pith. sign in

arxiv: 2606.21568 · v1 · pith:RQHJSTAFnew · submitted 2026-06-19 · 💻 cs.CV

A Smart Classroom Behavior Analysis Framework with a New Highly Congested Classroom Dataset

Wei Xu , Maoxiang Chu , Yuelong Fan , Guanghao Liao , Yinxiang Yu , Zhi Chen , Haotian Wang , Yutian Zhu This is my paper

Pith reviewed 2026-06-26 14:58 UTC · model grok-4.3

classification 💻 cs.CV
keywords classroom behavior detectioncrowded scene detectionocclusion handlingYOLO modificationnew datasetstudent action recognitiondense object detection
0
0 comments X

The pith

ODER-HSFNet with three custom modules outperforms standard YOLO detectors on crowded classroom behavior tasks.

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

The paper builds the Highly Congested Classroom Behavior dataset with over fifty thousand labeled student instances to expose the limits of current detectors in packed rooms. It introduces ODER-HSFNet, a YOLO variant that adds an occlusion-aware edge rectifier, a hypergraph-state fusion module, and an occlusion-calibrated detection head. These additions are presented as direct responses to dense overlaps, partial views, distance-based size changes, and weak semantics in distant students. A sympathetic reader would expect the new framework to produce fewer missed or confused behaviors than unmodified YOLO models when students sit shoulder-to-shoulder. The reported gains on both the new dataset and an existing one supply the concrete evidence offered for that expectation.

Core claim

The authors claim that the ODER-HSFNet framework, built around the Occlusion-aware Deformable Edge Rectifier for boundary strengthening, the Hypergraph-State Spatial Fusion module for local-to-global context integration, and the Occlusion-Calibrated Detection Head for pruning weak candidates, delivers higher mean average precision than mainstream YOLO detectors when locating seven categories of student behavior under the conditions captured in the HCCB dataset.

What carries the argument

ODER-HSFNet, the YOLO-based detector whose three modules (ODER for deformable edge correction under occlusion, HSSF for hypergraph and state-space fusion, and OCDetect for pre-NMS candidate filtering) target the specific failure modes of dense classroom scenes.

If this is right

  • Boundary evidence remains usable even when neighboring students heavily overlap.
  • High-order spatial relations among instances can be aggregated without separate post-processing steps.
  • False positives triggered by occlusion edges or adjacent students drop after candidate calibration.
  • The same three modules transfer to the SCB-D3-S classroom dataset and still improve over unmodified YOLO baselines.
  • Ablation results indicate that removing any one module measurably lowers performance on the congested benchmark.

Where Pith is reading between the lines

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

  • The module pattern could be tested on other dense detection problems such as counting attendees at events or monitoring livestock in pens.
  • The HCCB construction guidelines might be reused to create comparable benchmarks for fine-grained action categories in other crowded indoor settings.
  • Video extensions of the same head could track behavior sequences across frames once single-frame detection is stabilized.

Load-bearing premise

The four listed scene challenges are the primary reasons standard detectors fail, and the three added modules correct them without introducing new failure modes or overfitting to the HCCB data collection process.

What would settle it

A baseline YOLO model trained from scratch on the HCCB training split that reaches or exceeds the reported 60.60 percent mAP50:95 and 80.12 percent mAP50 would show that the custom modules are not necessary for the claimed improvement.

Figures

Figures reproduced from arXiv: 2606.21568 by Guanghao Liao, Haotian Wang, Maoxiang Chu, Wei Xu, Yinxiang Yu, Yuelong Fan, Yutian Zhu, Zhi Chen.

Figure 1
Figure 1. Figure 1: HCCB dataset construction and model-assisted annotation refinement pipeline. The yellow modules represent the manual annotation stage, the blue modules represent the data preprocessing stage, and the green modules represent the model-assisted annotation optimization stage. 3.2. Annotation Protocol and Dataset Split In highly congested classroom scenarios, the difficulty of student behavior detection comes … view at source ↗
Figure 2
Figure 2. Figure 2: Data acquisition environment and dual-view camera deployment of HCCB. The figure shows the relative positions and fields of view of the two cameras. The right side presents real lecture hall scenes captured from the two viewpoints. The blue and yellow regions indicate the primary coverage areas of the two viewpoints, respectively [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Local annotation examples of seven classroom behavior categories in HCCB [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Full-image annotation visualization under dual-view classroom scenes. important characteristics. First, the dominant categories show strong appearance similarity. For example, Heads Up and Looking Around both correspond to head-raised postures, whereas Reading, Writing, and Bowing Head all involve head-down states. This requires the model to perform fine-grained behavior discrimination. Second, the low-fre… view at source ↗
Figure 5
Figure 5. Figure 5: Image-level instance density distribution across datasets. The box indicates the interquartile range, the line inside the box indicates the median, and the whiskers indicate the distribution range. substantially, and their overall trends are less continuous than that of HCCB. This phenomenon is related to the spatial distance difference between front and rear rows in tiered lecture classrooms and the fixed… view at source ↗
Figure 6
Figure 6. Figure 6: Spatial heatmap of object centers in HCCB. The x- and y-axes denote the normalized coordinates of bounding-box centers. 4. Methodology and Design 4.1. System Overview The detection difficulty of HCCB is not a simple performance degradation caused by an increased number of objects, but arises from structural failure modes jointly induced by extremely high instance density, asymmetric occlusion, depth-induce… view at source ↗
Figure 7
Figure 7. Figure 7: Depth-binned occlusion ratio curves. The horizontal axis 𝑘 represents the depth-bin index from the front rows to the back rows. The vertical axis 𝑃 (𝑘) bin (𝛿) denotes the occlusion proportion in the corresponding depth region under threshold 𝛿. VSS Proxy [12, 33] are used to enhance local structural cues and long-range contextual dependencies. Then, scale￾aligned fusion is performed under a unified spatia… view at source ↗
Figure 8
Figure 8. Figure 8: Cross-dataset relationship between object scale and vertical position. Points represent individual object instances, the curve shows the binned average trend, and the shaded region indicates the corresponding statistical range [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Behavior-category distribution across different depth regions in HCCB. OCDetect is designed to address candidate-box noise and confidence distortion at the detection output stage. In highly congested classrooms, desk-chair edges, partial limbs, occlusion boundaries, and background textures can generate a large number of low-quality candidate boxes. In addition to the original classification and regression … view at source ↗
Figure 10
Figure 10. Figure 10: Overall framework of ODER-HSFNet. improves candidate-box ranking and post-processing stability in dense scenes by suppressing false responses induced by background noise and occlusion boundaries. From the perspective of the correspondence between structural difficulties and model components, ODER restores effective evidence around occlusion edges and far-field degraded regions at the local sampling level,… view at source ↗
Figure 11
Figure 11. Figure 11: Structure of the ODER module. prior feature 𝑃 . This process is formulated as 𝑍 = SiLU ( 𝑊in(𝑋) ) , (7) 𝑃 = 𝜙𝑝 (𝑍), (8) where 𝑊in denotes the input channel alignment mapping, 𝜙𝑝 (⋅) denotes the prior encoding mapping, and SiLU denotes the nonlinear activation function. The shared prior 𝑃 is simultaneously used for offset prediction, sampling￾weight estimation, and residual amplitude modulation, enabling d… view at source ↗
Figure 12
Figure 12. Figure 12: Detailed three-branch structure of the ODER module. The figure shows the internal computation flows of the DEAO, HDSG, and Scale branches, including offset generation, sampling-weight prediction, and residual-amplitude modulation. regions. Therefore, ODER further uses topology-aware routing to assign spatial weights to different sampling results. Specifically, the shared prior 𝑃 is flattened into spatial … view at source ↗
Figure 13
Figure 13. Figure 13: Overall architecture of the Hypergraph-State Spatial Fusion (HSSF) mechanism. (A) Cross-scale state proxy construction. (B) FuseModule bottleneck. (C) High-order relation fusion bottleneck. edges and far-field degraded regions through bounded edge resampling, topology-aware routing, and sample-level residual modulation, without disrupting the stability of backbone features. 4.3. Hypergraph-State Spatial F… view at source ↗
Figure 14
Figure 14. Figure 14: Illustration of the Occlusion-Calibrated Detection Head (OCDetect). where 𝜙𝑒 (⋅) and 𝜙𝑣 (⋅) denote hyperedge feature transformation and node feature transformation, respectively. A hyperedge in the hypergraph can simultaneously connect multiple spatial nodes, making it more suitable for describing multi-instance associations within the same depth region, behavior pattern, or occlusion structure. Meanwhile… view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of normalized confusion matrices for classroom action recognition. The categories include Reading (RD), Writing (WR), Heads Up (HU), Sleeping (SL), Looking Around (LA), Bowing Head (BH), Using Phone (UP), and Background (BG). classroom scenario. In the third sample, YOLOv10s obtains Box 26, Pred 30, Box-Cls 24, and FP 6, while YOLOv13s obtains Box 26, Pred 33, Box-Cls 25, and FP 8. ODER-HSFNet … view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative comparison of classroom behavior detection on the HCCB and SCB-D3-S datasets. The upper-left label indicates the original image or detector, and the upper-right label summarizes instance statistics: Total/Box denotes ground-truth annotations, Pred denotes predicted boxes, Box-Cls denotes correctly localized and classified detections, and FP denotes false positives. mAP50 does not continue to i… view at source ↗
Figure 17
Figure 17. Figure 17: Visualization of the ODER boundary-guided route in partially occluded scenes. The blue circles highlight local regions with occlusion interference, the orange dashed boxes indicate occluded targets, and the yellow routes represent the boundary-guided responses generated by ODER [PITH_FULL_IMAGE:figures/full_fig_p030_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Cross-scale response heatmap comparison between Adaptive Hyperedges and HSSF. Warmer colors indicate stronger feature responses. HSSF enhances responses on student instances and suppresses irrelevant background activations [PITH_FULL_IMAGE:figures/full_fig_p031_18.png] view at source ↗
read the original abstract

Student behavior detection is important for intelligent classroom analysis but remains challenging in large-class scenarios due to dense instance co-occurrence, asymmetric occlusion, depth-wise scale variation, and fine-grained semantic degradation in distant targets. Existing classroom behavior datasets and general-purpose detectors are insufficient to characterize and address these challenges. This paper constructs the Highly Congested Classroom Behavior (HCCB) dataset, containing 50,229 student behavior instances across seven categories: reading, writing, heads up, sleeping, looking around, bowing head, and using phone. HCCB provides a challenging benchmark that integrates dense distributions, severe occlusion, scale variation, and fine-grained behavioral semantics. To address these issues, we propose ODER-HSFNet, a YOLO-based detection framework tailored to highly crowded classrooms. At its core, ODER-HSFNet introduces three task-specific innovations: the Occlusion-aware Deformable Edge Rectifier (ODER), which strengthens boundary evidence under occlusion; the Hypergraph-State Spatial Fusion (HSSF) module, which integrates local structure enhancement, state-space contextual modeling, and high-order relation aggregation; and the Occlusion-Calibrated Detection Head (OCDetect), which suppresses low-quality Pre-NMS candidates and reduces false positives from occlusion boundaries and neighboring instances. Experiments on two classroom behavior detection datasets show that ODER-HSFNet outperforms mainstream YOLO-series methods, achieving 60.60%/80.12% mAP50:95/mAP50 on HCCB and 57.36%/74.65% on SCB-D3-S. Ablation studies further verify the effectiveness of the proposed design for highly crowded classroom behavior detection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces the Highly Congested Classroom Behavior (HCCB) dataset with 50,229 instances across seven behavior categories and proposes ODER-HSFNet, a YOLO-based detector incorporating the Occlusion-aware Deformable Edge Rectifier (ODER), Hypergraph-State Spatial Fusion (HSSF) module, and Occlusion-Calibrated Detection Head (OCDetect). It reports that ODER-HSFNet outperforms YOLO-series baselines, achieving 60.60%/80.12% mAP50:95/mAP50 on HCCB and 57.36%/74.65% on SCB-D3-S, with ablation studies verifying module contributions.

Significance. If the quantitative claims are reproducible, the work contributes a new challenging benchmark for crowded classroom scenes and task-specific modules addressing occlusion and scale issues. Credit is given for conducting ablation studies that isolate module effects and for evaluating on an external dataset (SCB-D3-S).

major comments (2)
  1. [Experiments] Experiments section: The mAP improvements are reported without error bars, standard deviations across runs, or statistical tests, which is required to establish that outperformance over YOLO baselines is reliable rather than due to random variation on the newly constructed HCCB dataset.
  2. [Experiments] Experimental setup: No details are provided on train/validation/test splits, cross-validation procedure, or hyperparameter selection for the HCCB and SCB-D3-S evaluations; this directly affects the load-bearing claim of module effectiveness and outperformance.
minor comments (2)
  1. [Introduction] The abstract and introduction list four challenges but do not explicitly map each to the three modules with supporting citations or preliminary experiments; a table linking challenges to modules would improve clarity.
  2. Notation for the three modules (ODER, HSSF, OCDetect) should be introduced consistently with full names on first use in all sections.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on experimental rigor. We address each major comment below and will revise the manuscript accordingly to strengthen reproducibility and statistical reliability.

read point-by-point responses

  1. Referee: [Experiments] Experiments section: The mAP improvements are reported without error bars, standard deviations across runs, or statistical tests, which is required to establish that outperformance over YOLO baselines is reliable rather than due to random variation on the newly constructed HCCB dataset.

    Authors: We agree that the absence of error bars and statistical tests limits the strength of the outperformance claims. In the revised manuscript, we will conduct additional training runs using different random seeds, report mean mAP50:95 and mAP50 values with standard deviations, and include paired statistical tests (e.g., t-tests) against the YOLO baselines to demonstrate that the gains are not due to random variation. revision: yes

  2. Referee: [Experiments] Experimental setup: No details are provided on train/validation/test splits, cross-validation procedure, or hyperparameter selection for the HCCB and SCB-D3-S evaluations; this directly affects the load-bearing claim of module effectiveness and outperformance.

    Authors: We acknowledge that these experimental details are necessary for full reproducibility. The revised manuscript will add a dedicated subsection in the Experiments section that explicitly describes the train/validation/test splits for both HCCB and SCB-D3-S, the cross-validation procedure (if used), and the hyperparameter selection process. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper introduces the HCCB dataset and proposes ODER-HSFNet with three modules (ODER, HSSF, OCDetect) whose effectiveness is shown via direct mAP comparisons to YOLO baselines on HCCB and the external SCB-D3-S dataset, plus ablation studies. No equations or claims reduce a prediction to a fitted input by construction, no self-citations bear the central load, and no uniqueness theorems or ansatzes are smuggled in. The reported results are standard empirical measurements on held-out and external data, making the evaluation chain self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 3 invented entities

The central claim rests on the empirical effectiveness of three newly introduced modules and the representativeness of the HCCB dataset; no free parameters, mathematical axioms, or externally validated invented entities are invoked beyond standard deep learning assumptions.

invented entities (3)
  • Occlusion-aware Deformable Edge Rectifier (ODER) no independent evidence
    purpose: Strengthens boundary evidence under occlusion
    New module introduced to address occlusion challenge in crowded scenes.
  • Hypergraph-State Spatial Fusion (HSSF) module no independent evidence
    purpose: Integrates local structure enhancement, state-space contextual modeling, and high-order relation aggregation
    Custom fusion module for contextual modeling in dense scenes.
  • Occlusion-Calibrated Detection Head (OCDetect) no independent evidence
    purpose: Suppresses low-quality Pre-NMS candidates and reduces false positives from occlusion boundaries
    New detection head design to handle neighboring instances.

pith-pipeline@v0.9.1-grok · 5849 in / 1499 out tokens · 37918 ms · 2026-06-26T14:58:23.541868+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

47 extracted references · 1 linked inside Pith

  1. [1]

    H. Zhou, F. Jiang, J. Si, L. Xiong, and H. Lu. StuArt: Individualized classroom observation of students with automatic behavior recognition and tracking, 2022. arXiv preprint

    2022

  2. [2]

    F. C. Lin, H. H. Ngo, C. R. Dow, K. H. Lam, and H. L. Le. Student behavior recognition system for the classroom environment based on skeleton pose estimation and person detection.Sensors, 21(16):5314, 2021

    2021

  3. [3]

    Yang and T

    F. Yang and T. Wang. SCB-Dataset3: A benchmark for detecting student classroom behavior, 2023. arXiv preprint

    2023

  4. [4]

    F. Yang. Student classroom behavior detection based on improved YOLOv7, 2023. arXiv preprint

    2023

  5. [5]

    Featurepyramidnetworksforobjectdetection

    T.Y.Lin,P.Doll’ar,R.Girshick,K.He,B.Hariharan,andS.Belongie. Featurepyramidnetworksforobjectdetection. InProceedingsofthe IEEE Conference on Computer Vision and Pattern Recognition, pages 2117–2125, 2017

    2017

  6. [6]

    Redmon, S

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You only look once: Unified, real-time object detection. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 779–788, 2016

    2016

  7. [7]

    YOLOv13:Real-timeobjectdetectionwithhypergraph- enhanced adaptive visual perception, 2025

    M.Lei,S.Li,Y.Wu,H.Hu,Y.Zhou,X.Zheng,G.Ding,S.Du,Z.Wu,andY.Gao. YOLOv13:Real-timeobjectdetectionwithhypergraph- enhanced adaptive visual perception, 2025. arXiv preprint

    2025

  8. [8]

    S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. InAdvances in Neural Information Processing Systems, 2015

    2015

  9. [9]

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C. Y. Fu, and A. C. Berg. SSD: Single shot multibox detector. InEuropean Conference on Computer Vision, pages 21–37, 2016

    2016

  10. [10]

    Carion, F

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko. End-to-end object detection with transformers. InEuropean Conference on Computer Vision, pages 213–229, 2020

    2020

  11. [11]

    J. Dai, H. Qi, Y. Xiong, Y. Li, G. Zhang, H. Hu, and Y. Wei. Deformable convolutional networks. InProceedings of the IEEE International Conference on Computer Vision, pages 764–773, 2017

    2017

  12. [12]

    Gu and T

    A. Gu and T. Dao. Mamba: Linear-time sequence modeling with selective state spaces, 2023. arXiv preprint

    2023

  13. [13]

    Hypergraphneuralnetworks

    Y.Feng,H.You,Z.Zhang,R.Ji,andY.Gao. Hypergraphneuralnetworks. InProceedingsoftheAAAIConferenceonArtificialIntelligence, volume 33, pages 3558–3565, 2019

    2019

  14. [14]

    StudentclassroombehaviordetectionbasedonYOLOv7-BRAandmulti-modelfusion,2023

    F.Yang,T.Wang,andX.Wang. StudentclassroombehaviordetectionbasedonYOLOv7-BRAandmulti-modelfusion,2023. arXivpreprint

    2023

  15. [15]

    S. Shao, Z. Zhao, B. Li, T. Xiao, G. Yu, X. Zhang, and J. Sun. CrowdHuman: A benchmark for detecting human in a crowd, 2018. arXiv preprint

    2018

  16. [16]

    Zhang, R

    S. Zhang, R. Benenson, and B. Schiele. CityPersons: A diverse dataset for pedestrian detection. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3213–3221, 2017

    2017

  17. [17]

    Zhang, Y

    S. Zhang, Y. Xie, J. Wan, H. Xia, S. Z. Li, and G. Guo. WiderPerson: A diverse dataset for dense pedestrian detection in the wild.IEEE Transactions on Multimedia, 22(2):380–393, 2020

    2020

  18. [18]

    V. A. Sindagi, R. Yasarla, and V. M. Patel. JHU-CROWD++: Large-scale crowd counting dataset and a benchmark method, 2020. arXiv preprint

    2020

  19. [19]

    Bodla, B

    N. Bodla, B. Singh, R. Chellappa, and L. S. Davis. Soft-NMS: Improving object detection with one line of code. InProceedings of the IEEE International Conference on Computer Vision, pages 5561–5569, 2017

    2017

  20. [20]

    S. Liu, D. Huang, and Y. Wang. Adaptive NMS: Refining pedestrian detection in a crowd. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6459–6468, 2019. Xu et al.:Preprint submitted to ElsevierPage 31 of 32 Smart Classroom Behavior Analysis with HCCB

    2019

  21. [21]

    X. Wang, T. Xiao, Y. Jiang, S. Shao, J. Sun, and C. Shen. Repulsion loss: Detecting pedestrians in a crowd. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7774–7783, 2018

    2018

  22. [22]

    X. Chu, A. Zheng, X. Zhang, and J. Sun. Detection in crowded scenes: One proposal, multiple predictions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12214–12223, 2020

    2020

  23. [23]

    S. Liu, L. Qi, H. Qin, J. Shi, and J. Jia. Path aggregation network for instance segmentation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8759–8768, 2018

    2018

  24. [24]

    EfficientDet:Scalableandefficientobjectdetection

    M.Tan,R.Pang,andQ.V.Le. EfficientDet:Scalableandefficientobjectdetection. InProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition, pages 10781–10790, 2020

    2020

  25. [25]

    CascadeR-CNN:Delvingintohighqualityobjectdetection

    Z.CaiandN.Vasconcelos. CascadeR-CNN:Delvingintohighqualityobjectdetection. InProceedingsoftheIEEEConferenceonComputer Vision and Pattern Recognition, pages 6154–6162, 2018

    2018

  26. [26]

    Zhang, C

    S. Zhang, C. Chi, Y. Yao, Z. Lei, and S. Z. Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9759–9768, 2020

    2020

  27. [27]

    Z. Ge, S. Liu, Z. Li, O. Yoshie, and J. Sun. OTA: Optimal transport assignment for object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 303–312, 2021

    2021

  28. [28]

    Generalizedfocalloss:Learningqualifiedanddistributedboundingboxes for dense object detection

    X.Li,W.Wang,L.Wu,S.Chen,X.Hu,J.Li,J.Tang,andJ.Yang. Generalizedfocalloss:Learningqualifiedanddistributedboundingboxes for dense object detection. InAdvances in Neural Information Processing Systems, volume 33, pages 21002–21012, 2020

    2020

  29. [29]

    X.Zhu,W.Su,L.Lu,B.Li,X.Wang,andJ.Dai.DeformableDETR:Deformabletransformersforend-to-endobjectdetection.InInternational Conference on Learning Representations, 2021

    2021

  30. [30]

    Pyramidvisiontransformer:Aversatilebackbonefordense prediction without convolutions

    W.Wang,E.Xie,X.Li,D.P.Fan,K.Song,D.Liang,T.Lu,P.Luo,andL.Shao. Pyramidvisiontransformer:Aversatilebackbonefordense prediction without convolutions. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 568–578, 2021

    2021

  31. [31]

    Swintransformer:Hierarchicalvisiontransformerusingshiftedwindows

    Z.Liu,Y.Lin,Y.Cao,H.Hu,Y.Wei,Z.Zhang,S.Lin,andB.Guo. Swintransformer:Hierarchicalvisiontransformerusingshiftedwindows. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012–10022, 2021

    2021

  32. [32]

    X. Wang, R. Girshick, A. Gupta, and K. He. Non-local neural networks. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7794–7803, 2018

    2018

  33. [33]

    Y. Liu, Y. Tian, Y. Zhao, H. Yu, L. Xie, Y. Wang, Q. Ye, J. Jiao, and Y. Liu. VMamba: Visual state space model. InAdvances in Neural Information Processing Systems, 2024

    2024

  34. [34]

    H. Hu, J. Gu, Z. Zhang, J. Dai, and Y. Wei. Relation networks for object detection. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3588–3597, 2018

    2018

  35. [35]

    T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. InInternational Conference on Learning Representations, 2017

    2017

  36. [36]

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, and H. Adam. MobileNets: Efficient convolutional neural networks for mobile vision applications, 2017. arXiv preprint

    2017

  37. [37]

    Jocher, A

    G. Jocher, A. Chaurasia, and J. Qiu. Ultralytics YOLO, 2023. Ultralytics

    2023

  38. [38]

    B. Sun, Y. Wu, K. Zhao, et al. Student class behavior dataset: A video dataset for recognizing, detecting, and captioning students’ behaviors in classroom scenes.Neural Computing and Applications, 33:8335–8354, 2021

    2021

  39. [39]

    J.ZhaoandH.Zhu.CBPH-Net:Asmallobjectdetectorforbehaviorrecognitioninclassroomscenarios.IEEETransactionsonInstrumentation and Measurement, 2023

    2023

  40. [40]

    MicrosoftCOCO:Commonobjectsincontext

    T.Y.Lin,M.Maire,S.Belongie,J.Hays,P.Perona,D.Ramanan,P.Doll’ar,andC.L.Zitnick. MicrosoftCOCO:Commonobjectsincontext. InEuropean Conference on Computer Vision, pages 740–755, 2014

    2014

  41. [41]

    C. Li, L. Li, H. Jiang, K. Weng, Y. Geng, L. Li, et al. YOLOv6: A single-stage object detection framework for industrial applications, 2022. arXiv preprint

    2022

  42. [42]

    Terven and D

    J. Terven and D. Cordova-Esparza. A comprehensive review of YOLO architectures in computer vision: From YOLOv1 to YOLOv8 and YOLO-NAS.Machine Learning and Knowledge Extraction, 5(4):1680–1716, 2023

    2023

  43. [43]

    YOLOv9:Learningwhatyouwanttolearnusingprogrammablegradientinformation

    C.Y.Wang,I.H.Yeh,andH.Y.M.Liao. YOLOv9:Learningwhatyouwanttolearnusingprogrammablegradientinformation. InEuropean Conference on Computer Vision, 2024

    2024

  44. [44]

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, and G. Ding. YOLOv10: Real-time end-to-end object detection. InAdvances in Neural Information Processing Systems, 2024

    2024

  45. [45]

    Khanam and M

    R. Khanam and M. Hussain. YOLOv11: An overview of the key architectural enhancements, 2024. arXiv preprint

    2024

  46. [46]

    Y. Tian, Q. Ye, and D. Doermann. YOLOv12: Attention-centric real-time object detectors, 2025. arXiv preprint

    2025

  47. [47]

    Ultralytics yolo26: Unified real-time end-to-end vision models.arXiv preprint arXiv:2606.03748, 2026

    Glenn Jocher, Jing Qiu, Mengyu Liu, Shuai Lyu, Fatih Cagatay Akyon, and Muhammet Esat Kalfaoglu. Ultralytics yolo26: Unified real-time end-to-end vision models.arXiv preprint arXiv:2606.03748, 2026. Xu et al.:Preprint submitted to ElsevierPage 32 of 32

    Pith/arXiv arXiv 2026