Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Wrapping specialized detectors in multi-frame temporal validation turns noisy AI into reliable real-time multi-task surveillance.

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

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

T0 review · grok-4.5

2026-07-12 04:40 UTC pith:5L2FD2MF

load-bearing objection Solid open multi-task surveillance systems paper: temporal validation gives real, measured FA/accuracy gains on video, two fine-tunes and a vandalism corpus are useful, but the headline deltas rest on hand-tuned thresholds and small video n. the 4 major comments →

arxiv 2607.03131 v1 pith:5L2FD2MF submitted 2026-07-03 cs.CV cs.AI

A Multi-Task Deep Learning Framework for Real-Time Intelligent Video Surveillance with Temporal Event Validation

Estera Dumitru , Stelian Sp\^inu This is my paper
classification cs.CV cs.AI
keywords intelligent video surveillancedeep learningtemporal event validationmulti-task real-time systemsweapon detectionhuman action recognitionautomatic license plate recognitionfire and smoke detection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Modern cameras produce far more video than people can watch. This paper claims the practical fix is not a new detector architecture but a shared-GPU multi-task system that runs face recognition with zone authorization, license-plate reading, weapon detection, fire/smoke detection, and action recognition concurrently, then never treats a single-frame output as an event. Instead every module is wrapped in a temporal validation layer that demands multi-frame confirmation, confidence-weighted voting, and cascaded filters before an alarm is raised. Two custom models fill public-data gaps: a single-class weapon detector (mAP@0.5 0.947) and a SlowFast action model trained on a new 614-clip vandalism set (94.33% accuracy). On continuous video the layer cuts fire/smoke false alarms from 52% to 4% and raises plate exact-match accuracy from 66.7% to 81.8%, while end-to-end latency stays under 100 ms on ordinary hardware. The authors argue this combination makes open, auditable surveillance competitive with closed commercial products.

Core claim

A unified multi-task pipeline that shares one GPU across five heterogeneous vision tasks becomes operationally reliable only when every detector is wrapped in a temporal event-validation layer (multi-frame confirmation, confidence-weighted voting, cascaded filtering, open-set rules). That layer, not raw model accuracy, is what converts frame-level noise into trustworthy security events while preserving real-time speed.

What carries the argument

Temporal event-validation architecture: each detector is post-processed by multi-frame IoU confirmation, confidence-weighted temporal voting, and cascaded size/color/cooldown filters so that only persistent, consistent predictions become alarms.

Load-bearing premise

The hand-chosen confirmation lengths, confidence cuts, cooldowns and format rules that produce the headline gains will transfer from the small evaluation video sets to continuous multi-camera deployment without large sensitivity loss or retuning.

What would settle it

Deploy the identical cascade thresholds and confirmation lengths on a new multi-day multi-camera stream whose fire, plate and weapon ground truth are independently labeled; if false-alarm or exact-match rates deviate sharply from the reported 4% and 81.8% without retuning, the transfer claim fails.

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

If this is right

  • False-alarm rates of commodity fire and smoke detectors can be driven into single digits on video without sacrificing near-perfect positive-clip detection.
  • License-plate accuracy on continuous streams rises substantially once single-frame OCR is replaced by confidence-weighted multi-frame voting.
  • A single commodity GPU can host five concurrent surveillance tasks at camera rate when display uses a last-available-frame fusion policy.
  • Open modular systems can match the operational reliability of closed commercial platforms while remaining fully inspectable and free of recurring licenses.
  • Purpose-built data for under-represented classes such as vandalism enables action models that public benchmarks alone do not support.

Where Pith is reading between the lines

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

  • The same temporal-validation wrapper could be applied to any other single-frame security detector (loitering, abandoned object, fall) with comparable false-alarm reduction.
  • Because the architecture already isolates each detector behind queues, adding a sixth task requires only another thread and a validation rule, not a redesign.
  • The reported hardware headroom under six-stream stress testing implies that modest model compression or multi-GPU distribution would scale the system to dozens of cameras without changing the validation logic.
  • If confirmation lengths must be retuned per site, the open modular design makes that retuning far cheaper than closed commercial alternatives.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a unified multi-task real-time video surveillance framework that runs five concurrent vision tasks (face recognition with zone authorization, ALPR, weapon detection, fire/smoke detection, and action recognition) on a shared GPU, with a temporal event-validation layer (multi-frame confirmation, confidence-weighted voting, cascaded filtering) converting frame-level detections into operational alarms. Two custom models are introduced: a single-class YOLOv8m weapon detector fine-tuned on a merged/relabeled dataset (mAP@0.5 = 0.947) and a SlowFast-R50 action model trained with a new 614-clip vandalism corpus (94.33% validation accuracy). Module-level results on LFW, D-Fire, FIRESENSE, and a UCF-Crime subset, plus end-to-end latency under concurrent load (<100 ms/frame on commodity hardware), are used to argue that specialized detectors plus temporal validation yield reliable real-time surveillance, with headline gains of fire/smoke false-alarm rate 52%→4% and video plate exact-match 66.7%→81.8%.

Significance. If the temporal-validation gains transfer beyond the reported clips, the work is a useful systems contribution: an open, modular alternative to closed commercial VMS stacks, with concrete operational metrics (false-alarm reduction, concurrent multi-task latency) that pure detector papers rarely report. Strengths that should be credited include public code, held-out weapon test metrics that match validation (mAP@0.5 0.943 vs 0.947), open-set face evaluation with 0% FAR on large impostor sets, and paired raw-vs-pipeline tables that isolate the validation layer. The purpose-built vandalism set and single-class weapon training fill genuine coverage gaps. Novelty is primarily architectural and operational rather than algorithmic; significance therefore hinges on whether the reliability claims are robustly supported, not on new network designs.

major comments (4)
  1. [§III-B–G, Tables IV and VI] §III-B–G and Tables IV/VI: The central claim that temporal event validation produces the headline reliability gains (fire/smoke FA 52%→4% on FIRESENSE; plate exact-match 66.7%→81.8% on 33 clips) rests on a stack of hand-chosen free parameters (fire conf. 0.40 / smoke 0.55, size band 0.3–85%, HSV gate, multi-frame IoU persistence, cooldowns; weapon secondary conf. 0.7; plate ≥3 readings + Romanian regex + conf-weighted vote; face L2<1.0 and conf≥0.4). No threshold sweep, confirmation-length ablation, or leave-one-camera/site hold-out is reported. With n=49 FIRESENSE clips and n=33 plate clips, the magnitude of the claimed deltas is not yet shown to transfer to continuous multi-camera deployment without retuning. A sensitivity analysis of confirmation length and cascade cutoffs, plus at least one held-out site or longer continuous stream, is needed to support the strongest claim.
  2. [§IV-C, Tables V–VI] Table V vs Table VI: On static D-Fire the same cascade already cuts detection from 89.7% to 53.6% (FA 19.4%→8.9%), so the operating point is known to be aggressive. The video result (23/24 detection, 1/25 FA) is encouraging but does not characterize delayed detections, missed short events, or the sensitivity–FA tradeoff as a function of confirmation length. Without that curve (or an ablation removing individual cascade stages), it is hard to judge whether the 4% FA is purchased at an operationally acceptable sensitivity cost under varied scene conditions.
  3. [§III-D, §IV-E, Tables VIII–IX] §III-D and §IV-E: Action recognition reports 94.33% on the authors’ validation split of a purpose-built three-class set (RWF-2000 normal/fight + 614 vandalism clips). UCF-Crime generalization (150 clips, 86.7% with “at least one” window aggregation) is a useful check but remains small and uses an aggregation rule that can inflate incident-level recall. The vandalism corpus construction (multilingual scraping, manual validation) is only briefly described; for a claimed contribution, clearer documentation of clip criteria, inter-annotator checks, and train/val leakage controls (same source video across splits) is needed so others can assess bias and reproduce the class.
  4. [§IV-F, Tables X–XII] §IV-F: Integrated evaluation is limited to one sequential functional clip and one six-quadrant stress clip. That demonstrates concurrent scheduling and latency headroom, but does not stress multi-camera RTSP variability, long-running false-alarm accumulation, or concurrent alarm deduplication under realistic event rates. A longer multi-stream run with logged alarm precision/recall over time would better support the “reliable operational system” claim.
minor comments (6)
  1. [Abstract, §I-C, Table VII] Abstract and §I-C state weapon mAP@0.5 = 0.947; Table VII test is 0.943. Align abstract/contribution bullets with the held-out test figure, or clearly label 0.947 as validation.
  2. [§III-B, Fig. 1] Fig. 1 is described in text but the cascade and last-available-frame policy would benefit from a short pseudocode box or algorithm listing for the temporal validation layer, to make the contribution implementable without reading the full module sections.
  3. [§III-F, §IV-B] Plate evaluation is Romanian-format-specific (regex + allowlist). State this limitation more prominently when claiming ALPR gains, and note expected transfer cost to other plate formats.
  4. [§II, References] Related work cites several 2025–2026 items; ensure final versions and DOIs are stable at camera-ready, and that commercial product URLs are archived or dated consistently.
  5. [Table XII] Table XII note on plate latency (median dominated by no-plate frames) is important; consider reporting conditional latency when EasyOCR fires as a primary metric for that module.
  6. [§I-A] Minor wording: “closedandproprietary” and similar missing spaces appear in §I-A; clean copy-editing pass needed.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper whose headline deltas are measured before/after comparisons on held-out public and custom splits, not forced by definition or self-citation.

full rationale

The paper is an engineering systems contribution that trains two custom detectors (YOLOv8m weapon on a merged/relabeled public corpus; SlowFast-R50 action on RWF-2000 plus a purpose-built 614-clip vandalism set), wraps off-the-shelf and custom models in hand-chosen temporal filters (multi-frame IoU confirmation, confidence-weighted voting, five-level cascade, double-threshold open-set rule), and reports measured performance on independent public benchmarks (LFW, D-Fire, FIRESENSE, UCF-Crime) plus held-out test partitions. The key numerical claims (fire/smoke FA 52%→4% on FIRESENSE Table VI; plate exact-match 66.7%→81.8% on 33 clips Table IV; weapon mAP@0.5 0.947; action 94.33% validation accuracy) are obtained by direct evaluation of the full pipeline versus the raw detector, not by algebraic identity with fitted inputs. Thresholds and confirmation lengths are free operational parameters introduced in §III-B–G; they are not derived from the target metrics, nor is any uniqueness theorem or ansatz imported via self-citation. Training curves, validation/test agreement for the weapon model, and external-set results for action recognition further confirm that reported numbers are not self-forced. No equation equates a claimed prediction to its own inputs; the work is therefore self-contained against external benchmarks and exhibits zero circularity under the stated criteria.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

Load-bearing content is engineering choices and hand-set operating thresholds around pretrained detectors, not new physical entities. Free parameters are the cascade and decision thresholds that produce the headline false-alarm and accuracy deltas. Axioms are standard CV/surveillance domain assumptions and reuse of public pretrained models. No new particles or mediators are invented; the 'temporal event-validation architecture' is a design pattern over known filters.

free parameters (5)
  • Face open-set L2 and confidence thresholds = L2=1.0, conf≥0.4
    Identity accepted only if L2 < 1.0 and cosine confidence ≥ 0.4; chosen via threshold sweep on LFW and used for 0% FAR claim.
  • Fire/smoke cascade thresholds = 0.40 / 0.55 conf; size 0.3–85%
    Per-class confidence (0.40 fire, 0.55 smoke), size band 0.3–85% of frame, HSV heuristic, multi-frame IoU persistence, per-location cooldown drive FA drop 52%→4%.
  • Weapon secondary confidence and multi-frame confirm = conf≥0.7 + multi-frame
    Inference keeps detections above secondary conf 0.7 plus multi-frame confirmation before alert.
  • Plate track voting rules = ≥3 readings, conf≥0.4, RO regex
    ≥3 format-valid readings, conf≥0.4, Romanian regex, confidence-weighted majority vote; produce 66.7%→81.8% exact-match gain.
  • Action alert confidence and cooldown = conf≥0.5, 5s cooldown
    Non-normal action alerted only above softmax conf 0.5 with 5 s cooldown; inference every 30 frames from 2.0 s window.
axioms (5)
  • domain assumption Pretrained InsightFace buffalo_s ArcFace embeddings, YOLO family detectors, EasyOCR, and Kinetics-pretrained SlowFast-R50 are adequate bases for surveillance fine-tuning and deployment.
    §II–III select these models as mature real-time backbones; all module results rest on that reuse.
  • domain assumption Multi-frame persistence and cascaded filters reduce operational false alarms without unacceptable true-positive loss on continuous video.
    Central design principle of §III-B; quantified on FIRESENSE and plate clips in §IV.
  • ad hoc to paper Single generic 'weapon' class is the right operational target (presence over subtype).
    §III-C merges heterogeneous labels into one class for binary presence alerts.
  • ad hoc to paper Author-curated vandalism clips (614) plus RWF-2000 normal/fight splits represent the three action classes of interest.
    §III-D dataset construction; validation and UCF-Crime transfer metrics depend on this labeling.
  • domain assumption Last-available-frame fusion on one shared GPU is an acceptable correctness/latency trade-off for multi-task display.
    §III-A architecture rationale; end-to-end claims use this policy.

pith-pipeline@v1.1.0-grok45 · 22897 in / 3712 out tokens · 40333 ms · 2026-07-12T04:40:02.694558+00:00 · methodology

0 comments
read the original abstract

Modern video surveillance systems generate far more video streams than human operators can effectively monitor, making automated analysis essential for timely detection of security events. This paper presents a unified multi-task deep learning framework that simultaneously performs face recognition with zone-based authorization, automatic license plate recognition, weapon detection, fire and smoke detection, and human action recognition on a shared GPU platform. Among the integrated modules, two task-specific deep-learning models are proposed in this work to address scenarios that are insufficiently represented in publicly available datasets: a single-class weapon detector fine-tuned on a merged and relabeled dataset, achieving a mean average precision (mAP@0.5) of 0.947, and a SlowFast-R50 action recognition model trained on a purpose-built vandalism dataset comprising 614 video clips, achieving 94.33% classification accuracy. To improve robustness in continuous video, all detection modules are integrated into a temporal event-validation architecture based on multi-frame confirmation, confidence-weighted voting, and cascaded filtering, transforming frame-level predictions into reliable security events. Each module is evaluated independently on established public datasets (LFW, D-Fire, FIRESENSE, and UCF-Crime), followed by integrated end-to-end system evaluation. The proposed temporal validation strategy reduces the fire and smoke false-alarm rate from 52% to 4% and improves video license plate exact-match accuracy from 66.7% to 81.8%, while the complete framework maintains real-time operation with a per-frame latency below 100 ms on commodity hardware. These results demonstrate that combining specialized deep-learning models with temporal event validation provides an effective and practical solution for reliable real-time intelligent video surveillance.

Figures

Figures reproduced from arXiv: 2607.03131 by Estera Dumitru, Stelian Sp\^inu.

Figure 1
Figure 1. Figure 1: Overall architecture of the proposed framework. Each frame from the video streams is dispatched, in parallel, to five detectors that run concurrently [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training curves of the weapon detector over 100 epochs. Top [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Normalized confusion matrix of the action recognizer on the [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training evolution of the action recognizer over 28 epochs. Left: [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. ObjectStream: Latent Objects as Memory Anchors for Streaming Video Understanding

    cs.CV 2026-07 conditional novelty 6.0

    Training-free latent-object memory anchors let frozen Video-LLMs retain object histories under a tight token budget and improve streaming and long-video QA.

  2. ObjectStream: Latent Objects as Memory Anchors for Streaming Video Understanding

    cs.CV 2026-07 conditional novelty 6.0

    A training-free memory framework that anchors streaming video memory to latent objects discovered from frozen Video-LLM features, improving streaming QA accuracy while cutting memory and latency.

Reference graph

Works this paper leans on

51 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    Surveillance camera statistics: which city has the most CCTV cameras?

    P. Bischoff, “Surveillance camera statistics: which city has the most CCTV cameras?” Comparitech, 2025. [Online]. Available: https://www.comparitech.com/vpn-privacy/the-worlds-most-surveilled- cities/

    2025

  2. [2]

    Intelligent video surveillance: a review through deep learning techniques for crowd analysis,

    G. Sreenu and M. A. Saleem Durai, “Intelligent video surveillance: a review through deep learning techniques for crowd analysis,” J. Big Data, vol. 6, no. 48, pp. 1–27, 2019

    2019

  3. [3]

    How effective is human video surveillance performance?

    N. Sulman, T. Sanocki, D. Goldgof, and R. Kasturi, “How effective is human video surveillance performance?” in Proc. 19th Int. Conf. Pattern Recognit. (ICPR), 2008, pp. 1–3

    2008

  4. [4]

    Pirate stealth or inattentional blindness? The effects of target relevance and sustained attention on security monitoring for experienced and na ¨ıve operators,

    E. N ¨asholm, S. Rohlfing, and J. D. Sauer, “Pirate stealth or inattentional blindness? The effects of target relevance and sustained attention on security monitoring for experienced and na ¨ıve operators,” PLoS ONE, vol. 9, no. 1, p. e86157, 2014

    2014

  5. [5]

    How the machine ‘thinks’: understanding opacity in machine learning algorithms,

    J. Burrell, “How the machine ‘thinks’: understanding opacity in machine learning algorithms,” Big Data & Society, vol. 3, no. 1, pp. 1–12, 2016

    2016

  6. [6]

    The value of open source software,

    M. Hoffmann, F. Nagle, and Y . Zhou, “The value of open source software,” Harvard Business School, Working Paper 24-038, 2024

    2024

  7. [7]

    Privacy and security vulnerabilities in edge intelligence: an analysis and countermeasures,

    A. Shafee, S. R. Hasan, and T. A. Awaad, “Privacy and security vulnerabilities in edge intelligence: an analysis and countermeasures,” Comput. Electr. Eng., vol. 123, p. 110146, 2025. doi: https://doi.org/10. 1016/j.compeleceng.2024.110146

    arXiv 2025

  8. [8]

    Ultralytics YOLOv8,

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics YOLOv8,” 2023. [Online]. Available: https://github.com/ultralytics/ultralytics

    2023

  9. [9]

    Labeled Faces in the Wild: a database for studying face recognition in uncon- strained environments,

    G. B. Huang, M. Ramesh, T. Berg, and E. Learned-Miller, “Labeled Faces in the Wild: a database for studying face recognition in uncon- strained environments,” Univ. Massachusetts, Amherst, Tech. Rep. 07- 49, 2007

    2007

  10. [10]

    An automatic fire detection system based on deep convolutional neural networks for low-power, resource-constrained devices,

    P. V . A. B. de Venˆancio, A. C. Lisboa, and A. V . Barbosa, “An automatic fire detection system based on deep convolutional neural networks for low-power, resource-constrained devices,” Neural Comput. Appl., vol. 34, pp. 15349–15368, 2022

    2022

  11. [11]

    FIRESENSE database of videos for flame and smoke detection,

    N. Grammalidis, K. Dimitropoulos, and E. Cetin, “FIRESENSE database of videos for flame and smoke detection,” Zenodo, 2017. [Online]. Available: https://doi.org/10.5281/zenodo.836749

    work page doi:10.5281/zenodo.836749 2017

  12. [12]

    Real-world anomaly detection in surveillance videos,

    W. Sultani, C. Chen, and M. Shah, “Real-world anomaly detection in surveillance videos,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 6479–6488

    2018

  13. [13]

    The security of IP- based video surveillance systems,

    N. Kalbo, Y . Mirsky, A. Shabtai, and Y . Elovici, “The security of IP- based video surveillance systems,” Sensors, vol. 20, no. 17, p. 4806, 2020

    2020

  14. [14]

    Video surveillance systems—current status and future trends,

    V . Tsakanikas and T. Dagiuklas, “Video surveillance systems—current status and future trends,” Comput. Electr. Eng., vol. 70, pp. 736–753, 2018

    2018

  15. [15]

    Abandoned object detection in video-surveillance: survey and comparison,

    E. Luna, J. C. San Miguel, D. Ortego, and J. M. Mart ´ınez, “Abandoned object detection in video-surveillance: survey and comparison,” Sensors, vol. 18, no. 12, p. 4290, 2018

    2018

  16. [16]

    Analytics and detection—Verkada camera documenta- tion,

    Verkada Inc., “Analytics and detection—Verkada camera documenta- tion,” 2026. [Online]. Available: https://help.verkada.com/

    2026

  17. [17]

    How secure is your video surveillance system? Cyberse- curity guide,

    Verkada Inc., “How secure is your video surveillance system? Cyberse- curity guide,” 2021. [Online]. Available: https://docs.verkada.com/

    2021

  18. [18]

    Avigilon Appearance Search technology,

    Motorola Solutions, “Avigilon Appearance Search technology,” Avigilon,

  19. [19]

    Available: https://docs.avigilon.com/

    [Online]. Available: https://docs.avigilon.com/

  20. [20]

    BriefCam platform—datasheet,

    BriefCam, “BriefCam platform—datasheet,” 2025. [Online]. Available: https://www.briefcam.com/

    2025

  21. [21]

    BriefCam appliance catalog,

    BCDVideo/BriefCam, “BriefCam appliance catalog,” 2018. [Online]. Available: https://www.bcdvideo.com/

    2018

  22. [22]

    Frigate: open source AI object detection for IP cameras,

    Frigate NVR, “Frigate: open source AI object detection for IP cameras,”

  23. [23]

    Available: https://frigate.video/

    [Online]. Available: https://frigate.video/

  24. [24]

    Microsoft COCO: common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll ´ar, and C. L. Zitnick, “Microsoft COCO: common objects in context,” in Proc. Eur. Conf. Comput. Vis. (ECCV), 2014, pp. 740–755

    2014

  25. [25]

    ZoneMinder documentation—user guide

    ZoneMinder, “ZoneMinder documentation—user guide.” [Online]. Available: https://zoneminder.readthedocs.io/

  26. [26]

    Comparative analysis of previous YOLO detectors and YOLOv26s for real-time weapon detection in video surveillance,

    C. J. Fierro Silva, C. Del-Valle-Soto, C. Bran, and J. Varela-Ald ´as, “Comparative analysis of previous YOLO detectors and YOLOv26s for real-time weapon detection in video surveillance,” Front. Comput. Sci., vol. 8, 2026. doi: https://doi.org/10.3389/fcomp.2026.1789702

    work page doi:10.3389/fcomp.2026.1789702 2026

  27. [27]

    Real-time threat detection from surveillance cameras using machine learning,

    G. Mandal, J. P. Patra, and P. Mahant, “Real-time threat detection from surveillance cameras using machine learning,” 2026, arXiv:2606.05708. [Online]. Available: https://arxiv.org/abs/2606.05708

    Pith/arXiv arXiv 2026

  28. [28]

    Audiovisual gun detection with automated lockdown and PA announcing IoT system for schools,

    T. Khan, “Audiovisual gun detection with automated lockdown and PA announcing IoT system for schools,” IoT, vol. 7, no. 1, p. 15, 2026. doi: https://doi.org/10.3390/iot7010015

    work page doi:10.3390/iot7010015 2026

  29. [29]

    ArcFace: additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “ArcFace: additive angular margin loss for deep face recognition,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 4690–4699

    2019

  30. [30]

    A review of machine learning and deep learning methods for person detection, tracking and identification, and face recognition with applications,

    B. Amirgaliyev, M. Mussabek, T. Rakhimzhanova, and A. Zhumadil- layeva, “A review of machine learning and deep learning methods for person detection, tracking and identification, and face recognition with applications,” Sensors, vol. 25, no. 5, p. 1410, 2025. doi: https: //doi.org/10.3390/s25051410

    work page doi:10.3390/s25051410 2025

  31. [31]

    A., Gomez-Rios, M

    Ronquillo-Figueroa, C., Quiroz-Martinez, M. A., Gomez-Rios, M. D., & Leyva-Vazquez, M. (2026). Comparative Evaluation of ArcFace and FaceNet Models for Real-World Facial Recognition Using the Analytic Hierarchy Process. ˆIn M. Botto Tobar et al. (eds), Trends in Artificial Intelligence and Computer Engineering. ICAETT 2025 . Lecture Notes in Networks and ...

    2026

  32. [32]

    Partial FC: training 10 million identities on a single machine,

    X. An et al., “Partial FC: training 10 million identities on a single machine,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. Workshops (ICCVW), 2021, pp. 1445–1449

    2021

  33. [33]

    Billion-scale similarity search with GPUs,

    J. Johnson, M. Douze, and H. J ´egou, “Billion-scale similarity search with GPUs,” IEEE Trans. Big Data, vol. 7, no. 3, pp. 535–547, 2021

    2021

  34. [34]

    A comprehensive review of YOLO architectures in computer vision: from YOLOv1 to YOLOv8 and YOLO-NAS,

    J. Terven, D.-M. C ´ordova-Esparza, and J.-A. Romero-Gonz ´alez, “A comprehensive review of YOLO architectures in computer vision: from YOLOv1 to YOLOv8 and YOLO-NAS,” Mach. Learn. Knowl. Extr., vol. 5, no. 4, pp. 1680–1716, 2023. doi: https://doi.org/10.3390/ make5040083

    2023

  35. [35]

    YOLOv10 to its genesis: a decadal and comprehensive review of the you only look once series,

    R. Sapkota et al., “YOLOv10 to its genesis: a decadal and comprehensive review of the you only look once series,” 2024, arXiv:2406.19407. [Online]. Available: https://arxiv.org/abs/2406.19407

    Pith/arXiv arXiv 2024

  36. [36]

    YOLOv10: real-time end-to-end object detection,

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, and G. Ding, “YOLOv10: real-time end-to-end object detection,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2024, arXiv:2405.14458. [Online]. Avail- able: https://arxiv.org/abs/2405.14458

    Pith/arXiv arXiv 2024

  37. [37]

    SlowFast networks for video recognition,

    C. Feichtenhofer, H. Fan, J. Malik, and K. He, “SlowFast networks for video recognition,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 6202–6211

    2019

  38. [38]

    RWF-2000: an open large scale video database for violence detection,

    M. Cheng, K. Cai, and M. Li, “RWF-2000: an open large scale video database for violence detection,” in Proc. 25th Int. Conf. Pattern Recognit. (ICPR), 2021, pp. 4183–4190

    2000

  39. [39]

    Literature review of deep-learning-based detec- tion of violence in video,

    P. Negre, R. S. Alonso, A. Gonz ´alez-Briones, J. Prieto, and S. Rodr´ıguez-Gonz´alez, “Literature review of deep-learning-based detec- tion of violence in video,” Sensors, vol. 24, no. 12, p. 4016, 2024. doi: https://doi.org/10.3390/s24124016

    work page doi:10.3390/s24124016 2024

  40. [40]

    The Kinetics human action video dataset,

    W. Kay et al., “The Kinetics human action video dataset,” 2017, arXiv:1705.06950

    Pith/arXiv arXiv 2017

  41. [41]

    Egyptian car plate recognition based on YOLOv8, Easy-OCR, and CNN,

    A. Sarhan et al., “Egyptian car plate recognition based on YOLOv8, Easy-OCR, and CNN,” J. Electr. Syst. Inf. Technol., vol. 11, no. 1, p. 32, 2024. doi: https://doi.org/10.1186/s43067-024-00156-y

    work page doi:10.1186/s43067-024-00156-y 2024

  42. [42]

    A deep learning-based system for automatic license plate recognition using YOLOv12 and PaddleOCR,

    B. Buleu, R. Robu, and I. Filip, “A deep learning-based system for automatic license plate recognition using YOLOv12 and PaddleOCR,” Appl. Sci., vol. 15, no. 14, p. 7833, 2025. doi: https://doi.org/10.3390/ app15147833

    2025

  43. [43]

    EasyOCR: ready-to-use optical character recognition

    JaidedAI, “EasyOCR: ready-to-use optical character recognition.” [On- line]. Available: https://github.com/JaidedAI/EasyOCR

  44. [44]

    Character region awareness for text detection,

    Y . Baek, B. Lee, D. Han, S. Yun, and H. Lee, “Character region awareness for text detection,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 9365–9374

    2019

  45. [45]

    An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition,

    B. Shi, X. Bai, and C. Yao, “An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, no. 11, pp. 2298–2304, 2017

    2017

  46. [46]

    Real-time convolu- tional neural networks for emotion and gender classification,

    O. Arriaga, M. Valdenegro-Toro, and P. Pl ¨oger, “Real-time convolu- tional neural networks for emotion and gender classification,” 2017, arXiv:1710.07557

    Pith/arXiv arXiv 2017

  47. [47]

    Dangerous Items Dataset for 5-Class Object Detection (YOLO annotation),

    Z. Omiotek, “Dangerous Items Dataset for 5-Class Object Detection (YOLO annotation),” Zenodo, 2025. [Online]. Available: https://doi.org/ 10.5281/zenodo.16422779

    work page doi:10.5281/zenodo.16422779 2025

  48. [48]

    Object detection binary classifiers methodology based on deep learning to identify small objects handled similarly: application in video surveillance,

    F. P ´erez-Hern´andez, S. Tabik, A. Lamas, R. Olmos, H. Fujita, and F. Herrera, “Object detection binary classifiers methodology based on deep learning to identify small objects handled similarly: application in video surveillance,” Knowl.-Based Syst., vol. 194, p. 105590, 2020. doi: https: //doi.org/10.1016/j.knosys.2020.105590

    work page doi:10.1016/j.knosys.2020.105590 2020

  49. [49]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2019

    2019

  50. [50]

    MobileFaceNets: efficient CNNs for accurate real-time face verification on mobile devices,

    S. Chen, Y . Liu, X. Gao, and Z. Han, “MobileFaceNets: efficient CNNs for accurate real-time face verification on mobile devices,” in Proc. Chin. Conf. Biometric Recognit. (CCBR), 2018, pp. 428–438

    2018

  51. [51]

    PyTorch: an imperative style, high-performance deep learning library,

    A. Paszke et al., “PyTorch: an imperative style, high-performance deep learning library,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2019, pp. 8024–8035

    2019