Pith. sign in

REVIEW 4 major objections 5 minor 52 references

Robust and Efficient Motion Reasoning for Privacy-Aware Classroom Incident Recognition

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A small pose-only model trained on generated classroom videos can recognize incidents and transfer to real CCTV footage.

desk verdict A useful pilot benchmark for classroom incident recognition, with a real evaluation-protocol flaw that leaves the headline gains unverified. read the letter →

arxiv 2608.05115 v1 pith:C6QX7VS6 submitted 2026-08-05 cs.CV cs.AIcs.ETcs.HCcs.LG

Paritosh Parmar , Landy Lan , Hong Yang , Chen Yi , Chiat Pin Tay This is my paper
classification cs.CVcs.AIcs.ETcs.HCcs.LG
keywords classroomincidentrecognitionprivacy-preservingactionskeleton-basedknowledgedistillationkinematicmotionfeaturessyntheticvideogenerationzero-shotdomaintransfer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a lightweight, privacy-preserving action-recognition model can identify classroom safety incidents from human joint positions alone, after being trained on a mix of generated and real classroom videos. The authors build a new benchmark of seven incident classes plus a background class, with 1,296 synthetic CCTV-style videos and 574 real samples, then train a multi-order teacher on joint positions, velocities, and accelerations. They distill that teacher into a small student that consumes only joint positions at inference. On the synthetic test split the student reaches 71.78% accuracy, ahead of pose-based baselines such as PoseC3D (70.54%), and on real footage without any fine-tuning it reaches 63.41%, ahead of MSG3D (59.23%), at roughly one-tenth the computational cost. If correct, this makes skeleton-only monitoring a viable privacy-aware option for classrooms and shows synthetic video can serve as training data for safety-critical events that are hard to collect.

What carries the argument

The load-bearing mechanism is a two-stage pipeline. First, joint trajectories are denoised (missing frames interpolated and trajectories smoothed with a filtering step) and then expanded into a hierarchy of kinematic orders via recursive finite differences: velocity $v_t^a=(j_t^a-j_{t-\delta}^a)/\delta$ and acceleration $a_t^a=(v_t^a-v_{t-\delta}^a)/\delta$, with the general $n$-th order defined recursively. Each order feeds an order-specific graph-convolutional backbone; their projected latent features are combined by weighted averaging with weights $w_0=1.0, w_1=0.8, w_2=0.2$. Second, a multi-objective distillation objective (temperature-scaled KL divergence between teacher and student softmax distributions plus a standard cross-entropy term, with weights $\lambda_{KD}=0.3$, $\lambda_{Cls}=1.0$) compresses the fused teacher into a student that sees only zeroth-order positions. The student then carries the multi-order motion reasoning implicitly, so inference runs a single small stream.

What would settle it

Compute keypoint confidence, frame-to-frame joint displacement, and sequence-length statistics for the synthetic and real sets, then retrain the student on synthetic poses perturbed to match the real statistics; if accuracy falls back to baseline levels, the transfer is explained by pose-distribution similarity, whereas if it stays near 63.41%, the kinematic-distillation claim is supported.

Watch

Extended reading notes

Core claim

The central claim is that classroom incidents—falling, punching, jumping, kicking, throwing, running, sitting—can be recognized from skeletal trajectories alone, and that a deliberately small model can do it better than larger pose-based baselines. The method works by representing each action at three kinematic orders (joint positions, velocities, accelerations), training separate backbones on each, and fusing their latent features into a teacher. A multi-objective distillation loss then trains a student, roughly one-sixth the teacher's size and one-tenth the inference cost of the compared baselines, to match the teacher's softened class probabilities while operating only on zeroth-order joint positions. On the paper's synthetic test split the student reports 71.78% accuracy versus 70.54% for the strongest baseline (PoseC3D); without any fine-tuning on real data it reports 63.41% on the real-world split versus 59.23% for the strongest baseline (MSG3D). The paper interprets these numbers as evidence that the distilled student internalizes velocity- and acceleration-like reasoning from the teacher, and that pose abstraction plus denoised kinematic features transfer across the synthetic-to-real gap.

Load-bearing premise

The transfer claim depends on the assumption that the joint positions extracted from generated videos resemble those extracted from real classroom CCTV footage well enough to train on; the paper does not report a distributional comparison of the extracted poses, so if the two differ systematically, the 63.41% figure may reflect dataset similarity rather than method robustness.

Editorial extensions

If this is right

  • A deployed recognizer could run on joint trajectories only, so the institution holding CCTV footage never has to share raw frames with the model.
  • The student's compute and parameter footprint is about one-tenth that of the compared pose-based baselines, which puts it in range of inexpensive edge hardware.
  • Training on generated CCTV-style videos transfers to real classroom footage without fine-tuning, meaning synthetic data can stand in for incidents that are too rare or unsafe to record.
  • Distillation lets a single-order student match or beat a multi-order teacher, so velocity and acceleration reasoning can be had without computing derivatives at inference.

Reading between the lines

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

  • Editorial inference: the same generate-then-distill recipe could be adapted to other privacy-sensitive monitoring settings—elder-care homes, changing rooms, or special-education classrooms—where incident data is scarce and identifiable footage is off-limits.
  • Editorial inference: the paper does not compare against models that see raw RGB frames or vision-language embeddings; such a comparison would show how much accuracy is traded away for the privacy and efficiency the pose-only student buys.
  • Editorial inference: a direct test of the transfer claim would be to train the same student on real poses only and compare with the synthetic-trained model; the gap would quantify how much of the 63.41% comes from the synthetic data rather than from the distillation itself.
  • Editorial inference: because the background class is a large meta-class, per-class confusion matrices for the seven incident classes plus background would reveal whether the headline accuracy hides systematic confusion between, say, kicking and running; the paper reports such gains mainly for kick, sit, and run in the appendix.
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

4 major / 5 minor

Summary. This pilot study targets privacy-aware classroom incident recognition from pose sequences. The authors construct a new hybrid benchmark of 1,296 synthetic CCTV-style classroom videos plus 574 real-world pose samples, and propose a method that builds hierarchical kinematic representations (positions, velocities, accelerations) into a multi-order teacher that is distilled into a small single-order student operating only on joint positions. Experiments on the synthetic split claim the student reaches 71.78% accuracy, outperforming pose-based baselines such as PoseC3D, and 63.41% zero-shot accuracy on the real-world split. The paper also reports efficiency gains and promises public release of the benchmark and code.

Significance. If the performance claims hold, the paper would make a useful contribution: a new privacy-oriented dataset, a clearly described and computationally efficient skeleton-based method, and a zero-shot synthetic-to-real transfer protocol. The manuscript is generally well written, the equations for kinematic features and distillation are standard and correctly formulated, and the authors are explicit about the pilot nature of the study. However, the central quantitative claims currently rest on an evaluation protocol that selects hyperparameters on the test set and on differences that are close to the sampling noise of the reported test sizes. The significance is therefore conditional on a stronger evaluation that separates model selection from final reporting and provides uncertainty estimates.

major comments (4)
  1. [§5.1, §5.3, Figure 8] The paper defines an 80/20 train/test split in §5.1, but §5.3 describes sweeping fusion weights (w0, w1, w2) and distillation loss weight λ_KD over a grid and selecting the best accuracy, all without mentioning any validation split. This means the reported 71.78% synthetic accuracy is a selected maximum on the same 20% test set, not an unbiased estimate. The student model carried into the zero-shot real-world evaluation (§5.2) was also chosen under this protocol. The authors should introduce a proper validation split (or nested cross-validation), select hyperparameters on it, and then report test accuracy with the final fixed configuration.
  2. [Table 1, §5.1] The claimed 1.24-point improvement over PoseC3D corresponds to roughly 3 clips on a test set of about 260 clips, and the 4.18-point lead over MSG3D in Table 2 is about 1.4 standard errors on the 574 real-world clips. No confidence intervals, repeated seeds, or significance tests are reported. Given the small sample sizes and the hyperparameter selection issue, the central claim that the method 'outperforms substantially larger baselines' is not statistically established. The authors should report accuracies with confidence intervals (e.g., bootstrap or multiple seeds) and, where possible, a paired test across the same test clips.
  3. [§3.1, §5.2] The zero-shot synthetic-to-real transfer results depend on the assumption that YOLO-extracted pose sequences from generated videos and real CCTV footage are comparable. The paper does not report pose-detection accuracy, per-domain pose statistics, or any distribution analysis for the pose features on either domain. Furthermore, the synthetic pipeline includes a manual quality filter in which three authors and preschool teachers discard videos that do not look realistic; this curation may reduce the domain gap by construction. Without reporting pose-quality metrics and a domain-shift analysis, the 63.41% transfer number cannot be separated from dataset similarity or curation effects. The authors should add quantitative pose-quality measures (e.g., detection confidence, joint visibility, trajectory smoothness) on both synthetic and real sets and compare their distributions.
  4. [§5.3, Figure 8 (3)] The ablation shows that the distilled student outperforms its own multi-order teacher. This is surprising and the paper's explanation (smaller size leads to more transferable representations) is speculative. Because the teacher and student are not evaluated with the same hyperparameter selection protocol—the teacher uses the manually set {w0,w1,w2}={1.0,0.8,0.2} and the student's λ_KD is tuned on the test set—this specific claim may be an artifact of test-set tuning. The authors should either provide a comparison where both teacher and student use the same validation-based selection, or temper the claim.
minor comments (5)
  1. [Figure 2, caption and text] The text in §1 says 'Performance & GFLOPs are along Y & X-axes' while the caption says 'Performance & GFLOPs are along Y & X-axes, respectively'; please make the axis labels explicit and consistent, and indicate whether the circle diameter represents parameters or model size.
  2. [Eq. (7)] The notation z_t and z_s for teacher and student logits could be confused with the time index t used in the kinematic equations. Consider renaming to z_tea and z_stu or using an unambiguous superscript.
  3. [§5.3, Figure 8 (1)] The text says 'we sweep each weight w from 0 to 1 with a step size of 0.1' but the figure appears to show a two-dimensional surface over w0 and w1 with w2 fixed at 0.2. Please clarify the exact grid and the number of trials.
  4. [§5.1 and Figure 2] The paper alternates between 'one-tenth of the parameters' and 'one-tenth of the computational cost.' These are different claims; please specify which is reported and provide both parameter counts and FLOPs in a consistent table.
  5. [§3.3, 'Realworld Dataset'] The real-world dataset section is very brief. Please specify how the 574 samples are distributed across the eight classes (seven actions plus background), how many distinct classrooms/subjects are involved, and how temporal boundaries were annotated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the reported accuracies are empirical measurements, not reductions to fitted inputs or self-citations.

full rationale

The derivation chain is self-contained. Equations (1)-(3) define velocity and acceleration as finite differences of joint positions; these are feature transformations, not redefinitions of the action labels. Equation (6) is a convex combination of order-specific representations, and Equation (10) is a standard classification-plus-KL distillation loss; neither encodes the reported accuracies by construction. The student's outputs are learned from teacher soft targets and ground-truth labels, so the 71.78% and 63.41% numbers are empirical measurements, not identities. No load-bearing self-citations appear; references [10] and [11] are external baselines and toolboxes, and NTURGBD pretraining is an external benchmark. The only near-concern is hyperparameter selection: Section 5.3 sweeps fusion weights and lambda_KD and reports the best test accuracy, which is a fitting and selection-bias issue rather than circularity, because the chosen weights do not mathematically force the reported performance. No uniqueness theorem, renamed known result, or ansatz-smuggling-by-self-citation is present. The zero-shot synthetic-to-real evaluation is an external target that does not reduce to any fitted parameter of the method.

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

The central claim rests on the reliability of automatically extracted pose, the fidelity of the human-filtered synthetic data as a proxy for real classrooms, and standard distillation assumptions. Test-set tuning of fusion and distillation weights adds free parameters that are not independently validated.

free parameters (4)
  • Fusion weights w0, w1, w2 = 1.0, 0.8, 0.2
    Manually set and tuned by grid search on the test split (Figure 8); weights control the contribution of positions, velocities, and accelerations in Equation 6.
  • Distillation loss weight lambda_KD = 0.3
    Chosen by sweeping on the test set (Figure 8); balances the KD and classification losses in Equation 10.
  • Distillation temperature tau = 4
    Set to a typical value; no sensitivity analysis is provided, and it affects the soft targets in Equations 7 and 8.
  • Temporal interval delta = not reported (assumed 1)
    Used in finite differences in Equations 1 through 3; the value is not stated and could change the scale of velocity and acceleration features.
assumptions (4)
  • domain assumption YOLO pose estimates on both synthetic and real footage are accurate enough that higher-order kinematic features encode true motion rather than detector noise.
    The kinematic teacher in Sections 4.3 through 4.5 depends on this; Figure 6 shows noise amplification but no quantitative pose-error analysis is reported.
  • domain assumption The manually filtered set of generated videos and teacher-validated annotations is a faithful proxy for real classroom incidents.
    Benchmark validity and zero-shot transfer (Sections 3.1, 3.2, and 5.2) rest on this; the selection is human-in-the-loop and could introduce bias.
  • standard math Standard knowledge distillation assumptions: softened teacher logits convey beneficial inter-class structure to the student.
    Equations 7 through 10 follow Hinton et al.; this is accepted in the literature and not proved in the paper.
  • domain assumption Anonymized pose keypoints are sufficient to protect student privacy in deployment when raw video is discarded.
    The privacy-aware framing in Section 4.2 assumes keypoints are non-identifiable, but pose data can still be re-identifiable; no de-identification analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust and Efficient Motion Reasoning for Privacy-Aware Classroom Incident Recognition." pith.science (2026). https://pith.science/paper/C6QX7VS6

@misc{pith2026260805115,
  author       = {Pith},
  title        = {Pith review of: Robust and Efficient Motion Reasoning for Privacy-Aware Classroom Incident Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6QX7VS6}},
  note         = {Machine review of arXiv:2608.05115}
}
read the original abstract

Can computer vision help make classrooms safer? In this pilot study, we investigate privacy-aware and computationally efficient classroom incident recognition from CCTV-style observations. This setting remains underexplored, with limited benchmarks and few methods designed for the privacy, efficiency, and generalization demands of real-world deployment. We introduce a novel hybrid benchmark combining generative CCTV-style videos with real-world classroom pose data, and propose a lightweight, but robust motion-reasoning framework motivated by the observation that many incidents differ more in motion direction, speed, acceleration, and intensity than in pose alone. To that end, our method first constructs hierarchical kinematic representations of human actions. Our method then distills hierarchical, multi-order kinematic reasoning from a large teacher into a much smaller single-order student, enabling efficient per-person inference while preserving expressive motion understanding. Experiments show that our model outperforms substantially larger baselines at less than one-tenth of their computational cost, while also demonstrating stronger out-of-domain motion reasoning and zero-shot synthetic-to-real generalization. We will publicly release the benchmark, codebase, and supporting tools to facilitate further research in privacy-aware classroom safety.

Figures

Figures reproduced from arXiv: 2608.05115 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Performance vs. Efficiency for various models on our synthetic dataset. Performance & GFLOPs are along Y & X￾axes, respectively; while circle diameter indicates the model size in million parameters (shorter diameter, smaller model size). typically underrepresented in standard action recognition datasets; and domain-specific action categories, such as kicking, punching, and other safety-relevant behaviors, which may … view at source ↗
Figure 3
Figure 3. Our synthetic dataset generation pipeline. We capitalize on these advances to design an efficient stu￾dent model. Privacy-Preserving Skeleton-Based Action Recognition. Skeleton-based action recognition has emerged as a privacy-conscious alternative to RGB-based video analy￾sis, as pose sequences encode human motion while reduc￾ing reliance on the appearance, background, and identity cues present in raw video. Repres… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Generated dataset samples. Our synthetic dataset contains diverse viewpoints, lighting conditions, subjects, objects, classroom settings, clothing, etc. Please zoom in for better view. Annotations. After generating all videos, we annotated the ac￾tion classes and their…
Figure 5
Figure 5. Figure 5: Privacy-preserving pose-based classroom incident recognition approach. and accelerations. Each kinematic order is processed by a ded￾icated action-recognition backbone, and the resulting representa￾tions are fused to form a stronger hierarchical multi-order teacher mod…
Figure 6
Figure 6. Figure 6: Example Noise amplification when naively deriving higher order features. Trajectories of three orders of kinematic data for all joints is plotted (different colors). The backbone is architecture-agnostic and may be instantiated using, for example, spatiotemporal graph …
Figure 7
Figure 7. Figure 7: Our full approach. DO represents order of the input kinematic data; FO represents teacher feature’s order. Specifically, we propose a multi-objective knowledge distilla￾tion (KD) framework that transfers information from the multi￾order fused model into a substantially…
Figure 8
Figure 8. Figure 8: Ablation study results. (1) accuracy surface across a range of w0 & w1; keeping w2 constant at its optimal value of 0.2. (2) plot of accuracy across a range of distillation loss weight, λKD. (3) effect of multiorder hierarchical kinematic representation and distillatio…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 36 canonical work pages

  1. [1]

    A Sheik Abdullah, S Geetha, AB Abdul Aziz, and Utkarsh Mishra. Design of automated model for inspecting and eval- uating handwritten answer scripts: A pedagogical approach with nlp and deep learning.Alexandria engineering journal, 108:764–788, 2024. 2

    work page 2024

  2. [2]

    Transforming pedagogical assess- ment: Ai and computer vision-enhanced classroom observa- tions for experiment-centric learning environments

    Blessing Isoyiza Adeika, Pelumi Olaitan Abiodun, and Olu- dare Adegbola Owolabi. Transforming pedagogical assess- ment: Ai and computer vision-enhanced classroom observa- tions for experiment-centric learning environments. In2024 ASEE Annual Conference & Exposition, 2024. 2

    work page 2024

  3. [3]

    Do deep nets really need to be deep?Advances in neural information processing systems, 27, 2014

    Lei J Ba and Rich Caruana. Do deep nets really need to be deep?Advances in neural information processing systems, 27, 2014. 2

    work page 2014

  4. [4]

    Model compression

    Cristian Bucilu ˇa, Rich Caruana, and Alexandru Niculescu- Mizil. Model compression. InProceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535–541, 2006. 2

    work page 2006

  5. [5]

    Exploring au- tomated recognition of instructional activity and discourse from multimodal classroom data

    Ivo Bueno, Ruikun Hou, Babette B¨uhler, Tim F¨utterer, James Drimalla, Jonathan K Foster, Peter Youngs, Peter Gerjets, Ulrich Trautwein, and Enkelejda Kasneci. Exploring au- tomated recognition of instructional activity and discourse from multimodal classroom data. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6...

    work page 2026

  6. [6]

    VideoCrafter1: Open diffusion models for high-quality video generation.arXiv preprint arXiv:2310.19512, 2023

    Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter1: Open diffusion models for high-quality video generation.arXiv preprint arXiv:2310.19512, 2023. 3

    arXiv

  7. [7]

    Channel-wise topology refinement graph convolution for skeleton-based action recognition

    Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13359–13368, 2021. 9

    work page 2021

  8. [8]

    Engagement detection in online learning: a review.Smart Learning Envi- ronments, 6(1):1–20, 2019

    M Dewan, Mahbub Murshed, and Fuhua Lin. Engagement detection in online learning: a review.Smart Learning Envi- ronments, 6(1):1–20, 2019. 2

    work page 2019

Show all 52 references
  1. [9]

    Bullying10k: A large-scale neuromorphic dataset towards privacy-preserving bullying recognition.Advances in Neural Information Processing Systems, 36:1923–1937,

    Yiting Dong, Yang Li, Dongcheng Zhao, Guobin Shen, and Yi Zeng. Bullying10k: A large-scale neuromorphic dataset towards privacy-preserving bullying recognition.Advances in Neural Information Processing Systems, 36:1923–1937,

    1923

  2. [10]

    Pyskl: Towards good practices for skeleton action recogni- tion

    Haodong Duan, Jiaqi Wang, Kai Chen, and Dahua Lin. Pyskl: Towards good practices for skeleton action recogni- tion. InProceedings of the 30th ACM International Confer- ence on Multimedia, pages 7351–7354, 2022. 8, 9, 10

    2022

  3. [11]

    Revisiting skeleton-based action recognition

    Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2969–2978, 2022. 3, 9

    2022

  4. [12]

    Tim F ¨utterer, Patricia Goldberg, Babette B ¨uhler, Vlasta Sikimi´c, Ulrich Trautwein, Peter Gerjets, Kathleen St ¨urmer, and Enkelejda Kasneci. Artificial intelligence in classroom management: A systematic review on educational purposes, technical implementations, and ethica...

  5. [13]

    Seedance 1.0: Exploring the boundaries of video generation models.arXiv preprint arXiv:2506.09113,

    Yu Gao, Haoyuan Guo, Tuyen Hoang, Weilin Huang, Lu Jiang, Fangyuan Kong, Huixia Li, Jiashi Li, Liang Li, Xi- aojie Li, et al. Seedance 1.0: Exploring the boundaries of video generation models.arXiv preprint arXiv:2506.09113,

    arXiv

  6. [14]

    Bk-sad: A large scale dataset for student activity recognition.JST: Smart Systems and Devices, 2023

    Nguyen Van Giang, Nguyen Gia Chan, Nguyen Tuan Hung, Quang Dich, and Trinh Cong Dong. Bk-sad: A large scale dataset for student activity recognition.JST: Smart Systems and Devices, 2023. 3

    2023

  7. [15]

    Attentive or not? toward a machine learning approach to assessing students’ visible en- gagement in classroom instruction.Educational Psychology Review, 33(1):27–49, 2021

    Patricia Goldberg, ¨Omer S ¨umer, Kathleen St ¨urmer, Wolf- gang Wagner, Richard G¨ollner, Peter Gerjets, Enkelejda Kas- neci, and Ulrich Trautwein. Attentive or not? toward a machine learning approach to assessing students’ visible en- gagement in classroom instruction.Educat...

    2021

  8. [16]

    LTX-Video: Realtime video latent diffusion.arXiv preprint arXiv:2501.00103, 2025

    Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weiss- buch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime video latent diffusion....

    2025 arXiv

  9. [17]

    Distill- ing the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distill- ing the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015. 2

    2015 arXiv

  10. [18]

    Kingma, Ben Poole, Mohammad Norouzi, David J

    Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Sali- mans. Imagen Video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303, 2022. 3

    arXiv

  11. [19]

    CogVideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022

    Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-scale pretraining for text-to-video generation via transformers.arXiv preprint arXiv:2205.15868, 2022. 3

    2022 arXiv

  12. [20]

    Ultralytics YOLO, 2023

    Glenn Jocher, Jing Qiu, and Ayush Chaurasia. Ultralytics YOLO, 2023. 8

    2023

  13. [21]

    Black, David W

    Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. InComputer Vision and Pattern Recognition (CVPR),

  14. [22]

    The kinetics hu- 11 man action video dataset.arXiv preprint arXiv:1705.06950,

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics hu- 11 man action video dataset.arXiv preprint arXiv:1705.06950,

    arXiv

  15. [23]

    Kling-Omni technical report.arXiv preprint arXiv:2512.16776, 2025

    Kling Team. Kling-Omni technical report.arXiv preprint arXiv:2512.16776, 2025. 3

    2025 arXiv

  16. [24]

    Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. InThe IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2020. 3

    2020

  17. [25]

    Computer vision in stem education research

    Gyeonggeon Lee and Xiaoming Zhai. Computer vision in stem education research. InArtificial Intelligence for STEM Education Research: Advanced Methods and Applications, pages 113–141. Springer, 2026. 2

    2026

  18. [26]

    Ntu rgb+ d 120: A large- scale benchmark for 3d human activity understanding.IEEE transactions on pattern analysis and machine intelligence, 42(10):2684–2701, 2019

    Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large- scale benchmark for 3d human activity understanding.IEEE transactions on pattern analysis and machine intelligence, 42(10):2684–2701, 2019. 3, 9

    2019

  19. [27]

    Classroom be- havior recognition using computer vision: A systematic re- view.Sensors, 25(2), 2025

    Qingtang Liu, Xinyu Jiang, and Ruyi Jiang. Classroom be- havior recognition using computer vision: A systematic re- view.Sensors, 25(2), 2025. 2

    2025

  20. [28]

    Disentangling and unifying graph convo- lutions for skeleton-based action recognition

    Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convo- lutions for skeleton-based action recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143–152, 2020. 9

    2020

  21. [29]

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi- person linear model.ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 34(6):248:1–248:16, 2015. 3

    2015

  22. [30]

    Video generation models as world simula- tors.https : / / openai

    OpenAI. Video generation models as world simula- tors.https : / / openai . com / index / video - generation - models - as - world - simulators/,

  23. [31]

    Sora 2 is here.https://openai.com/ index/sora-2/, 2025

    OpenAI. Sora 2 is here.https://openai.com/ index/sora-2/, 2025. Accessed: 2026-05-11. 3

    2025

  24. [32]

    Curran Associates Inc., Red Hook, NY , USA, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K ¨opf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu F...

    2019

  25. [33]

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. InProceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. 3

    2019

  26. [34]

    Learning graph convolutional network for skeleton-based human action recognition by neural searching

    Wei Peng, Xiaopeng Hong, Haoyu Chen, and Guoying Zhao. Learning graph convolutional network for skeleton-based human action recognition by neural searching. InProceed- ings of the AAAI conference on artificial intelligence, pages 2669–2676, 2020. 3

    2020

  27. [35]

    You only look once: Unified, real-time object de- tection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object de- tection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 779–788, 2016. 8

    2016

  28. [36]

    Ntu rgb+d: A large scale dataset for 3d human activity anal- ysis

    Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity anal- ysis. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1010–1019, 2016. 3

    2016

  29. [37]

    Finegym: A hierarchical video dataset for fine-grained action under- standing

    Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action under- standing. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2616–2625,

  30. [38]

    Two- stream adaptive graph convolutional networks for skeleton- based action recognition

    Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two- stream adaptive graph convolutional networks for skeleton- based action recognition. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12018–12027. IEEE, 2019. 3

    2019

  31. [39]

    Skeleton-based action recognition with multi-stream adap- tive graph convolutional networks.IEEE Transactions on Image Processing, 29:9532–9545, 2020

    Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adap- tive graph convolutional networks.IEEE Transactions on Image Processing, 29:9532–9545, 2020. 9

    2020

  32. [40]

    Make-a-video: Text-to-video generation without text-video data

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. InInternational Conference on Learning Represen...

    2023

  33. [41]

    Wan: Open and ad- vanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    Team Wan, Ang Wang, Baole Ai, et al. Wan: Open and ad- vanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025. 3

    2025 arXiv

  34. [42]

    Ordinal-aware multimodal engagement recognition for collaborative learning

    Nha Tran, Dat Ly, Phi Ta, Hung Nguyen, and Hien D Nguyen. Ordinal-aware multimodal engagement recognition for collaborative learning. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2424–2433, 2026. 2

    2026

  35. [43]

    MoCoGAN: Decomposing motion and content for video generation

    Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. MoCoGAN: Decomposing motion and content for video generation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1526– 1535, 2018. 3

    2018

  36. [44]

    Learning from synthetic humans

    Gul Varol, Javier Romero, Xavier Martin, Naureen Mah- mood, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 109–117, 2017. 3

    2017

  37. [45]

    Synthetic humans for action recognition from un- seen viewpoints

    G ¨ul Varol, Ivan Laptev, Cordelia Schmid, and Andrew Zis- serman. Synthetic humans for action recognition from un- seen viewpoints. InIJCV, 2021. 3

    2021

  38. [46]

    Phenaki: Variable length video generation from open domain textual description.arXiv preprint arXiv:2210.02399, 2022

    Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kin- dermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description.arXiv preprint arXiv:2210.02399, 2022. 3

    arXiv

  39. [47]

    Generating videos with scene dynamics

    Carl V ondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. InAdvances in Neu- ral Information Processing Systems, pages 613–621, 2016. 3

    2016

  40. [48]

    Aric: An activity recognition dataset in classroom surveillance images, 2024

    Linfeng Xu, Fanman Meng, Qingbo Wu, Lili Pan, Heqian Qiu, Lanxiao Wang, Kailong Chen, Kanglei Geng, Yilei Qian, Haojie Wang, Shuchang Zhou, Shimou Ling, Zejia 12 Liu, Nanlin Chen, Yingjie Xu, Shaoxu Cheng, Bowen Tan, Ziyong Xu, and Hongliang Li. Aric: An activity recognition d...

    2024

  41. [49]

    Spatial tempo- ral graph convolutional networks for skeleton-based action recognition

    Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial tempo- ral graph convolutional networks for skeleton-based action recognition. InThirty-second AAAI conference on artificial intelligence, 2018. 3

    2018

  42. [50]

    VideoGPT: Video generation using VQ-V AE and transformers.arXiv preprint arXiv:2104.10157, 2021

    Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video generation using VQ-V AE and transformers.arXiv preprint arXiv:2104.10157, 2021. 3

    2021 arXiv

  43. [51]

    Scb-dataset: A dataset for detecting student class- room behavior.arXiv preprint arXiv:2304.02488, 2023

    Fan Yang. Scb-dataset: A dataset for detecting student class- room behavior.arXiv preprint arXiv:2304.02488, 2023. 3 13

    2023 arXiv

  44. [2024]

    Accessed: 2026-05-11. 3

    2026

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.