Pith. sign in

REVIEW 3 major objections 7 minor 51 references

Forward Consistency Learning with Gated Context Aggregation for Video Anomaly Detection

T0 review · 3 major / 7 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A 2.17-million-parameter U-Net that predicts both the immediate next frame and a longer-term frame can outperform much larger video anomaly detectors.

desk verdict Solid efficiency story, but the accuracy claims rest on per-dataset test-set tuning and are overstated; worth peer review with a validation protocol. read the letter →

arxiv 2601.18135 v1 pith:L4UE3BGE submitted 2026-01-26 cs.CV

Jiahao Lyu , Minghua Zhao , Xuewen Huang , Yifei Chen , Shuangli Du , Jing Hu , Cheng Shi , Zhiyong Lv This is my paper
classification cs.CV
keywords videoanomalydetectionframepredictionforwardconsistencylossgatedattentionU-Netlightweightmodelmulti-scaleevaluationedgecomputing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a lightweight video anomaly detection model can be both more accurate and faster than heavy state-of-the-art alternatives. The key idea is to make a single U-Net predict two future frames at different horizons, combine their prediction errors into one anomaly score, and train with a forward consistency loss. On four standard benchmarks, the method reports top-tier frame-level AUC while running at 110 FPS, suggesting that compact models do not have to sacrifice detection quality. If correct, this would make real-time anomaly detection practical for edge devices with limited memory and compute.

What carries the argument

The central object is a dual-horizon prediction head that makes two future-frame predictions from one decoder, paired with the Gated Context Aggregation Module (GCAM) embedded in U-Net skip connections. GCAM combines a multi-scale context feature aggregation branch and an efficient gated attention unit that reweights channel and spatial features. The forward consistency loss, including an SSIM term, aligns the two predictions so that the longer-horizon error becomes a more sensitive anomaly signal.

What would settle it

One concrete check is to rerun FoGA on the same benchmarks with a single fixed λ (for example, 0.2) and σ=4, using only a validation split for any tuning, and compare the resulting AUC to the paper's reported numbers and to the best baselines. If the AUC drops to the level of prior methods, the central performance claim is weakened.

Watch

Extended reading notes

Core claim

FoGA is a U-Net with about 2.17 million parameters that takes t consecutive frames as input and outputs both the immediate next frame Ît+1 and a forward frame Ît+σ from the same decoding stage. A gated context aggregation module is inserted into skip connections to adaptively fuse encoder and decoder features, and the training loss combines intensity and gradient losses on both predictions with an SSIM-based consistency loss between them. At test time, the anomaly score fuses the immediate and forward prediction errors, weighted by λ. On UCSD Ped1, Ped2, CUHK Avenue, and ShanghaiTech, the paper reports frame-level AUC of 87.4%, 98.9%, 90.1%, and 76.2%, best or second-best among the compared

Load-bearing premise

The main load-bearing premise is that the reported AUC advantage comes with λ, σ, and input length chosen per dataset using test-set results; if those hyperparameters had to be fixed beforehand or chosen only on validation data, the performance margin could shrink or disappear.

Editorial extensions

If this is right

  • If the reported results hold, lightweight models with around 2M parameters can achieve state-of-the-art video anomaly detection accuracy, making edge-device deployment feasible.
  • The dual-horizon design means anomalies whose early motion looks normal can still be caught by the larger error at a longer prediction horizon.
  • A single autoencoder suffices for multi-frame temporal modeling, avoiding the doubled compute and memory of bidirectional two-network methods.
  • The forward consistency loss provides a simple, general training constraint that can be added to existing prediction-based VAD frameworks.
  • Multi-scale error evaluation, while slightly slower, gives a consistent accuracy boost and can be disabled for higher throughput.

Reading between the lines

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

  • The prediction horizon σ acts like a tunable temporal receptive field; choosing σ too small collapses the two predictions into one, and the reported results suggest that a moderate horizon improves discrimination.
  • A natural extension would be to adaptively weight the immediate and forward errors at test time rather than using a fixed λ per dataset, which might remove the need for per-dataset tuning.
  • The forward consistency loss could be transferred as a regularizer to reconstruction-based anomaly detectors, where a similar consistency between short and long temporal views might strengthen normality modeling.
  • A fairer comparison would involve selecting λ and σ on a validation split rather than the test set; the paper does not report such a protocol, so the generalization margin is uncertain.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes FoGA, a lightweight U-Net-based video anomaly detection model that predicts both an immediate next frame and a longer-term forward frame, using a gated context aggregation module (GCAM) in the skip connections and a hybrid anomaly score combining the two prediction errors. On UCSD Ped1, Ped2, CUHK Avenue, and ShanghaiTech, it reports frame-level AUCs of 87.4%, 98.9%, 90.1%, and 76.2% at 110 FPS (155 FPS without multi-scale scoring), claiming a better accuracy-efficiency trade-off than existing methods. The central claim is that simultaneous short- and long-term prediction with forward consistency and lightweight gated aggregation yields state-of-the-art accuracy at real-time speed.

Significance. If the reported results are robust, FoGA would be a valuable contribution to edge-oriented VAD: the model has only 2.17M parameters, 5.85 GFLOPs, and the architecture is simple enough to be reproducible. The idea of using two prediction horizons with a consistency term is a reasonable extension of prediction-based VAD, and the component ablations suggest that both the gated module and the forward loss contribute positively. However, the headline performance rests on per-dataset hyperparameter selection using test-set labels, with no validation protocol, error bars, or code. Under those conditions, the claimed margins over the strongest competitor BiSP are small (+1.1, +0.3, +0.6, −0.2 AUC), so the paper's central quantitative claim is not yet secured. The efficiency numbers are plausible but need independent confirmation.

major comments (3)
  1. [Sec. IV-B and IV-D, Eq. (17)] The fusion weight λ is set per dataset (0.06/1.0/0.2/0.06) and σ=4, t=4/8 are fixed after inspecting test-set AUC sweeps in Fig. 3-4. No validation split is described. This means the reported AUCs are the result of selecting hyperparameters on the same test data used to compute the final metric, which is circular. Since the margins over BiSP are small (Table I: +1.1, +0.3, +0.6, −0.2), the 'substantially outperforming' claim is not supported. Please provide a proper protocol: fix hyperparameters a priori, or use a validation split, or report sensitivity over a grid with error bars and show that the advantage holds for a wide range of choices.
  2. [Eq. (15), Sec. III-C] The loss L_con = 1 − SSIM(Î_{t+1}, Î_{t+σ}) enforces appearance similarity between two predicted future frames, not 'forward consistency' in a motion sense. This is a static structural similarity between two predictions at different horizons. It may encourage Î_{t+σ} to copy Î_{t+1}, potentially collapsing the long-term prediction and undermining the stated purpose of modeling longer-term temporal dynamics. The ablation in Table IV shows a modest gain from L_con, but the mechanism is unclear. Please rename the loss, or replace it with a genuine motion-consistency term (e.g., based on temporal gradients or optical flow), or justify why appearance similarity between two future frames helps anomaly detection.
  3. [Sec. IV-C, Tables I-II] The paper reports single point estimates without error bars, multiple seeds, or code release. Given the small performance margins and the dataset-specific tuning, this is insufficient to establish a reliable ranking. Please report mean ± std over at least three independent training runs, provide per-dataset hyperparameter choices on a validation split, and release code and trained models to allow reproduction. Without this, the efficiency-vs-accuracy claims cannot be independently verified.
minor comments (7)
  1. [Eq. (18)] The PSNR formula is unclear: the denominator is written as (1/∑_{i=0}^{N} v_i) with N=3, and v_i is the maximum patch-wise MSE at scale i. Standard PSNR uses the mean squared error over all pixels. Please define the multi-scale error pyramid precisely and explain how it yields a PSNR; otherwise the anomaly score is not reproducible.
  2. [Eq. (3), Eq. (7)] The notation for the adaptive kernel size is confusing: the floor/odd operator is not clearly defined, and in Eq. (7) the max with 3 is written as max(3, |...|_odd), but the text says k is determined by spatial resolution. Please clarify the exact computation.
  3. [Sec. IV-B] Please state explicitly whether the test-set anomalies were ever used during development, e.g., for choosing t, σ, λ, or early stopping. If any test information leaked, the comparison in Table I becomes unfair.
  4. [Table I] Some reported baselines are from the authors' own recent publications ([8], [15], [17])—this is fine, but the relation and architectural differences should be discussed more carefully, since FoGA appears to be a direct extension of BiSP with a gated module and forward consistency.
  5. [Abstract / Sec. I] The abstract says 'substantially outperforms state-of-the-art competing methods,' but on ShanghaiTech FoGA is 0.2 AUC worse than BiSP (76.2 vs 76.4). Please qualify the claim to reflect the actual ranking.
  6. [Fig. 3-4] The AUC-vs-λ and AUC-vs-σ curves have no error bars, and the x-axis ranges are partially hidden. Please include error bars and label all axes clearly.
  7. [General] There are several typos and inconsistent notations (e.g., 'predictes' in Fig. 5 caption, 'the FOGA' in Sec. IV-D, and the use of I_t both as input sequence and as time index in Eqs. 11-12). A thorough proofread is needed.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline AUCs are selected on the test set via per-dataset fusion weight λ (and σ/t ablations), so the reported accuracy advantage is partially an optimized fit rather than a free prediction.

  1. fitted input called prediction [Sec. IV-B (Experimental Setup) and Sec. IV-D (Ablation Study), Eq. (17), Fig. 3, Table I]
    "The error fusing weights λ for Ped1, Ped2, Avenue, and Sh-Tech are set to 0.06, 1.0, 0.2, and 0.06, respectively. ... Notably, the optimal λ selected from the first study is used in the remaining three ablation experiments. ... Fig. 3. AUC comparison under different λ across four datasets."

    The final anomaly score is E = E_i + λ·E_f (Eq. 17). λ is chosen per dataset to maximize frame-level AUC on the benchmark test set (Fig. 3), and the Table I AUCs (87.4/98.9/90.1/76.2) are then computed on that same test set with the selected λ. No validation split is reported. Hence the headline 'substantially outperforms' numbers are not free predictions of a fixed model; they are selected maxima of a one-parameter family over the same labels used to compute the reported metric. The same test-set ablation is used to pick σ (Fig. 4) and Avenue t=8.

full rationale

The derivation of the model itself (U-Net + GCAM + forward prediction losses L_pred, L_fc, L_con and the hybrid error Eq. 17) is not circular in the sense of an equation reducing to its own inputs; the forward-consistency loss constrains two predictions and the anomaly score is a deterministic function of prediction errors. There is no self-citation chain or imported uniqueness theorem: citations to the authors' prior BiSP/MoBA/VAD-Mamba papers are only comparative/related work, not load-bearing. The only serious reduction is experimental: the scalar fusion weight λ is tuned on the test set (per-dataset values 0.06/1.0/0.2/0.06) by the AUC curves in Fig. 3, and σ and input length t are likewise selected from test-set ablations. Because the reported Table I AUCs are evaluated under these test-selected hyperparameters, the performance advantage over BiSP etc. is partly a fitted optimum rather than a fixed-model prediction. This does not make the architecture's qualitative behavior vacuous, but it means the central 'state-of-the-art accuracy' claim is not fully free-standing. I therefore score it 6 rather than higher: the prediction mechanism still has independent content, but the headline comparison is partially constructed by test-label selection.

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

The model's performance is carried by per-dataset hyperparameters (λ, σ, t) tuned against test labels, and by the untested assumption that an SSIM consistency between two predicted frames captures temporal 'forward consistency'. These are the main external loads beyond standard VAD premises.

free parameters (3)
  • λ (error fusion weight) = Ped1: 0.06; Ped2: 1.0; Avenue: 0.2; Sh-Tech: 0.06
    Chosen per dataset by sweeping values and picking the one that maximizes test AUC (Sec. IV-B, IV-D, Fig. 3). This is test-set fitting.
  • σ (forward prediction step) = 4 for all datasets
    Selected via ablation on test sets (Fig. 4); σ=1 makes the two predictions identical and 'constraint fails'.
  • t (input frame length) = 4 for Ped1/Ped2/Sh-Tech; 8 for Avenue
    Set per dataset without a validation-based justification; likely tuned on test data.
assumptions (4)
  • domain assumption Anomalous events yield larger prediction errors than normal events
    Core premise of prediction-based VAD; used to define the anomaly score (Sec. I, III-D).
  • ad hoc to paper SSIM between two predicted future frames enforces 'forward consistency' and improves temporal modeling
    L_con = 1 - SSIM(Ît+1, Ît+σ) (Eq. 15) is a static image-similarity term, not a motion-based consistency; no theoretical support that aligning two predictions at different time horizons separates anomalies.
  • domain assumption The multi-scale error pyramid from [34] is a valid anomaly measure
    Adopted from prior work (Eq. 18) without re-validation in this setting.
  • domain assumption U-Net-based prediction generalizes to unseen normal patterns
    Standard architectural assumption; relies on training data covering normal variation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Forward Consistency Learning with Gated Context Aggregation for Video Anomaly Detection." pith.science (2026). https://pith.science/paper/L4UE3BGE

@misc{pith2026260118135,
  author       = {Pith},
  title        = {Pith review of: Forward Consistency Learning with Gated Context Aggregation for Video Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L4UE3BGE}},
  note         = {Machine review of arXiv:2601.18135}
}
read the original abstract

As a crucial element of public security, video anomaly detection (VAD) aims to measure deviations from normal patterns for various events in real-time surveillance systems. However, most existing VAD methods rely on large-scale models to pursue extreme accuracy, limiting their feasibility on resource-limited edge devices. Moreover, mainstream prediction-based VAD detects anomalies using only single-frame future prediction errors, overlooking the richer constraints from longer-term temporal forward information. In this paper, we introduce FoGA, a lightweight VAD model that performs Forward consistency learning with Gated context Aggregation, containing about 2M parameters and tailored for potential edge devices. Specifically, we propose a Unet-based method that performs feature extraction on consecutive frames to generate both immediate and forward predictions. Then, we introduce a gated context aggregation module into the skip connections to dynamically fuse encoder and decoder features at the same spatial scale. Finally, the model is jointly optimized with a novel forward consistency loss, and a hybrid anomaly measurement strategy is adopted to integrate errors from both immediate and forward frames for more accurate detection. Extensive experiments demonstrate the effectiveness of the proposed method, which substantially outperforms state-of-the-art competing methods, running up to 155 FPS. Hence, our FoGA achieves an excellent trade-off between performance and the efficiency metric.

Figures

Figures reproduced from arXiv: 2601.18135 by the authors.

Figure 1
Figure 1. Left: Three prediction modeling types of VAD. (i) Prediction-based (Pred.) method. (ii) Bidirectional prediction-based (Bi-Pred.) method. (iii) The proposed forward consistency prediction method (For-Pred.). Right: The radar chart provides comparison among Frame-Pred [7], BiSP [8] and the proposed FoGA on four datasets performance (AUC ↑), as well as efficiency metric (FPS ↑, Params(M) ↓ and Flops(G) ↓). real-world … view at source ↗
Figure 2
Figure 2. Overall framework of the proposed FoGA that simultaneously predicts an immediate frame [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. AUC comparison under different λ across four datasets [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Effect of different forward consistency measurement parameter [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of Ground Truth (GT) frames, Immediate predictes, Forward predictes, and Hybrid errors on four datasets. The objects marked with red [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visualization of Attention map in three different types. The top [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visualization of score curves on four datasets. The x-axis indicates [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 1 linked inside Pith

  1. [1]

    Generalized video anomaly event detection: Systematic taxonomy and comparison of deep models,

    Y . Liu, D. Yang, Y . Wang, J. Liu, J. Liu, A. Boukerche, P. Sun, and L. Song, “Generalized video anomaly event detection: Systematic taxonomy and comparison of deep models,”ACM Computing Surveys, vol. 56, no. 7, pp. 1–38, 2024

    2024

  2. [2]

    Memorizing normality to detect anomaly: Memory- augmented deep autoencoder for unsupervised anomaly detection,

    D. Gong, L. Liu, V . Le, B. Saha, M. R. Mansour, S. Venkatesh, and A. v. d. Hengel, “Memorizing normality to detect anomaly: Memory- augmented deep autoencoder for unsupervised anomaly detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 1705–1714

    2019

  3. [3]

    Learning memory-guided normality for anomaly detection,

    H. Park, J. Noh, and B. Ham, “Learning memory-guided normality for anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 14 372–14 381

    2020

  4. [4]

    Hierarchical semantic contrast for scene-aware video anomaly detection,

    S. Sun and X. Gong, “Hierarchical semantic contrast for scene-aware video anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 22 846–22 856

    2023

  5. [5]

    Video anomaly detection based on spatio-temporal relationships among objects,

    Y . Wang, T. Liu, J. Zhou, and J. Guan, “Video anomaly detection based on spatio-temporal relationships among objects,”Neurocomputing, vol. 532, pp. 141–151, 2023

    2023

  6. [6]

    Towards interpretable video anomaly detec- tion,

    K. Doshi and Y . Yilmaz, “Towards interpretable video anomaly detec- tion,” inProceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, 2023, pp. 2655–2664

    2023

  7. [7]

    Future frame prediction for anomaly detection–a new baseline,

    W. Liu, W. Luo, D. Lian, and S. Gao, “Future frame prediction for anomaly detection–a new baseline,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6536– 6545

    2018

  8. [8]

    Bidirectional skip-frame prediction for video anomaly detection with intra-domain disparity-driven attention,

    J. Lyu, M. Zhao, J. Hu, R. Xi, X. Huang, S. Du, C. Shi, and T. Ma, “Bidirectional skip-frame prediction for video anomaly detection with intra-domain disparity-driven attention,”Pattern Recognition, vol. 170, p. 112010, 2026

    2026

Show all 51 references
  1. [9]

    Clustering driven deep au- toencoder for video anomaly detection,

    Y . Chang, Z. Tu, W. Xie, and J. Yuan, “Clustering driven deep au- toencoder for video anomaly detection,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16. Springer, 2020, pp. 329–345

    2020

  2. [10]

    A cascade reconstruction model with generalization ability evaluation for anomaly detection in videos,

    Y . Zhong, X. Chen, J. Jiang, and F. Ren, “A cascade reconstruction model with generalization ability evaluation for anomaly detection in videos,”Pattern Recognition, vol. 122, p. 108336, 2022

    2022

  3. [11]

    Dast-net: Dense visual attention augmented spatio-temporal network for unsupervised video anomaly detection,

    R. Kommanduri and M. Ghorai, “Dast-net: Dense visual attention augmented spatio-temporal network for unsupervised video anomaly detection,”Neurocomputing, p. 127444, 2024

    2024

  4. [12]

    Pseudobound: Limiting the anomaly reconstruction capability of one-class classifiers using pseudo anomalies,

    M. Astrid, M. Z. Zaheer, and S.-I. Lee, “Pseudobound: Limiting the anomaly reconstruction capability of one-class classifiers using pseudo anomalies,”Neurocomputing, vol. 534, pp. 147–160, 2023

    2023

  5. [13]

    Appearance-motion memory consistency network for video anomaly detection,

    R. Cai, H. Zhang, W. Liu, S. Gao, and Z. Hao, “Appearance-motion memory consistency network for video anomaly detection,” inProceed- ings of the AAAI conference on artificial intelligence, vol. 35, no. 2, 2021, pp. 938–946. 9

    2021

  6. [14]

    Spatiotemporal consistency-enhanced network for video anomaly detection,

    Y . Hao, J. Li, N. Wang, X. Wang, and X. Gao, “Spatiotemporal consistency-enhanced network for video anomaly detection,”Pattern Recognition, vol. 121, p. 108232, 2022

    2022

  7. [15]

    Rethinking prediction- based video anomaly detection from local–global normality perspective,

    M. Zhao, X. Zeng, Y . Liu, J. Liu, and C. Pang, “Rethinking prediction- based video anomaly detection from local–global normality perspective,” Expert Systems with Applications, vol. 262, p. 125581, 2025

    2025

  8. [16]

    Fast video anomaly detection via context-aware shortcut exploration and abnormal feature distance learning,

    C. Park, D. Kim, M. Cho, M. Kim, M. Lee, S. Park, and S. Lee, “Fast video anomaly detection via context-aware shortcut exploration and abnormal feature distance learning,”Pattern Recognition, vol. 157, p. 110877, 2025

    2025

  9. [17]

    Moba: Motion memory-augmented deblurring autoencoder for video anomaly detection,

    J. Lyu, M. Zhao, J. Hu, X. Huang, S. Du, C. Shi, and Z. Lv, “Moba: Motion memory-augmented deblurring autoencoder for video anomaly detection,”Knowledge-Based Systems, p. 115218, 2025

    2025

  10. [18]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. ...

    2015

  11. [19]

    Generative adversarial nets,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in neural information processing systems, vol. 27, 2014

    2014

  12. [20]

    Self-supervised attentive generative adversarial networks for video anomaly detection,

    C. Huang, J. Wen, Y . Xu, Q. Jiang, J. Yang, Y . Wang, and D. Zhang, “Self-supervised attentive generative adversarial networks for video anomaly detection,”IEEE transactions on neural networks and learning systems, vol. 34, no. 11, pp. 9389–9403, 2022

    2022

  13. [21]

    Video event restoration based on keyframes for video anomaly detection,

    Z. Yang, J. Liu, Z. Wu, P. Wu, and X. Liu, “Video event restoration based on keyframes for video anomaly detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 14 592–14 601

    2023

  14. [22]

    Object-guided and motion-refined atten- tion network for video anomaly detection,

    W. Zhou, Y . Li, and C. Zhao, “Object-guided and motion-refined atten- tion network for video anomaly detection,” in2022 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2022, pp. 1–6

    2022

  15. [23]

    Anomaly detection with bidirectional consistency in videos,

    Z. Fang, J. Liang, J. T. Zhou, Y . Xiao, and F. Yang, “Anomaly detection with bidirectional consistency in videos,”IEEE transactions on neural networks and learning systems, vol. 33, no. 3, pp. 1079–1092, 2020

    2020

  16. [24]

    Multi-encoder towards effective anomaly detection in videos,

    Z. Fang, J. T. Zhou, Y . Xiao, Y . Li, and F. Yang, “Multi-encoder towards effective anomaly detection in videos,”IEEE Transactions on Multimedia, vol. 23, pp. 4106–4116, 2021

    2021

  17. [25]

    Learning regularity in skeleton trajectories for anomaly detection in videos,

    R. Morais, V . Le, T. Tran, B. Saha, M. Mansour, and S. Venkatesh, “Learning regularity in skeleton trajectories for anomaly detection in videos,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 11 996–12 004

    2019

  18. [26]

    Making reconstruction-based method great again for video anomaly detection,

    Y . Wang, C. Qin, Y . Bai, Y . Xu, X. Ma, and Y . Fu, “Making reconstruction-based method great again for video anomaly detection,” in2022 IEEE International Conference on Data Mining (ICDM). IEEE, 2022, pp. 1215–1220

    2022

  19. [27]

    Anomaly detection in video se- quence with appearance-motion correspondence,

    T.-N. Nguyen and J. Meunier, “Anomaly detection in video se- quence with appearance-motion correspondence,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1273– 1283

    2019

  20. [28]

    Abnormal event detection in videos using generative adversarial nets,

    M. Ravanbakhsh, M. Nabi, E. Sangineto, L. Marcenaro, C. Regazzoni, and N. Sebe, “Abnormal event detection in videos using generative adversarial nets,” in2017 IEEE international conference on image processing (ICIP). IEEE, 2017, pp. 1577–1581

    2017

  21. [29]

    Nm-gan: Noise- modulated generative adversarial network for video anomaly detection,

    D. Chen, L. Yue, X. Chang, M. Xu, and T. Jia, “Nm-gan: Noise- modulated generative adversarial network for video anomaly detection,” Pattern Recognition, vol. 116, p. 107969, 2021

    2021

  22. [30]

    Detecting abnormality with separated foreground and background: Mutual generative adversarial networks for video abnormal event detection,

    Z. Zhang, S.-h. Zhong, A. Fares, and Y . Liu, “Detecting abnormality with separated foreground and background: Mutual generative adversarial networks for video abnormal event detection,”Computer Vision and Image Understanding, vol. 219, p. 103416, 2022

    2022

  23. [31]

    Improving video anomaly detection performance by mining useful data from unseen video frames,

    R. Wu, S. Li, C. Chen, and A. Hao, “Improving video anomaly detection performance by mining useful data from unseen video frames,” Neurocomputing, vol. 462, pp. 523–533, 2021

    2021

  24. [32]

    Future frame prediction network for video anomaly detection,

    W. Luo, W. Liu, D. Lian, and S. Gao, “Future frame prediction network for video anomaly detection,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 11, pp. 7505–7520, 2021

    2021

  25. [33]

    A video anomaly detection framework based on appearance-motion semantics representation consistency,

    X. Huang, C. Zhao, and Z. Wu, “A video anomaly detection framework based on appearance-motion semantics representation consistency,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  26. [34]

    Bidirectional spatio- temporal feature learning with multiscale evaluation for video anomaly detection,

    Y . Zhong, X. Chen, Y . Hu, P. Tang, and F. Ren, “Bidirectional spatio- temporal feature learning with multiscale evaluation for video anomaly detection,”IEEE Transactions on Circuits and Systems for Video Tech- nology, vol. 32, no. 12, pp. 8285–8296, 2022

    2022

  27. [35]

    Bman: Bidirectional multi-scale aggregation networks for abnormal event detection,

    S. Lee, H. G. Kim, and Y . M. Ro, “Bman: Bidirectional multi-scale aggregation networks for abnormal event detection,”IEEE Transactions on Image Processing, vol. 29, pp. 2395–2408, 2019

    2019

  28. [36]

    Lightning fast video anomaly detection via multi-scale adversarial distillation,

    F.-A. Croitoru, N.-C. Ristea, D. D ˘asc˘alescu, R. T. Ionescu, F. S. Khan, and M. Shah, “Lightning fast video anomaly detection via multi-scale adversarial distillation,”Computer Vision and Image Understanding, vol. 247, p. 104074, 2024

    2024

  29. [37]

    Prior knowledge guided network for video anomaly detection,

    Z. Deng, D. Chen, and S. Deng, “Prior knowledge guided network for video anomaly detection,” inProceedings of the 5th ACM International Conference on Multimedia in Asia, 2023, pp. 1–7

    2023

  30. [38]

    Vadmamba: Exploring state space models for fast video anomaly detection,

    J. Lyu, M. Zhao, J. Hu, X. Huang, Y . Chen, and S. Du, “Vadmamba: Exploring state space models for fast video anomaly detection,” in2025 IEEE International Conference on Multimedia and Expo (ICME), 2025, pp. 1–6

    2025

  31. [39]

    Gated attention for large language models: Non-linearity, sparsity, and attention-sink-free,

    Z. Qiu, Z. Wang, B. Zheng, Z. Huang, K. Wen, S. Yang, R. Men, L. Yu, F. Huang, S. Huanget al., “Gated attention for large language models: Non-linearity, sparsity, and attention-sink-free,”arXiv preprint arXiv:2505.06708, 2025

    2025 arXiv

  32. [40]

    Gt-had: Gated transformer for hyperspectral anomaly detection,

    J. Lian, L. Wang, H. Sun, and H. Huang, “Gt-had: Gated transformer for hyperspectral anomaly detection,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 2, pp. 3631–3645, 2024

    2024

  33. [41]

    Dual-branch learning with prior information for surface anomaly detection,

    S. Wang, C. Lv, Z. Zhang, and X. Wei, “Dual-branch learning with prior information for surface anomaly detection,”IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–11, 2023

    2023

  34. [42]

    Attention-based residual autoencoder for video anomaly detection,

    V .-T. Le and Y .-G. Kim, “Attention-based residual autoencoder for video anomaly detection,”Applied Intelligence, vol. 53, no. 3, pp. 3240–3254, 2023

    2023

  35. [43]

    Anomaly de- tection in crowded scenes,

    V . Mahadevan, W. Li, V . Bhalodia, and N. Vasconcelos, “Anomaly de- tection in crowded scenes,” in2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2010, pp. 1975–1981

    2010

  36. [44]

    Abnormal event detection at 150 fps in matlab,

    C. Lu, J. Shi, and J. Jia, “Abnormal event detection at 150 fps in matlab,” inProceedings of the IEEE international conference on computer vision, 2013, pp. 2720–2727

    2013

  37. [45]

    A revisit of sparse coding based anomaly detection in stacked rnn framework,

    W. Luo, W. Liu, and S. Gao, “A revisit of sparse coding based anomaly detection in stacked rnn framework,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 341–349

    2017

  38. [46]

    Dynamic local aggregation network with adaptive clusterer for anomaly detection,

    Z. Yang, P. Wu, J. Liu, and X. Liu, “Dynamic local aggregation network with adaptive clusterer for anomaly detection,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 404–421

    2022

  39. [47]

    Hybrid attention and motion constraint for anomaly detection in crowded scenes,

    X. Zhang, J. Fang, B. Yang, S. Chen, and B. Li, “Hybrid attention and motion constraint for anomaly detection in crowded scenes,”IEEE Transactions on Circuits and Systems for Video Technology, 2022

    2022

  40. [48]

    Spatial- temporal graph convolutional network boosted flow-frame prediction for video anomaly detection,

    K. Cheng, X. Zeng, Y . Liu, M. Zhao, C. Pang, and X. Hu, “Spatial- temporal graph convolutional network boosted flow-frame prediction for video anomaly detection,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  41. [49]

    Memorizing normality to detect anomaly: Memory-augmented deep autoencoder video anomaly detection,

    Y . Liu, D. Li, W. Zhu, D. Yang, J. Liu, and L. Song, “Memorizing normality to detect anomaly: Memory-augmented deep autoencoder video anomaly detection,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  42. [50]

    Long short-term dynamic proto- type alignment learning for video anomaly detection,

    C. Huang, J. Wen, C. Liu, and Y . Liu, “Long short-term dynamic proto- type alignment learning for video anomaly detection,” inProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024, pp. 866–874

    2024

  43. [51]

    Multi-branch gan-based abnormal events detection via context learning in surveillance videos,

    D. Li, X. Nie, R. Gong, X. Lin, and H. Yu, “Multi-branch gan-based abnormal events detection via context learning in surveillance videos,” IEEE Transactions on Circuits and Systems for Video Technology, 2023

    2023

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.