47. [55]

    Vadclip: Adapting vision-language models for weakly supervised video anomaly detection.AAAI, 2024

    Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection.AAAI, 2024

  48. [56]

    Avadclip: Audio-visualcollaborationforrobustvideoanomalydetection

    Peng Wu, Wanshun Su, Guansong Pang, Yujia Sun, Qingsen Yan, Peng Wang, and Yanning Zhang. Avadclip: Audio-visualcollaborationforrobustvideoanomalydetection. arXivpreprintarXiv:2504.04495 , 2025

  49. [57]

    Wu andet al.Open-vocabulary video anomaly detection

    Y. Wu andet al.Open-vocabulary video anomaly detection. InCVPR, 2024

  50. [58]

    Vera: Explainable video anomaly detection via verbalized learning of vision-language models.arXiv preprint arXiv:2412.01095, 2024

    Muchao Ye, Weiyang Liu, and Pan He. Vera: Explainable video anomaly detection via verbalized learning of vision-language models.arXiv preprint arXiv:2412.01095, 2024

  51. [59]

    Generative cooperative learning for unsupervised video anomaly detection

    M.ZaighamZaheer,ArifMahmood,M.HarisKhan,MattiaSegu,FisherYu,andSeung-IkLee. Generative cooperative learning for unsupervised video anomaly detection. InCVPR, 2022

  52. [60]

    Harnessing large language models for training-free video anomaly detection

    Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. InCVPR, 2024

  53. [61]

    Holmes-vad: Towards unbiased and explainable video anomaly detection via multi-modal llm.arXiv preprint arXiv:2406.12235, 2024

    Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Chuchu Han, Xiaonan Huang, Changxin Gao, Yuehuan Wang, and Nong Sang. Holmes-vad: Towards unbiased and explainable video anomaly detection via multi-modal llm.arXiv preprint arXiv:2406.12235, 2024

  54. [62]

    Temporal convolutional network with complementary inner bag loss for weakly supervised anomaly detection

    Jiangong Zhang, Laiyun Qing, and Jun Miao. Temporal convolutional network with complementary inner bag loss for weakly supervised anomaly detection. InICIP, 2019

  55. [63]

    Video anomaly detection for smart surveillance

    Sijie Zhu, Chen Chen, and Waqas Sultani. Video anomaly detection for smart surveillance. InComputer Vision: A Reference Guide. Springer, 2021. 12

Pith tools

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