Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

Surrogate-Enhanced Modeling and Adaptive Modular Control of All-Electric Heavy-Duty Robotic Manipulators

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Weakly-supervised video anomaly detection reaches 91.58% AUC on UCF-Crime when a mixture of per-class experts is trained on temporal Gaussian pseudo-labels derived from the model's own score peaks.

desk verdict The record is two different papers: the abstract promises robotics, the full text is a WSVAD paper; as a submission it is broken, though the body has real merit on its own. read the letter →

arxiv 2508.06313 v1 pith:BWEEGL2P submitted 2025-08-08 cs.RO

Amir Hossein Barjini , Mohammad Bahari , Mahdi Hejrati , Jouni Mattila This is my paper
classification cs.RO
keywords weakly-supervisedvideoanomalydetectiontemporalGaussiansplattingmixtureofexpertspseudo-labelingUCF-CrimeXD-ViolenceMSADmulti-instancelearning
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's full text is a weakly-supervised video anomaly detection (WSVAD) study. It tries to show that the two main weaknesses of current WSVAD—diverse anomaly types and weak labels that lack temporal detail—can be addressed together: temporal Gaussian splatting reconstructs each anomalous event as a Gaussian pseudo-label over time, and a mixture-of-experts model routes each anomaly class to a dedicated expert whose scores are fused by a gate. On UCF-Crime, this reaches 91.58% AUC and 83.86% anomaly-only AUC; on XD-Violence, the anomaly-only AP reaches 85.74%; on MSAD, AUC is 87.72%. If correct, the result means subtle anomalies such as shoplifting can be localized with only video-level labels, without extra frame-level annotation.

What carries the argument

The central mechanism is the Temporal Gaussian Splatting (TGS) loss. It detects local maxima (peaks) in the anomaly score sequence, builds a Gaussian kernel around each peak with a width determined by monotonic score increases and decreases, and renders these kernels over the whole video length to create a dense target score for every snippet. This is paired with a Mixture-of-Experts architecture in which each anomaly class has a dedicated transformer expert and a gate model fuses expert scores with task-aware features via bidirectional cross-attention. The TGS loss turns weak video-level labels into time-continuous targets; the experts specialize per anomaly class; the gate integrates that

What would settle it

A diagnostic experiment: on the UCF-Crime training set, build the Gaussian pseudo-labels from the ground-truth anomaly time windows (an 'oracle peak') instead of from the model's predicted peaks, and compare downstream AUC. If oracle peaks do not raise AUC above 91.58%, then the gain attributed to TGS is not coming from the accuracy of peak detection; if they do, the model's self-derived peaks are the bottleneck.

Watch

Extended reading notes

Core claim

The paper claims that a single shared detector trained with top-k multiple-instance learning overlooks the variety of anomaly classes and the full time span of each event. GS-MoE replaces top-k supervision with Temporal Gaussian Splatting: peaks in the model's own anomaly scores become Gaussian pseudo-labels covering each entire anomaly. A dedicated transformer expert per anomaly class learns fine-grained cues, and a gate model with bidirectional cross-attention fuses expert scores with coarse anomaly-aware features. Reported results: 91.58% AUC and 83.86% anomaly-only AUC on UCF-Crime, with gains on XD-Violence and MSAD; ablations show removing the gate, experts, or TGS loss each lowers per

Load-bearing premise

The load-bearing premise is that peaks in the model's own anomaly scores—after a one-epoch MIL warm-up and under a 0.2 prominence threshold—reliably outline real anomalous events, because those peaks become the Gaussian pseudo-labels that train the experts and gate; if early peaks miss or invent events, the self-training loop locks in the error.

Editorial extensions

If this is right

  • Weakly-supervised video anomaly detection can be advanced by pseudo-labeling the whole temporal extent of an anomaly rather than only its top-scoring snippets; the paper demonstrates this with the TGS loss on three datasets.
  • Per-class expert models, not just a single shared model, are what let the framework localize subtle anomalies like shoplifting and stealing; ablations show that masking an expert collapses that class's AUC to about 50%.
  • Existing detectors can be upgraded by fine-tuning with the TGS loss: the UR-DMU baseline used as a feature extractor improves when trained with TGS before the experts are added.
  • When class labels are unavailable, cluster-based experts still beat published baselines on UCF-Crime, suggesting the method does not strictly require predefined anomaly classes.
  • The reported computational cost of 13 experts plus the gate stays workable (9.57 FPS in the sequential implementation, with a parallel version noted as near real-time), so the accuracy gains do not require impractical hardware.

Reading between the lines

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

  • The record's title and abstract describe a robotics control paper (sub-centimeter tracking, Lyapunov stability), but the body text is this video-anomaly paper; if the title and abstract are taken as the declaration, that tracking claim has no derivation or experiment in the body.
  • Because TGS pseudo-labels are produced by the model's own scores, the training loop is self-referential; comparing against oracle peak locations would show how much headroom remains in the peak-detection step.
  • The TGS-plus-experts recipe should transfer to other weakly supervised temporal localization tasks, such as surgical phase detection or action segmentation, where video-level labels are cheap and event boundaries are ambiguous.
  • The cluster-experts results suggest that class labels are not strictly necessary; users with unknown anomaly taxonomies could choose cluster count as a hyperparameter, accepting a smaller gain over the best single-model baseline.
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 / 4 minor

Summary. The manuscript as submitted (arXiv:2508.06313) declares in its title and abstract a surrogate-enhanced modeling and adaptive modular control framework for all-electric heavy-duty robotic manipulators, claiming sub-centimeter Cartesian tracking accuracy, a Lyapunov-based stability proof, and 1-DoF experimental validation. The full text, however, is an entirely different paper, 'Mixture of Experts Guided by Gaussian Splatters Matters', on weakly-supervised video anomaly detection (WSVAD). The body contains no mention of electromechanical actuators, virtual decomposition control, Lyapunov functions, or Cartesian tracking. Instead, it proposes a GS-MoE architecture: temporal Gaussian splatting (TGS) generates pseudo-labels from the model's own anomaly-score peaks, and class-specific experts plus a gate produce final anomaly scores. Experiments on UCF-Crime, XD-Violence, and MSAD report a UCF-Crime AUC of 91.58% and an anomaly-only AUC_A of 83.86%. Thus the submitted record consists of a declared robotics contribution with no supporting derivation and a body that addresses a different problem.

Significance. If the WSVAD body were the intended submission, the GS-MoE contribution would have moderate interest: the idea of using Gaussian-kernel pseudo-labels derived from score peaks to densify the weak supervision signal is reasonable, the component ablations in Table 3 are monotone, the expert-masking experiment in Table 4 is informative, and the supplementary threshold sweep is an honest sensitivity check. The code link for GS-MoE is provided. However, the significance of the declared robotics result cannot be evaluated because no equation, experiment, or derivation in the body concerns the claimed system. Moreover, the WSVAD results rest on a self-training loop whose pseudo-label reliability is not directly validated, and all numbers are single-run point estimates without error bars or seed variation. These issues would need to be resolved before the body could be considered a solid standalone contribution, and they are entirely separate from the absent robotics content.

major comments (3)
  1. [Title and Abstract vs. Full Text] The declared central claim of the paper—sub-centimeter Cartesian tracking accuracy of an all-electric heavy-duty robotic manipulator via a surrogate-enhanced EMLA model embedded in virtual decomposition control with a Lyapunov-based stability proof—has no supporting content anywhere in the body. Equations (3)–(7) define Gaussian pseudo-labels and the WSVAD loss; Tables 1–6 report AUC/AP on UCF-Crime, XD-Violence, and MSAD. No section, equation, or experiment addresses electromechanical actuators, virtual decomposition control, Lyapunov functions, or Cartesian tracking. The abstract's central claim is therefore unverifiable from the submitted text; this is a load-bearing missing derivation, not a presentation issue.
  2. [Sec. 3.1, Eqs. (3)–(5)] The TGS pseudo-labels are generated by detecting peaks in the model's own anomaly scores after only a one-epoch MIL warm-up, and these pseudo-labels are then used to train the same experts and gate. This is a self-training loop: if early peaks are missing or spurious, the error is propagated and locked in. The paper does not report peak precision/recall against frame-level ground truth, and the supplementary threshold ablation (Table 1) shows AUC sensitivity but does not validate whether the detected peaks correspond to real anomalous events. Since the headline results, especially AUC_A 83.86% on UCF-Crime, depend on pseudo-label quality, this missing validation is load-bearing.
  3. [Tables 1, 3, and Suppl. Table 1] All reported experiments are single-run point estimates. No seeds, error bars, or significance tests are provided for the UCF-Crime, XD-Violence, or MSAD results. The claimed state-of-the-art gains (e.g., 91.58 vs. 86.97 AUC, and 83.86 vs. 70.81 AUC_A in Table 1) and the component ablations in Table 3 could be affected by run-to-run variance, which is not assessed. This is especially relevant because the self-training procedure introduces sensitivity to initialization and peak-detection thresholds. At minimum, multiple seeds and standard deviations are needed to support the magnitude of the claimed improvements.
minor comments (4)
  1. [Eqs. (1)–(7)] Several equations contain unreadable or placeholder characters (e.g., the MIL loss formulas and the 'top k' notation), making it difficult to verify the exact optimization objective. Please render all equations cleanly and ensure Eq. (6) and Eq. (7) are clearly distinguished.
  2. [Table 4 caption] The caption says 'GS-MoE without the expert model for a given class' but the rows labeled 'Mask' and 'W/o Mask' seem reversed in the text description. Please clarify the experimental setup: is 'Mask' the condition where the expert's score is removed from the gate's input?
  3. [Section 4.1 and Figure 5] The text says performance boosts 'up to ��� ���' and other percentages appear as '�� ����' or '�� ����' throughout Section 4.3. These values are missing or corrupted in the submitted text and need to be filled in.
  4. [Supplementary Material, Sec. 2] The threshold ablation reports AUC for thresholds 0.1–0.3 but does not describe how many peaks are detected per video or how peak detection quality changes during training. A short analysis would help the reader trust the pseudo-labeling step.

Circularity Check

1 steps flagged · score 4.0 of 10

TGS pseudo-labels are derived from the model's own anomaly scores, creating a partial self-training loop; the headline benchmark remains externally validated.

  1. self definitional [Section 3.1, 'Peak detection' and Eqs. (3)-(6); Supplementary Table 1]
    "Gaussian kernels are extracted in correspondence with local maxima, called "peaks", in the temporal axis of the abnormal scores predicted by a model. ... Finally, the pseudo-labels y_i are generated by rendering each of the K extracted kernels over the length of the video [Eq. 5]. ... the TGS loss function used to train the experts and the MoE is formulated as [Eq. 6]."

    The training target in Eq. 5 is a deterministic function of the model's own current anomaly scores: peaks are local maxima of s_t, widths are monotonic runs of s_t, and kernels are thresholded against s_t (Eqs. 3-4). Eq. 6 then trains the same model to match that target. Minimization therefore drives s toward a fixed point s = smooth(peaks(s)), rewarding self-consistent peak structure rather than agreement with external frame labels. The MIL warm-up and prominence threshold (0.2) mitigate spurious peaks, but the loop is still closed; the supplementary threshold ablation (Table 1) reports only AUC, not peak precision/recall against ground truth, so the targets' correctness is not independently verified. The held-out benchmark AUCs provide independent evidence that the loop did not collapse,

full rationale

The derivation chain of GS-MoE's central novelty (TGS) is partially self-referential: the pseudo-labels that supervise the experts and gate are generated from peaks in the model's own scores, so the method rewards its own confident patterns. This is a genuine self-definitional loop (Eqs. 3-6). However, TGS is only a training-time label-construction device; final AUC/AP are measured on held-out test sets with standard protocols, and the I3D/UR-DMU features and MIL warm-up inject external signal, so the headline 91.58% AUC is not forced by the loop. I find no load-bearing self-citation, no imported uniqueness theorem, and no renaming of a known result; the 'following [50]' citation is an independent prior. Separately, I flag a severe non-circularity flaw: the abstract claims a surrogate-enhanced EMLA/VDC robotics controller with Lyapunov proof and sub-centimeter tracking, but the full text is an unrelated WSVAD paper (arXiv:2508.06318) with no HDRM equations or experiments; that missing support is an integrity/correctness issue, not circularity, and I do not add it to the circularity score. Net score 4/10: partial self-training circularity, with the central benchmark claim still independently testable.

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

The body text's approach rests on four inherited or assumed building blocks: pretrained I3D features, the MIL warm-up objective, the assumption that anomalies have peak-like temporal profiles (Gaussian kernels), and the reliability of fine-tuned UR-DMU features. The trainable parameters of the experts and gate (about 13-16M total) are learned, while the pseudo-label generator introduces hand-set hyperparameters: peak prominence threshold (0.2, tuned by validation AUC), kernel bandwidth rule, expert count (13 classes or cluster count tuned 5-15), and one-epoch MIL warm-up. No new physical entities are introduced. For the declared robotics abstract, the ledger cannot be completed because the body does not describe the surrogate model, its training data, or the control gains.

free parameters (4)
  • Peak prominence threshold = 0.2 (ablation range 0.1-0.3)
    Threshold that decides which local maxima of the model's anomaly score become Gaussian kernel seeds (Section 3.1); supplementary Table 1 shows 0.2 is selected by validation AUC, so it is tuned to the benchmark.
  • Gaussian kernel width and sigma = per-peak, computed from score profile
    Kernel bandwidth in Eq. 3-4 is a modeling choice that shapes the pseudo-label targets; not fixed by theory.
  • Number of experts / clusters = 13 class experts; 5-15 clusters for cluster variants
    Class count defines experts for UCF-Crime; for cluster-experts the number K is chosen by AUC (Table 6 picks 7 clusters), i.e., tuned to the test metric.
  • MIL warm-up duration = 1 epoch
    First epoch uses only the MIL objective before TGS activation (supplementary Section 1); a schedule choice affecting which peaks exist early in training.
assumptions (4)
  • domain assumption I3D features pretrained on Kinetics-400 are a sufficient task-agnostic video representation
    Backbone for all features (supp. Section 1); standard in WSVAD, inherited from cited works.
  • ad hoc to paper Anomalous temporal windows are well-approximated by Gaussian kernels centered on score peaks
    Core postulate of TGS (Eq. 3-5); if real anomalies produce flat or multi-modal score profiles, the pseudo-labels are inaccurate.
  • domain assumption The MIL top-k objective is a valid warm-up baseline
    The paper warms up with the MIL loss (Section 3.1, supp.) and assumes it reliably bootstraps discriminative scores before TGS takes over.
  • domain assumption UR-DMU fine-tuned features carry class-discriminative anomaly signal
    Experts and gate inputs are UR-DMU task-aware features (Section 3.2); errors in the task encoder propagate into the experts and gate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Surrogate-Enhanced Modeling and Adaptive Modular Control of All-Electric Heavy-Duty Robotic Manipulators." pith.science (2026). https://pith.science/paper/BWEEGL2P

@misc{pith2026250806313,
  author       = {Pith},
  title        = {Pith review of: Surrogate-Enhanced Modeling and Adaptive Modular Control of All-Electric Heavy-Duty Robotic Manipulators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BWEEGL2P}},
  note         = {Machine review of arXiv:2508.06313}
}
read the original abstract

This paper presents a unified system-level modeling and control framework for an all-electric heavy-duty robotic manipulator (HDRM) driven by electromechanical linear actuators (EMLAs). A surrogate-enhanced actuator model, combining integrated electromechanical dynamics with a neural network trained on a dedicated testbed, is integrated into an extended virtual decomposition control (VDC) architecture augmented by a natural adaptation law. The derived analytical HDRM model supports a hierarchical control structure that seamlessly maps high-level force and velocity objectives to real-time actuator commands, accompanied by a Lyapunov-based stability proof. In multi-domain simulations of both cubic and a custom planar triangular trajectory, the proposed adaptive modular controller achieves sub-centimeter Cartesian tracking accuracy. Experimental validation of the same 1-DoF platform under realistic load emulation confirms the efficacy of the proposed control strategy. These findings demonstrate that a surrogate-enhanced EMLA model embedded in the VDC approach can enable modular, real-time control of an all-electric HDRM, supporting its deployment in next-generation mobile working machines.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Adaptive Modular Geometric Control of Robotic Manipulators

    eess.SY 2026-03 unverdicted novelty 7.0 of 10

    An adaptive modular geometric controller for robotic manipulators achieves exponential stability in nominal cases and reduces RMS position error by at least 12.2% versus state-of-the-art methods in simulations while u...

  2. Adaptive Modular Geometric Control of Robotic Manipulators

    eess.SY 2026-03 conditional novelty 6.0 of 10

    A modular SE(3)-based adaptive controller for manipulators is claimed to guarantee exponential tracking in the nominal case and uniformly ultimately bounded tracking under uncertain inertia.

Reference graph

Works this paper leans on

65 extracted references · 58 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ubnor- mal: New benchmark for supervised open-set video anomaly detection

    Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Ubnor- mal: New benchmark for supervised open-set video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 3

    work page 2022

  2. [2]

    What’s the point: Semantic segmentation with point supervision

    Amy Bearman, Olga Russakovsky, Vittorio Ferrari, and Li Fei-Fei. What’s the point: Semantic segmentation with point supervision. In European conference on computer vision , pages 549–565. Springer, 2016. 2

    work page 2016

  3. [3]

    Is space-time attention all you need for video understanding? In ICML, page 4, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 3

    work page 2021

  4. [4]

    Quo vadis, action recognition? a new model and the kinetics dataset

    Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In The IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2017. 1

    work page 2017

  5. [5]

    Deep analysis of cnn-based spatio-temporal representa- tions for action recognition

    Chun-Fu Richard Chen, Rameswar Panda, Kandan Ramakr- ishnan, Rogerio Feris, John Cohn, Aude Oliva, and Quanfu Fan. Deep analysis of cnn-based spatio-temporal representa- tions for action recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 6165–6175, 2021. 1

    work page 2021

  6. [6]

    Tevad: Improved video anomaly de- tection with captions

    Weiling Chen, Keng Teck Ma, Zi Jian Yew, Minhoe Hur, and David Aik-Aun Khoo. Tevad: Improved video anomaly de- tection with captions. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 5549–5559, 2023. 6

    work page 2023

  7. [7]

    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. In Proceedings of the AAAI Con- ference on Artificial Intelligence, pages 387–395, 2023. 6

    work page 2023

  8. [8]

    Look around for anomalies: Weakly-supervised anomaly detection via context-motion relational learning

    MyeongAh Cho, Minjung Kim, Sangwon Hwang, Chae- won Park, Kyungjae Lee, and Sangyoun Lee. Look around for anomalies: Weakly-supervised anomaly detection via context-motion relational learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12137–12146, 2023. 6

    work page 2023

Show all 65 references
  1. [9]

    Learning factored representations in a deep mixture of ex- perts

    David Eigen, Marc’Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of ex- perts. arXiv preprint arXiv:1312.4314, 2013. 3

    2013 arXiv

  2. [10]

    Switch transformers: Scaling to trillion parameter models with sim- ple and efficient sparsity

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with sim- ple and efficient sparsity. Journal of Machine Learning Re- search, 23(120):1–39, 2022. 3

    2022

  3. [11]

    Mist: Multiple instance self-training framework for video anomaly detection

    Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 14009– 14018, 2021. 6

    2021

  4. [12]

    Anomaly detection in video via self- supervised and multi-task learning

    Mariana-Iuliana Georgescu, Antonio Barbalau, Radu Tu- dor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly detection in video via self- supervised and multi-task learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognitio...

    2021

  5. [13]

    A background-agnostic framework with adversarial training for abnormal event detection in video

    Mariana Iuliana Georgescu, Radu Tudor Ionescu, Fa- had Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE transactions on pattern analysis and machine intelligence, 44(9):4505–4523,

  6. [14]

    Bridging nonlinearities and stochastic regularizers with gaussian error linear units

    Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. CoRR, abs/1606.08415, 2016. 5 9

    2016 arXiv

  7. [15]

    Mixture of nested experts: Adaptive processing of visual to- kens

    Gagan Jain, Nidhi Hegde, Aditya Kusupati, Arsha Nagrani, Shyamal Buch, Prateek Jain, Anurag Arnab, and Sujoy Paul. Mixture of nested experts: Adaptive processing of visual to- kens. arXiv preprint arXiv:2407.19985, 2024. 3

    2024 arXiv

  8. [16]

    Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection

    Hyekang Kevin Joo, Khoa V o, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230–3234. IEEE, 2023. 6, 7, 9

    2023

  9. [17]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1,

  10. [18]

    Point-based neural rendering with per- view optimization

    Georgios Kopanas, Julien Philip, Thomas Leimk ¨uhler, and George Drettakis. Point-based neural rendering with per- view optimization. In Computer Graphics Forum, pages 29–

  11. [19]

    Gshard: Scaling giant models with conditional computation and automatic sharding

    Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 3

    2006 arXiv

  12. [20]

    St-4dgs: Spatial-temporally consistent 4d gaus- sian splatting for efficient dynamic scene rendering

    Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. St-4dgs: Spatial-temporally consistent 4d gaus- sian splatting for efficient dynamic scene rendering. In ACM SIGGRAPH 2024 Conference Papers, New York, NY , USA,

    2024

  13. [21]

    Scale-aware spatio-temporal relation learning for video anomaly detection

    Guoqiu Li, Guanxiong Cai, Xingyu Zeng, and Rui Zhao. Scale-aware spatio-temporal relation learning for video anomaly detection. In European Conference on Computer Vision, pages 333–350. Springer, 2022. 6

    2022

  14. [22]

    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. In Proceedings of the AAAI Con- ference on Artificial Intelligence, pages 1395–1403, 2022. 6

    2022

  15. [23]

    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. In Proceedings of the AAAI Con- ference on Artificial Intelligence, pages 1395–1403, 2022. 2

    2022

  16. [24]

    Spacetime gaus- sian feature splatting for real-time dynamic view synthesis

    Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaus- sian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508–8520, 2024. 3

    2024

  17. [25]

    Least squares quantization in pcm

    Stuart Lloyd. Least squares quantization in pcm. IEEE trans- actions on information theory, 28(2):129–137, 1982. 8

    1982

  18. [26]

    Fixing weight decay reg- ularization in adam

    Ilya Loshchilov and Frank Hutter. Fixing weight decay reg- ularization in adam. CoRR, abs/1711.05101, 2017. 1

    2017 arXiv

  19. [27]

    Localizing anomalies from weakly-labeled videos

    Hui Lv, Chuanwei Zhou, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Localizing anomalies from weakly-labeled videos. IEEE transactions on image processing , 30:4505– 4515, 2021. 6

    2021

  20. [28]

    Unbiased multiple instance learning for weakly supervised video anomaly detection

    Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 8022–8031, 2023. 6

    2023

  21. [29]

    Human- scene network: A novel baseline with self-rectifying loss for weakly supervised video anomaly detection

    Snehashis Majhi, Rui Dai, Quan Kong, Lorenzo Garat- toni, Gianpiero Francesca, and Francois Bremond. Human- scene network: A novel baseline with self-rectifying loss for weakly supervised video anomaly detection. Computer Vi- sion and Image Understanding, 241:103955, 2024. 6

    2024

  22. [30]

    Oe- ctst: Outlier-embedded cross temporal scale transformer for weakly-supervised video anomaly detection

    Snehashis Majhi, Rui Dai, Quan Kong, Lorenzo Garat- toni, Gianpiero Francesca, and Franc ¸ois Br ´emond. Oe- ctst: Outlier-embedded cross temporal scale transformer for weakly-supervised video anomaly detection. In Proceedings of the IEEE/CVF winter conference on applications ...

    2024

  23. [31]

    Just dance with pi! a poly-modal inductor for weakly-supervised video anomaly detection

    Snehashis Majhi, Giacomo D’Amicantonio, Antitza Dantcheva, Quan Kong, Lorenzo Garattoni, Gianpiero Francesca, Egor Bondarev, and Francois Bremond. Just dance with pi! a poly-modal inductor for weakly-supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference...

    2025

  24. [32]

    Multimodal contrastive learn- ing with limoe: the language-image mixture of experts

    Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learn- ing with limoe: the language-image mixture of experts. Ad- vances in Neural Information Processing Systems, 35:9564– 9576, 2022. 3

    2022

  25. [33]

    From sparse to soft mixtures of experts, 2024

    Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts, 2024. 3, 2

    2024

  26. [34]

    Dance with self-attention: A new look of conditional ran- dom fields on anomaly detection in videos

    Didik Purwanto, Yie-Tarng Chen, and Wen-Hsien Fang. Dance with self-attention: A new look of conditional ran- dom fields on anomaly detection in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 173–183, 2021. 6

    2021

  27. [35]

    Scaling vision with sparse mix- ture of experts

    Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, Andr ´e Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mix- ture of experts. Advances in Neural Information Processing Systems, 34:8583–8595, 2021. 3

    2021

  28. [36]

    Real-world anomaly detection in surveillance videos

    Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6479–6488, 2018. 2, 6

    2018

  29. [37]

    Real-world anomaly detection in surveillance videos

    Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Proceedings of the IEEE conference on computer vision and pattern recog- nition, pages 6479–6488, 2018. 1, 2, 3

    2018

  30. [38]

    Weakly-supervised video anomaly detection with robust temporal feature magni- tude learning

    Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magni- tude learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4975–49...

    2021

  31. [39]

    Weakly-supervised video anomaly detection with robust temporal feature magni- tude learning

    Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magni- tude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975–49...

    2021

  32. [40]

    Visualizing data using t-sne

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579–2605, 2008. 7

    2008

  33. [41]

    Not only look, but also 10 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 10 listen: Learning multimodal violence detection under weak supervision. In European Conference on Computer Vision , pages 322–339. Springer, 2020. 2, 6

    2020

  34. [42]

    Weakly supervised audio-visual violence detection

    Peng Wu, Xiaotao Liu, and Jing Liu. Weakly supervised audio-visual violence detection. IEEE Transactions on Mul- timedia, 25:1674–1685, 2022. 1

    2022

  35. [43]

    Wiley Online Library, 2021. 3

    2021

  36. [44]

    Towards high- quality hdr deghosting with conditional diffusion models

    Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high- quality hdr deghosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Tech- nology, 2023. 1

    2023

  37. [45]

    Vadclip: Adapting vision-language models for weakly supervised video anomaly detection

    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. In Proceedings of the AAAI Con- ference on Artificial Intelligence, pages 6074–6082, 2024. 6, 9

    2024

  38. [46]

    Cloze test helps: Effec- tive video anomaly detection via learning to complete video events

    Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effec- tive video anomaly detection via learning to complete video events. In Proceedings of the 28th ACM international con- ference on multimedia, pages 583–591, 2020. 1

    2020

  39. [47]

    Text prompt with nor- mality guidance for weakly supervised video anomaly detec- tion

    Zhiwei Yang, Jing Liu, and Peng Wu. Text prompt with nor- mality guidance for weakly supervised video anomaly detec- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 18899– 18908, 2024. 6, 9

    2024

  40. [48]

    Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection

    Chen Zhang, Guorong Li, Yuankai Qi, Shuhui Wang, Laiyun Qing, Qingming Huang, and Ming-Hsuan Yang. Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2023

  41. [49]

    Modality-aware contrastive instance learning with self-distillation for weakly-supervised audio-visual violence detection

    Jiashuo Yu, Jinyu Liu, Ying Cheng, Rui Feng, and Yuejie Zhang. Modality-aware contrastive instance learning with self-distillation for weakly-supervised audio-visual violence detection. In Proceedings of the 30th ACM international con- ference on multimedia, pages 6278–6287, 2022. 6

    2022

  42. [50]

    Glancevad: Exploring glance su- pervision for label-efficient video anomaly detection

    Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance su- pervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 2, 4

    2024 arXiv

  43. [51]

    Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection

    Chen Zhang, Guorong Li, Yuankai Qi, Shuhui Wang, Laiyun Qing, Qingming Huang, and Ming-Hsuan Yang. Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2023

  44. [52]

    Togs: Gaussian splatting with temporal opacity offset for real-time 4d dsa rendering, 2024

    Shuai Zhang, Huangxuan Zhao, Zhenghong Zhou, Guanjun Wu, Chuansheng Zheng, Xinggang Wang, and Wenyu Liu. Togs: Gaussian splatting with temporal opacity offset for real-time 4d dsa rendering, 2024. 3

    2024

  45. [53]

    Temporal con- volutional network with complementary inner bag loss for weakly supervised anomaly detection

    Jiangong Zhang, Laiyun Qing, and Jun Miao. Temporal con- volutional network with complementary inner bag loss for weakly supervised anomaly detection. In 2019 IEEE In- ternational Conference on Image Processing (ICIP) , pages 4030–4034. IEEE, 2019. 6

    2019

  46. [54]

    Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection

    Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1237–1246,

  47. [55]

    Li, and Ge Li

    Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H. Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6

    2019

  48. [56]

    Dual memory units with uncertainty regulation for weakly supervised video anomaly detection

    Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3769–3777, 2023. 5

    2023

  49. [57]

    Dual memory units with uncertainty regulation for weakly supervised video anomaly detection

    Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection. arXiv preprint arXiv:2302.05160, 2023. 5, 6, 7, 9, 3

    2023 arXiv

  50. [58]

    Motion-aware feature for improved video anomaly detection

    Yi Zhu and Shawn Newsam. Motion-aware feature for improved video anomaly detection. arXiv preprint arXiv:1907.10211, 2019. 6 11 Mixture of Experts Guided by Gaussian Splatters Matters: A new Approach to Weakly-Supervised Video Anomaly Detection Supplementary Material We includ...

    1907 arXiv

  51. [59]

    Advancing video anomaly detection: A concise re- view and a new dataset

    Liyun Zhu, Lei Wang, Arjun Raj, Tom Gedeon, and Chen Chen. Advancing video anomaly detection: A concise re- view and a new dataset. In The Thirty-eighth Conference on Neural Information Processing Systems Datasets and Bench- marks Track, 2024. 2, 6

    2024

  52. [61]

    The I3D implementation chosen is the ResNet50, which is proven to be one of the best-performing [5]

    Implementation Details The video features were obtained with the I3D model [4] pre-trained on Kinetics-400 with sliding windows of 16 frames. The I3D implementation chosen is the ResNet50, which is proven to be one of the best-performing [5]. The transformer blocks implemented...

  53. [62]

    We conducted experiments with different peaks thresholds to evaluate the sensitivity of our approach to the selection of peaks

    TGS Ablation Study As mentioned in Section 3.1, in order to mitigate the pres- ence of spurious peaks, a model trained with TGS has to be warmed up using the standard MIL loss function or the � ���� � ���� component. We conducted experiments with different peaks thresholds to ...

  54. [63]

    The framework, shown in Figure 1, differs from the Gating model detailed in the main paper by the strategy used to leverage the expert’s predictions

    Soft MoE In order to provide an overview of the capabilities of the proposed GS-MoE framework, we implement the same training strategy with soft-MoE, a modern MoE architec- ture introduced by [33]. The framework, shown in Figure 1, differs from the Gating model detailed in the...

  55. [64]

    Shoplifting-015

    Qualitative results - Failure Cases In Figure 3 we report some examples of videos on which GS-MoE is unable to correctly detect the anomalous portion of the video. We identify three main failure cases: false negative, false positives and long peaks. In WSV AD, a false negative...

  56. [65]

    This dataset is composed of synthetic videos generated in 29 different scenes

    UBnormal Experiments In order to present a more comprehensive overview of the performance of GS-MoE, we experiment on the UBNormal dataset [1]. This dataset is composed of synthetic videos generated in 29 different scenes. We experiment on this dataset in order to show the eff...

  57. [2024]

    Association for Computing Machinery. 3

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.