Pith. sign in

REVIEW 4 major objections 4 minor 37 references

Machine and Deep Learning for Crowd Analytics

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a spatio-temporal Gaussian dynamic-texture model can flag anomalous crowd events by thresholding the likelihood of learned normal motion, reporting 88.83% average accuracy on the University of Minnesota crowd dataset.

desk verdict This is a placeholder draft, not a paper: it promises deep learning, delivers no equations or figures, and its only quantitative claim is unverifiable and likely in-sample. read the letter →

arxiv 1909.04150 v1 pith:WWFSKP35 submitted 2019-08-25 cs.CV

Muhammad Siraj This is my paper · ORCID
classification cs.CV
keywords crowdanalyticsanomalydetectiondynamictexturemodelspatio-temporalcubesGaussianlikelihoodthresholdingUniversityofMinnesotadatasetdeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a traditional machine-learning model—one that learns how pixels move together across space and time—can detect anomalous crowd events in surveillance video without tracking individuals. Its method models each local block of video as a dynamic texture, a linear dynamic system whose learned mean vector and covariance matrix describe normal crowd motion; a frame is judged anomalous when its likelihood under that learned distribution falls below a threshold. The paper reports that this method reaches 88.83% average accuracy on the University of Minnesota crowd dataset, using the GROUND sequence for performance analysis. If the claim holds, it matters because it offers a lightweight, scene-specific complement to deep-learning crowd analytics, which require large labeled datasets. The paper also argues that the learned distribution can be extended from a single frame to larger chunks through AR, MA, or ARMA process models.

What carries the argument

The central object is the spatio-temporal Gaussian/dynamic-texture model: video is cut into cubes with spatial size $p$ and temporal size $q$, each cube is fit by a linear dynamic system, and the crowd's normal behavior is encoded as a mean vector and covariance matrix for the motion pattern. The argument runs on likelihood thresholding—compute the probability that a new frame's motion comes from the learned distribution, and call it anomalous when that probability is low—with model parameters updated through partial derivatives of the model with respect to feature functions. This is the machinery that carries the reported 88.83% accuracy figure.

What would settle it

First, check whether the exact dynamic-texture equations can be reconstructed from the manuscript, since they are not displayed; then run the described procedure on the University of Minnesota GROUND sequence with an explicit disjoint train/test split and compare the likelihood-threshold labels to the ground-truth anomaly tags. If the reported 88.83% cannot be reproduced under a disjoint split, the number is a training fit, not a generalization result.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that crowd motion can be absorbed into a Gaussian spatio-temporal model: the collection of moving elements in a video is represented as spatio-temporal cubes of size $p \times p \times q$, each analyzed by a dynamic texture model, and normal activity is summarized by a learned mean vector and covariance matrix. Anomaly detection then reduces to thresholding the likelihood of a new motion pattern under this learned distribution. Applying this recipe to the University of Minnesota crowd dataset yields 88.83% average accuracy across the video sequences, with the GROUND sequence used for performance analysis. The paper further claims that once the distribution is learned for a definite frame it can be prolonged to larger frame chunks via AR, MA, or ARMA process models, lowering the number of parameters and the learning variance.

Load-bearing premise

The 88.83% accuracy result rests on the assumption that the video frames used for learning and the video frames used for testing were properly separated, so the reported number measures prediction rather than memorization; the paper does not say how that split was made.

Editorial extensions

If this is right

  • If the reported accuracy holds, a dynamic-texture likelihood with a learned mean and covariance is sufficient to flag anomalous events in fixed surveillance scenes without tracking individuals.
  • The learned distribution for a single frame can be extended to larger frame chunks via AR, MA, or ARMA process models, which reduces the number of learned parameters and the variance of learning.
  • The method provides a traditional-machine-learning alternative to deep crowd analytics that does not depend on large training datasets.
  • The same spatio-temporal texture representation could support crowd density estimation and crowd event recognition in scenes whose normal motion is stable.

Reading between the lines

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

  • A natural reading the paper leaves open is that the 88.83% figure measures how well the learned Gaussian parameters fit videos from the same dataset, not how well they transfer to a different scene; cross-scene accuracy is therefore an unmeasured quantity.
  • The likelihood-threshold recipe could serve as a cheap, scene-specific baseline for crowd anomaly detection against which deep models are compared, because it needs no large annotated training corpus.
  • An online extension that re-estimates the mean and covariance as new frames arrive would test whether the model tracks gradual scene changes, a testable variant not reported in the paper.
  • Since the paper identifies scene-specificity as the main weakness of existing methods, the decisive next experiment is training on one crowd scene and testing on another; the paper reports no such transfer result.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The manuscript claims to investigate traditional and deep learning models for crowd analytics, centering on a proposed spatio-temporal Gaussian/dynamic-texture model for crowd anomaly detection. The only quantitative claim is in Section 3, where the author states that the proposed method achieved 88.83% average accuracy on the University of Minnesota crowd dataset. However, the manuscript as submitted contains no actual equations, figures, or tables; the method is described only verbally, and the experimental section reports a single aggregate accuracy without any specification of train/test separation, evaluation protocol, or comparison baselines. The central claim is therefore unverifiable from the submitted text.

Significance. If the reported accuracy were properly supported by a reproducible experimental protocol, the paper would provide a modest benchmark for a classical dynamic-texture/Gaussian modeling approach on the UMN crowd dataset. Such a result could be a useful data point for the crowd-analytics community, where deep learning methods dominate but classical baselines remain relevant. However, as submitted, the manuscript makes no verifiable technical contribution: there is no method specification, no empirical protocol, no analysis of hyperparameters, no error bars, and no comparison to existing methods. The significance of the claimed result cannot be assessed because the evidence is absent.

major comments (4)
  1. [Section 2, Proposed Method] The core method is never actually specified: the text contains phrases such as 'Eq. can be formulated as:' and 'as formulated in the eq.' followed by blank space, and the flow diagram and figures are missing. The description of spatio-temporal cubes of dimension p×p×q, the dynamic texture model, and the Gaussian mean/covariance estimation is purely verbal. Without the explicit equations, the reader cannot check the derivation, the number of free parameters, or the claimed relationship to the AR/MA/ARMA process model. This is load-bearing because the entire contribution rests on this model specification.
  2. [Section 3, Experimental Analysis and Evaluations] The paper reports 'On average our proposed method achieved 88.83% accuracy when applied on the videos from the same dataset' but never describes a division between videos used to estimate the Gaussian/dynamic-texture parameters and videos used for evaluation. Since Section 2 says the model learns a mean vector and covariance matrix for each motion pattern, and the evaluation uses videos from the same dataset, the 88.83% figure appears to reflect in-sample fit quality rather than generalization to unseen anomalies. No likelihood threshold selection procedure is described, so the result cannot be interpreted as anomaly-detection performance.
  3. [Section 3, Experimental Analysis and Evaluations] The manuscript refers to 'the Table shows the experimental analysis results' and 'Both graphs below show,' but no table or graphs are actually present in the submitted text. The reader cannot inspect per-sequence accuracies, their variance, or qualitative output frames. Consequently, even the descriptive statistic of 88.83% cannot be independently verified, and there is no way to assess the robustness of the method across the different UMN sequences.
  4. [Abstract and Introduction] The abstract and introduction promise to 'propose many models of deep neural networks and training approaches' and to investigate diverse scene crowd analytics with traditional and deep learning models. However, Sections 2 and 3 contain no deep learning model descriptions, no network architectures, no training procedures, and no deep-learning experiments. The scope of the actual contribution is therefore unclear, and the manuscript does not deliver what it announces.
minor comments (4)
  1. [Section 1, Introduction] There are numerous typographical errors, including 'lastdecade,' 'Therese kind of approaches,' and 'the odel' in Section 2. The manuscript would benefit from thorough proofreading.
  2. [References] Reference [36] is listed in the bibliography but never cited in the text; the citation sequence jumps from [35] to [37].
  3. [Section 1, Introduction] The sentence 'the distribution does not matter, it could be stationary over the learning interval or it could be mobile' is ambiguous. Please clarify whether 'mobile' means non-stationary and how the learning window or the AR/MA/ARMA process models handle non-stationarity.
  4. [Section 2, Proposed Method] The blank space after 'the flow diagram is presented as' indicates that the figure is missing. All figures and tables should be embedded in the manuscript.

Circularity Check

1 steps flagged · score 6.0 of 10

The sole quantitative accuracy claim is an in-sample fit: the learned Gaussian/dynamic-texture parameters are evaluated on 'videos from the same dataset' with no train/test split described.

  1. fitted input called prediction [Section 3, Experimental Analysis and Evaluations]
    "The results demonstrate that most of video sequences are accurately understood and learned by the algorithm . On average our proposed method achieved 88.83% accuracy when applied on the videos from the same dataset ."

    The paper's only quantitative result is the 88.83% accuracy figure. Section 2 states that the important parameters learned from the crowd video are the 'mean vector and covariance matrix specific to crowd scene under observation.' Section 3 then reports accuracy on 'the videos from the same dataset' — the same dataset from which those parameters were learned — and never describes a split into training and held-out test videos. Consequently, the reported accuracy measures how well the learned Gaussian/dynamic-texture model fits the very data it was estimated from, rather than how well it generalizes to unseen anomalous frames. The fitted parameters are effectively relabeled as a prediction.

full rationale

The paper's derivation chain is largely unverifiable because the equations are not actually displayed, and the only concrete evaluation claim is the reported 88.83% accuracy on the University of Minnesota crowd dataset. That claim is circular as written: the model learns a mean vector and covariance matrix from the crowd videos, and the evaluation is performed on 'videos from the same dataset' with no stated train/test separation. Without a defined split, the accuracy cannot be distinguished from in-sample fitting quality, so the central quantitative claim reduces to a fit rather than an independent prediction. No load-bearing self-citation chain is identifiable from the text: references [6] and [10] are cited for the likelihood-thresholding and dynamic-texture formulations, but there is no evidence that the present author is the author of those works, and the paper does not rely on a uniqueness theorem imported from the authors' prior work. The deeper limitation is that the manuscript is too incomplete to support the claimed evaluation; however, the specific circularity that is directly evidenced is the same-dataset evaluation of learned parameters, which warrants a score of 6 rather than a higher score for a fully self-citation-driven derivation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on an unstated Gaussian distributional assumption for local spatio-temporal patches, the choice of patch sizes p and q, and an evaluation protocol that trains and tests on the same dataset. No equations for the dynamic texture model or the mean/covariance update are actually given.

free parameters (5)
  • Spatial patch size p = not specified
    Spacio-temporal cube dimension p, no value or selection criterion is given in the text.
  • Temporal patch size q = not specified
    Spacio-temporal cube temporal dimension q, no value or selection criterion is given in the text.
  • Gaussian mean vector = fitted to training frames
    The model assumes a Gaussian distribution and estimates mean and covariance from training frames; these values are not reported.
  • Gaussian covariance matrix = fitted to training frames
    The covariance matrix is learned from data and is central to the likelihood-based anomaly detection, but its values are not reported.
  • Likelihood threshold = not specified
    Thresholding the likelihood is mentioned as theoretically encouraged, but no threshold value or calibration method is provided.
assumptions (4)
  • domain assumption General postulate that the pixel distribution may be stationary or mobile and can be learned over a window.
    Stated in the Introduction: 'we make the general postulate that the distribution does not matter, it could be stationary over the learning interval or it could be mobile.' This is unproven and load-bearing.
  • domain assumption Local patch motion is adequately represented by a linear dynamic system (dynamic texture model).
    The method follows reference [10] and assumes spatio-temporal cubes encode motion; no validation is provided.
  • domain assumption The crowd model distribution can be approximated as Gaussian.
    The paper states 'the model approximately consider Gaussian distribution' without derivation or empirical check.
  • domain assumption Decreasing training window reduces learning variance.
    Stated in the Introduction without proof: 'Decreasing or imposing conditions on the training session reduce the number of learned parameters... hence reducing the learning variance.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine and Deep Learning for Crowd Analytics." pith.science (2026). https://pith.science/paper/WWFSKP35

@misc{pith2026190904150,
  author       = {Pith},
  title        = {Pith review of: Machine and Deep Learning for Crowd Analytics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WWFSKP35}},
  note         = {Machine review of arXiv:1909.04150}
}
read the original abstract

In high population cities, the gatherings of large crowds in public places and public areas accelerate or jeopardize people safety and transportation, which is a key challenge to the researchers. Although much research has been carried out on crowd analytics, many of existing methods are problem-specific, i.e., methods learned from a specific scene cannot be properly adopted to other videos. Therefore, this presents weakness and the discovery of these researches, since additional training samples have to be found from diverse videos. This paper will investigate diverse scene crowd analytics with traditional and deep learning models. We will also consider pros and cons of these approaches. However, once general deep methods are investigated from large datasets, they can be consider to investigate different crowd videos and images. Therefore, it would be able to cope with the problem including to not limited to crowd density estimation, crowd people counting, and crowd event recognition. Deep learning models and approaches are required to have large datasets for training and testing. Many datasets are collected taking into account many different and various problems related to building crowd datasets, including manual annotations and increasing diversity of videos and images. In this paper, we will also propose many models of deep neural networks and training approaches to learn the feature modeling for crowd analytics.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

  1. [6]

    D., & Blumenstein, M

    Saqib, M., Khan, S. D., & Blumenstein, M. (2016, November). Texture -based feature mining for crowd density estimation: A study. In Image and Vision Computing New Zealand (IVCNZ), 2016International Conference on (pp. 1-6). IEEE

    work page 2016

  2. [10]

    Ullah, H.,Uzair, M.,Ullah, M., Khan, A., Ahmad, A., & Khan, W. (2017). Density independent hydrodynamics model for crowd coherency detection. Neurocomputing, 242, 28-39

    work page 2017

  3. [1]

    Towards a Crowd Analytic Framework For Crowd Management in Majid-al-Haram

    Khan, S. D., Tayyab, M., Amin, M. K., Nour, A., Basalamah, A., Basalamah, S., & Khan, S. A. (2017). Towards a Crowd Analytic Framework For Crowd Management in Majid -al-Haram. arXiv preprint arXiv:1709.05952

    work page Pith review arXiv 2017

  4. [2]

    Ahmad, K., Conci, N., & De Natale, F. G. (2018). A s aliency-based approach toevent recognition. Signal Processing: Image Communication, 60, 42-51

    work page 2018

  5. [3]

    B., Uzair, M., & Ullah, M

    Ullah, H., Altamimi, A. B., Uzair, M., & Ullah, M. (2018). Anomalous entities detection and localization in pedestrian flows. Neurocomputing, 290, 74-86

    work page 2018

  6. [4]

    D., Sharma, N., & Blumenstein, M

    Saqib, M., Khan, S. D., Sharma, N., & Blumenstein, M. (2017, December). Extracting descriptive motion information from crowd scenes. In 2017International Conference on Image and Vision Computing New Zealand (IVCNZ) (pp. 1-6). IEEE

    work page 2017

  7. [5]

    D., & Ullah, H

    Basalamah, S., Khan, S. D., & Ullah, H. (2019). Scale Driven Convolutional Neural Network Model For People Counting and Localization in Crowd Scenes. IEEE Access

    work page 2019

  8. [7]

    Ullah, H., Ullah, M., & Uzair, M. (2018). A hybrid social influence model for pedestrian motion segmentation. Neural Computing and Applications, 1-17

    work page 2018

Show all 37 references
  1. [8]

    (2018, September)

    Bisagno, N., Zhang, B., & Conci, N. (2018, September). Group LSTM: Group Trajectory Prediction in Crowded Scenarios. InEuropean Conference on Computer Vision(pp. 213 -225). Springer, Cham

    2018

  2. [9]

    U., Uzair, M., & Ullah, H

    Ahmad, F., Khan, A., Islam, I. U., Uzair, M., & Ullah, H. (2017). Illumination normalization using independent component analysis and filtering. The Imaging Science Journal, 65(5), 308-313

    2017

  3. [11]

    (2017, November)

    Trabelsi, R., Jabri, I., Melgani, F., Smach, F., Conci, N., & Bouallegue, A. (2017, November). Complex -Valued Representation for RGB -D Object Recognition. In Pacific -Rim Symposium on Image and Video Technology (pp. 17-27). Springer, Cham

    2017

  4. [12]

    Ullah, M., Ullah, H., & Alseadonn, I. M. (2017). Human action recognition i n videos using stable features

    2017

  5. [13]

    Xu, M., Ge, Z., Jiang, X., Cui, G., Zhou, B., & Xu, C. (2019). Depth Information Guided Crowd Counting for Complex Crowd Scenes. Pattern Recognition Letters

    2019

  6. [14]

    Alameda-Pineda, X., Ricci, E., & Sebe, N. (2019). Multimodal behavior analysis in the wild: An introduction. In Multimodal Behavior Analysis in the Wild (pp. 1-8). Academic Press

    2019

  7. [15]

    Ullah, M., Ullah, H., Conci, N., &De Natale, F. G. (2016, September). Crowd behavior identification. In Image Processing (ICIP), 2016IEEE International Conference on(pp. 1195-1199). IEEE

    2016

  8. [16]

    Kim, H., Han, J., & Han, S. (2019). Analysis of evacuation simulation considering crowd density and the effect of a fallen person. Journal of A mbient Intelligence and Humanized Computing, 1-11

    2019

  9. [17]

    J., Liu, Y., Wang, J., & Fan, J

    Hao, Y., Xu, Z. J., Liu, Y., Wang, J., & Fan, J. L. (2019). Effective crowd anomaly detection through spatio -temporal texture analysis. International Journal of Automatio n and Computing, 16(1), 27-39

    2019

  10. [18]

    Ullah, H., Ullah, M., Afridi, H., Conci, N., & De Natale, F. G. (2015, September). Traffic accident detection through a hydrodynamic lens. In Image Processing (ICIP), 2015IEEE International Conference on (pp. 2470-2474). IEEE

    2015

  11. [19]

    D., Bandini, S., & Nishinari, K

    Shimura, K., Khan, S. D., Bandini, S., & Nishinari, K. (2016). Simulation and Evaluation of Spiral Movement of Pedestrians: Towards the Tawaf Simulator. Journal of Cellular Automata, 11(4)

    2016

  12. [20]

    Ullah, H. (2015). Cro wd Motion Analysis: Segmentation, Anomaly Detection, and Behavior Classification (Doctoral dissertation, University of Trento)

    2015

  13. [21]

    Kang, D., Ma, Z., & Chan, A. B. (2018). Beyond counting: Comparisons of density maps for crowd analysis tasks -counting, detection, and tracking. IEEE Transactionson Circuits and Systems for Video Technology

    2018

  14. [22]

    Rota, P., Ullah, H., Conci, N., Sebe, N., & De Natale, F. G. (2013, September). Particles cross -influence forentity grouping. In Signal Processing Conference (EUSIPCO), 2013Proceedings of the 21st European (pp. 1-5). IEEE

    2013

  15. [23]

    (2014, March)

    Ullah, H., Ullah, M., & Conci, N. (2014, March). Real -time anomaly detection in dense crowded scenes. In Video Surveillance and Transportation Imaging Applications 2014(Vol. 9026, p. 902608). International Society for Optics and Photonics

    2014

  16. [24]

    Arif, M., Daud, S., & Basalamah, S. (2013). Counting of people in the extremely dense crow d using genetic algorithm and blobs counting. IAES International Journal of Artificial Intelligence, 2(2), 51

    2013

  17. [25]

    Ullah, M., & Alaya Cheikh, F. (2018). A Directed SparseGraphical Model for Multi - Target Tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (pp. 1816-1823)

    2018

  18. [26]

    D., & Ullah, H

    Khan, S. D., & Ullah, H. (2019). A survey of advances in vision -based vehicle re - identification. Computer Vision and Image Understanding, 182, 50-63

    2019

  19. [27]

    Ullah, M., Mohammed, A., & Alaya Cheikh, F. (2018). PedNet: A Spatio-Temporal Deep Convolutional Neural Network for Pedestrian Segmentation. Journal of Imaging, 4(9), 107

    2018

  20. [28]

    Mining motion atom s and phrases for complex action recognition

    Wang, LiMin, Yu Qiao, and Xiaoou Tang. "Mining motion atom s and phrases for complex action recognition." Proceedings of the IEEE international conference on computer vision. 2013

    2013

  21. [29]

    Dense trajectories and motion boundary descriptors for action recognition

    Wang, Heng, Alexander Kläser, Cordelia Schmid, and Cheng -Lin Liu. "Dense trajectories and motion boundary descriptors for action recognition." International journal of computer vision 103, no. 1(2013): 60-79

    2013

  22. [30]

    A comparative study of encoding, pooling and normalization methods for action recog nition

    Wang, Xingxing, LiMin Wang, and Yu Qiao. "A comparative study of encoding, pooling and normalization methods for action recog nition." In Asian Conference on Computer Vision, pp. 572-585. Springer, Berlin, Heidelberg, 2012

    2012

  23. [31]

    A tutorial on support vector regression

    Smola, Alex J., and Bernhard Schölkopf. "A tutorial on support vector regression." Statistics and computing 14, no. 3(2004): 199-222

    2004

  24. [32]

    In: IEEEconference on computer vision and pattern recognition (CVPR), pp 1–8

    Mahadevan V, Li W, Bhalodia V, Vasconcelos N (2010) Anomaly detection in crowded scenes. In: IEEEconference on computer vision and pattern recognition (CVPR), pp 1–8

    2010

  25. [33]

    A survey of advances in vision -based vehiclere - identification

    Khan, Sultan Daud, and Habib Ullah. "A survey of advances in vision -based vehiclere - identification." Computer Vision and Image Understanding (2019)

    2019

  26. [34]

    Internal Emotion Classification Using EEG Signal with S parse Discriminative Ensemble

    Ullah, Habib, Muhammad Uzair, Arif Mahmood, Mohib Ullah, Sultan Daud Khan, and Faouzi Alaya Cheikh. "Internal Emotion Classification Using EEG Signal with S parse Discriminative Ensemble." IEEE Access (2019)

    2019

  27. [35]

    D., Sharma, N., & Blumenstein, M

    Saqib, M., Khan, S. D., Sharma, N., & Blumenstein, M. (2017, December). Extracting descriptive motion information from crowd scenes. In 2017 International Conference on Image and Vision Computing New Zealand (IVCNZ) (pp. 1-6). IEEE

    2017

  28. [36]

    & Amandi, R

    Coluccia, A., Ghenescu, M., Piatrik, T., De Cubber, G., Schumann, A., Sommer, L., ... & Amandi, R. (2017, August). Drone-vs-bird detection challenge at IEEE AVSS2017. In 2017 14th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS) (pp. 1-6). IEEE

    2017

  29. [37]

    Ullah, M., Ullah, H., & Cheikh, F. A. (2019). SINGLE SHOT APPEARANCE MODEL (SSAM) FOR MULTI-TARGET TRACKING. Electronic Imaging, 2019(7), 466-1

    2019

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.