Pith. sign in

REVIEW 4 major objections 6 minor 47 references

MoExDA: Domain Adaptation for Edge-based Action Recognition

T0 review · 4 major / 6 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 MoExDA claims that aligning RGB and edge feature moments in a two-stream ViT suppresses static bias in action recognition at low computational cost.

desk verdict A clean, honest exploratory study of moment exchange between RGB and edge streams; the core idea is simple and the BOR gains look real, but the headline cost/robustness claims outrun the measurements. read the letter →

arxiv 2508.02981 v1 pith:RLOJET6G submitted 2025-08-05 cs.CV

Takuya Sugimoto , Ning Ding , Toru Tamaki This is my paper
classification cs.CV
keywords actionrecognitionstaticbiasmomentexchangedomainadaptationedgeframesvisiontransformertwo-streamnetworkbackground-onlyratio
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 is trying to establish that a two-stream vision transformer can suppress static bias in action recognition by exchanging feature moments between an RGB stream and an edge stream that is computed online with a cheap edge detector. The proposed MoExDA module takes the intermediate features of the two streams and renormalizes one stream's distribution to the other's mean and variance, treating the RGB stream as the classification-strong but bias-prone partner and the edge stream as the bias-light partner. On a Kinetics50 subset, the authors report that every MoExDA configuration improves the edge stream's top-1 accuracy by 2 to 5 points over the edge-only baseline, and they report reduced Background-Only Ratio relative to baselines, at lower computational cost than the random-dot optical-flow preprocessing used by earlier appearance-free methods. The paper also reports that the method does not fully solve the problem: one IN configuration collapses to 33.74% top-1, and Human-Only Ratio stays below the S3Aug baseline, so the contribution is framed as a lightweight step toward bias suppression.

What carries the argument

The central object is the moment-exchange operation of Eq. (1): $$$h^{{(RGB)l}}$_{\mathrm{edge}} = \frac{h^l_{\mathrm{edge}} - \mu^l_{\mathrm{edge}}}{\$\sigma$^l_{\mathrm{edge}}} \$\sigma$^l_{\mathrm{RGB}} + \mu^l_{\mathrm{RGB}}.$$ It renormalizes one stream's feature tensor with the other stream's mean and standard deviation, computed either per patch over channels (PONO) or per channel over patches (Instance Normalization). The module is inserted in the ViT block between the residual connection after multi-head attention and the LayerNorm of the MLP, and it is the mechanism that transfers classification strength from the RGB stream to the edge stream while reshaping the representation away from static appearance.

What would settle it

Run the same two-stream training on a set of videos whose static backgrounds are deliberately edge-rich (highly textured scenes), then measure Background-Only Ratio with and without the moment exchange; if the exchange does not lower BOR relative to the RGB baseline, the central claim would be falsified.

Watch

Extended reading notes

Core claim

The central discovery is that first- and second-order moment alignment between RGB and edge streams is enough to shift representation away from static appearance. Equation (1) defines the operation: the edge feature distribution is renormalized so that its mean and standard deviation match those of the RGB stream. The authors' evidence is that the edge stream's top-1 accuracy rises 2 to 5 points over an edge-only ViT baseline in every configuration, and that Background-Only Ratio is reduced relative to the baselines they compare against. They interpret this as the edge stream receiving classification-relevant statistics from RGB without inheriting its static bias, while remaining usable with models trained on ordinary RGB frames.

Load-bearing premise

The load-bearing premise is that matching only the first and second moments of the RGB and edge feature distributions is enough to transfer useful, bias-free knowledge to the edge stream without damaging classification; if higher-order statistics or the direction of exchange matter more than the moments, the method loses its grounding.

Editorial extensions

If this is right

  • Action recognition can suppress static bias without expensive random-dot and optical-flow preprocessing, because edge frames are computed on the fly.
  • Models trained on ordinary RGB frames can benefit from an edge stream, unlike approaches that require training on appearance-free random-dot videos.
  • Every reported MoExDA configuration improves the edge stream's top-1 accuracy by 2 to 5 points over the edge-only ViT baseline.
  • PONO-based configurations are more stable than IN-based ones; the authors note IN can degrade severely (edge stream, bidirectional exchange, with stop gradient), reaching 33.74% top-1, and they leave that collapse to future work.
  • Human-Only Ratio remains below the S3Aug baseline, so the method reduces background reliance but has not fully shifted prediction onto human actors; the paper names this trade-off as future work.

Reading between the lines

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

  • Inference: because Eq. (1) matches only mean and variance, the method should be sensitive to cases where the background itself is edge-dense, such as patterned or textured scenes; a testable extension is to combine MoExDA with an explicit foreground-attention loss.
  • Inference: the authors mention estimating moments during training like batch normalization to allow edge-only evaluation at test time; this would make the method directly deployable where RGB frames are unavailable, not just a training-time regularizer.
  • Inference: the same two-stream moment-exchange recipe could transfer to other modalities that are less biased than RGB, such as depth or optical flow, whenever one stream is appearance-light and the other carries scene bias.
  • Inference: the observed IN collapse (edge stream at 33.74% top-1 under bidirectional exchange with stop gradient) suggests gradient flow through moment statistics can destabilize training; a robustness test would repeat the sweep with several seeds to see whether that collapse is deterministic or a training instability.
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 / 6 minor

Summary. The paper proposes MoExDA, a two-stream ViT architecture for action recognition that operates on both RGB frames and Sobel edge frames. The two streams are coupled by exchanging feature moments (mean and standard deviation) at intermediate layers, following the MoEx idea, with design choices over normalization type (IN vs PONO), exchange direction, and stop-gradient. Experiments are conducted on a self-constructed Kinetics50 subset derived from Mimetics classes. The paper reports top-1 accuracy, Background-Only Ratio (BOR), and Human-Only Ratio (HOR), and claims that MoExDA suppresses static bias at lower computational cost than prior appearance-free methods such as AFD.

Significance. If the claimed BOR reductions are real and the method is indeed cheaper than AFD, the paper would make a small but useful contribution: a simple, plug-in moment-exchange mechanism for edge-based action recognition that reduces background reliance. The ablations over normalization, exchange direction, and stop-gradient are informative, and the use of BOR/HOR metrics is appropriate for the stated goal. However, the current evidence does not establish the comparative claims: AFD is never evaluated, no run-time or FLOPs measurements are reported, and no error bars or repeated runs are provided. The central mechanism is borrowed from MoEx without analysis of when moment matching should transfer bias-free knowledge, and one configuration collapses to 33.74% top-1, indicating sensitivity to hyperparameters. The contribution is potentially sound but needs stronger empirical validation before the comparative abstract claims can be accepted.

major comments (4)
  1. [§4, Table 1] The paper provides no error bars, multiple seeds, or repeated runs for any of the reported numbers. Given that the claimed improvements are 2–5% in top-1 accuracy and a few points in BOR, these differences may be within run-to-run variance, especially since Table 1 shows extreme sensitivity to configuration (e.g., IN Bidirection with stop-gradient collapses to 33.74% top-1 for the edge stream and 32.45% for the RGB stream). Please report mean and standard deviation over at least three seeds, or otherwise justify that the reported differences are stable.
  2. [§1, §4, Abstract] The abstract's comparative claim—'lower computational cost' and 'more robust action recognition than previous approaches'—is not supported by the experiments. AFD [11], the approach identified in the introduction as the expensive alternative, is never run on the same split, and no FLOPs, parameter counts, or wall-clock times are provided. Since MoExDA trains and evaluates two full ViT encoders, its inference cost is likely about twice a single-stream model, so the 'lower computational cost' claim is unverified and possibly false. Either add an AFD comparison with cost measurements, or remove/qualify the comparative claims.
  3. [§4, Table 1] The BOR/HOR evidence for 'suppressing static bias' is inconsistent and the table formatting makes it hard to evaluate. The baseline rows contain only three numeric entries while MoExDA rows contain four, and the text says 'the lowest BOR falls below 70%' but the only non-collapsed configuration below 70% is PONO RGB-to-edge with stop-gradient (69.48), while most MoExDA configurations have BORs between 73 and 81, only a few points below the RGB baseline's 83.66. Clarify the table structure, report the full distribution of BOR across configurations, and provide error bars before claiming a 'superior BOR'.
  4. [§3.1, Eq. (1)] The edge-stream normalization statistics are computed from a single Kinetics class ('abseiling') rather than from the training set or dataset statistics, and no sensitivity analysis is performed. Since Eq. (1) directly uses these moments for the exchange, the method's behavior may depend on this arbitrary choice. Please report results with alternative statistics (e.g., computed on the training split of Kinetics50, or using standard ImageNet statistics for both streams) to show robustness.
minor comments (6)
  1. [§3.2] The notation 'h(RGB)l_edge' and 'hl_RGB, hl_edge' is difficult to read; use consistent superscripts for layer index and clear subscripts for the stream, e.g., h_{edge}^{(l;RGB)}.
  2. [Table 1] The table header should explicitly state that '1 top-1' and '1~12 top-1' apply only to MoExDA rows, and that baseline rows report top-1, BOR, and HOR; the current layout is ambiguous.
  3. [§4] The text says 'IN tends to perform better with the RGB-to-edge direction', but in Table 1 the IN edge-to-RGB rows achieve higher top-1 for the RGB stream (e.g., 82.21 and 82.37) than the RGB-to-edge rows (78-79). Please check whether the direction labels in the text or the table are reversed, and make the description consistent.
  4. [§5] The future-work statement that moment estimation could 'allow the use of only the edge stream during evaluation' implies that the current method requires both streams at test time; this should be stated explicitly in §3 because it is relevant to the computational-cost claim.
  5. [Throughout] Minor language issues: 'ordinal RGB frames' should be 'ordinary RGB frames', 'an ordinal' should be 'an ordinary', and 'contains less static appearance' should be 'contains less static appearance information'.
  6. [§4] The loss weights alpha_RGB = 0.5 and alpha_edge = 1.0 are chosen from preliminary experiments; please report a small sensitivity study or at least state the range over which conclusions are unchanged.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the reported BOR/HOR are measured on held-out validation and are not algebraically determined by Eq. (1), Eq. (2), or the edge normalization statistics.

full rationale

The paper makes no first-principles derivation; its contribution is an empirical application of the MoEx moment exchange (Eq. 1, explicitly cited to [13]) to RGB/edge streams. The claimed bias suppression is supported by BOR/HOR values measured on validation data, not by the fitted quantities. The loss weights in Eq. (2) are selected from preliminary experiments, but the final metrics are not equal to or constructed from these weights; this is ordinary hyperparameter selection. The edge normalization statistics are computed from Kinetics400 abseiling training frames, but they are fixed input pre-processing scalars and do not by themselves determine the BOR/HOR outcomes. The only self-citations are related-work and baseline items ([7], [25], [42]); none carries the argument. The abstract's 'lower computational cost' and 'more robust than previous approaches' claims are unmeasured, but lack of evidence is a correctness gap, not circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper adds no invented entities. It depends on three domain assumptions (edge maps carry action information, moment alignment transfers useful knowledge, BOR/HOR measure static bias) and on two fitted quantities (loss weights and edge normalization statistics), both chosen empirically rather than derived.

free parameters (3)
  • Loss weights alpha_RGB, alpha_edge = alpha_RGB=0.5, alpha_edge=1.0
    Set from preliminary experiments (Eq. 2); controls balance between RGB and edge classification losses.
  • Edge-frame normalization statistics (mu, sigma) = (0.026, 0.037)
    Computed from all frames of the 'abseiling' class in Kinetics400, not from the evaluation distribution; used to normalize all edge inputs (§3.1).
  • MoExDA hyper-configuration = IN/PONO, direction, stop-gradient, layer placement selected per row
    The paper presents an ablation and selects favorable settings; the final recommendation is edge-to-RGB with PONO for the RGB stream, but no single configuration is derived.
assumptions (3)
  • domain assumption Edge frames computed by Sobel filtering retain enough action-discriminative information while suppressing static appearance.
    The entire two-stream design depends on this; §1 and §3.1 assert it, and the edge-only baseline (68.84% top-1) is the supporting evidence rather than an external proof.
  • domain assumption Aligning the first and second moments of intermediate features between RGB and edge streams reduces harmful domain shift and improves bias resistance.
    Borrowed from MoEx [13] and normalization literature; Eq. (1) is applied without a formal justification, and results vary strongly across normalization choices.
  • domain assumption BOR and HOR from [5] are valid proxies for static bias.
    The paper uses these metrics to support the central claim but does not analyze their sensitivity or definition; they are taken from Chung et al. [5].

how reviews work

0 comments
Cite this review

Pith. "Pith review of MoExDA: Domain Adaptation for Edge-based Action Recognition." pith.science (2026). https://pith.science/paper/RLOJET6G

@misc{pith2026250802981,
  author       = {Pith},
  title        = {Pith review of: MoExDA: Domain Adaptation for Edge-based Action Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RLOJET6G}},
  note         = {Machine review of arXiv:2508.02981}
}
read the original abstract

Modern action recognition models suffer from static bias, leading to reduced generalization performance. In this paper, we propose MoExDA, a lightweight domain adaptation between RGB and edge information using edge frames in addition to RGB frames to counter the static bias issue. Experiments demonstrate that the proposed method effectively suppresses static bias with a lower computational cost, allowing for more robust action recognition than previous approaches.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 42 canonical work pages

  1. [11]

    Is appearance free action recognition possible?

    F. Ilic, T. Pock, and R. P. Wildes, “Is appearance free action recognition possible?” in Computer Vision – ECCV 2022, S. Avidan, G. Brostow, M. Ciss´ e, G. M. Farinella, and T. Hassner, Eds. Cham: Springer Na- ture Switzerland, 2022, pp. 156–173

    work page 2022

  2. [1]

    Human action recognition and prediction: A survey,

    Y. Kong and Y. Fu, “Human action recognition and prediction: A survey,” International Journal of Computer Vision, vol. 130, no. 5, pp. 1366–1401, may

    work page

  3. [2]

    Human action recognition and prediction: A survey,

    ——, “Human action recognition and prediction: A survey,” International Journal of Computer Vision, vol. 130, no. 5, pp. 1366–1401, 2022

    work page 2022

  4. [3]

    Video ac- tion understanding,

    M. S. Hutchinson and V. N. Gadepally, “Video ac- tion understanding,” IEEE Access, vol. 9, pp. 134 611– 134 637, 2021

    work page 2021

  5. [4]

    Can spatiotem- poral 3d cnns retrace the history of 2d cnns and im- agenet?

    K. Hara, H. Kataoka, and Y. Satoh, “Can spatiotem- poral 3d cnns retrace the history of 2d cnns and im- agenet?” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    work page 2018

  6. [5]

    Enabling detailed action recognition evaluation through video dataset augmentation,

    J. Chung, Y. Wu, and O. Russakovsky, “Enabling detailed action recognition evaluation through video dataset augmentation,” in Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track , 2022. [Online]. Available: https://openreview.net/forum?id=eOnQ2etkxto

    work page 2022

  7. [6]

    Resound: Towards action recognition without representation bias,

    Y. Li, Y. Li, and N. Vasconcelos, “Resound: Towards action recognition without representation bias,” inPro- ceedings of the European Conference on Computer Vi- sion (ECCV), September 2018

    work page 2018

  8. [7]

    Can masking background and object reduce static bias for zero-shot action recognition?

    T. Fukuzawa, K. Hara, H. Kataoka, and T. Tamaki, “Can masking background and object reduce static bias for zero-shot action recognition?” in MultiMe- dia Modeling, I. Ide, I. Kompatsiaris, C. Xu, K. Yanai, W.-T. Chu, N. Nitta, M. Riegler, and T. Yamasaki, Eds. Singapore: Springer Nature Singapore, 2025, pp. 366–379

    work page 2025

Show all 47 references
  1. [8]

    Hu- man action recognition without human,

    Y. He, S. Shirakabe, Y. Satoh, and H. Kataoka, “Hu- man action recognition without human,” in Computer Vision – ECCV 2016 Workshops, G. Hua and H. J´ egou, Eds. Cham: Springer International Publishing, 2016, pp. 11–17

    2016

  2. [9]

    What do 15,000 object categories tell us about classifying and localizing actions?

    M. Jain, J. C. van Gemert, and C. G. M. Snoek, “What do 15,000 object categories tell us about classifying and localizing actions?” in Proceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), June 2015

    2015

  3. [10]

    Harnessing object and scene semantics for large-scale video under- standing,

    Z. Wu, Y. Fu, Y.-G. Jiang, and L. Sigal, “Harnessing object and scene semantics for large-scale video under- standing,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016

    2016

  4. [12]

    An image is worth 16x16 words: Transformers for image recogni- tion at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weis- senborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszko- reit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recogni- tion at scale,” in International Conference on Le...

    2021

  5. [13]

    On feature normalization and data augmenta- tion,

    B. Li, F. Wu, S.-N. Lim, S. Belongie, and K. Q. Wein- berger, “On feature normalization and data augmenta- tion,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021, pp. 12 383–12 392

    2021

  6. [14]

    Quo vadis, action recog- nition? a new model and the kinetics dataset,

    J. Carreira and A. Zisserman, “Quo vadis, action recog- nition? a new model and the kinetics dataset,” in Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017

    2017

  7. [15]

    X3d: Expanding architectures for efficient video recognition,

    C. Feichtenhofer, “X3d: Expanding architectures for efficient video recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), June 2020. [Online]. Avail- able: https://openaccess.thecvf.com/content CVPR 2020/html/Feichtenhofer...

    2020

  8. [16]

    Slow- fast networks for video recognition,

    C. Feichtenhofer, H. Fan, J. Malik, and K. He, “Slow- fast networks for video recognition,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019

    2019

  9. [17]

    Recent advances in video action recognition with 3d convolutions,

    K. Hara, “Recent advances in video action recognition with 3d convolutions,” IEICE Transactions on Fun- damentals of Electronics, Communications and Com- puter Sciences, vol. E104.A, no. 6, pp. 846–856, 2021

    2021

  10. [18]

    Vivit: A video vision transformer,

    A. Arnab, M. Dehghani, G. Heigold, C. Sun, M. Luˇ ci´ c, and C. Schmid, “Vivit: A video vision transformer,” in Proceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), October 2021, pp. 6836–6846

    2021

  11. [19]

    Is space- time attention all you need for video understanding?

    G. Bertasius, H. Wang, and L. Torresani, “Is space- time attention all you need for video understanding?” in Proceedings of the 38th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, M. Meila and T. Zhang, Eds., vol. 139. PMLR, 18–24 ...

    2021

  12. [20]

    Video transform- ers: A survey,

    J. Selva, A. S. Johansen, S. Escalera, K. Nasrollahi, T. B. Moeslund, and A. Clapes, “Video transform- ers: A survey,” IEEE Transactions on Pattern Anal- ysis & Machine Intelligence, vol. 45, no. 11, pp. 12 922–12 943, nov 2023

    2023

  13. [21]

    Video swin transformer,

    Z. Liu, J. Ning, Y. Cao, Y. Wei, Z. Zhang, S. Lin, and H. Hu, “Video swin transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 3202– 3211

    2022

  14. [22]

    VideoCLIP: Contrastive pre- training for zero-shot video-text understanding,

    H. Xu, G. Ghosh, P.-Y. Huang, D. Okhonko, A. Aghajanyan, F. Metze, L. Zettlemoyer, and C. Feichtenhofer, “VideoCLIP: Contrastive pre- training for zero-shot video-text understanding,” in Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. On...

    2021

  15. [23]

    Actionclip: A new paradigm for video action recognition,

    M. Wang, J. Xing, and Y. Liu, “Actionclip: A new paradigm for video action recognition,” CoRR, vol. abs/2109.08472, 2021. [Online]. Available: https://arxiv.org/abs/2109.08472

    2021 arXiv

  16. [24]

    Fine-tuned clip models are efficient video learners,

    H. Rasheed, M. U. Khattak, M. Maaz, S. Khan, and F. S. Khan, “Fine-tuned clip models are efficient video learners,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 6545–6554

    2023

  17. [25]

    S3aug: Segmenta- tion, sampling, and shift for action recognition,

    T. Sugiura and T. Tamaki, “S3aug: Segmenta- tion, sampling, and shift for action recognition,” in Proceedings of the 19th International Joint Confer- ence on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 2: VISAPP, INSTICC. SciTePress, 2024, p...

    2024 doi

  18. [26]

    Mitigating and evaluating static bias of action representations in the background and the foreground,

    H. Li, Y. Liu, H. Zhang, and B. Li, “Mitigating and evaluating static bias of action representations in the background and the foreground,” in Proceedings of the IEEE/CVF International Conference on Computer Vi- sion (ICCV), October 2023, pp. 19 911–19 923

    2023

  19. [27]

    Removing the background by adding the background: Towards back- ground robust self-supervised video representation learn- ing,

    J. Wang, Y. Gao, K. Li, Y. Lin, A. J. Ma, H. Cheng, P. Peng, F. Huang, R. Ji, and X. Sun, “Removing the background by adding the background: Towards back- ground robust self-supervised video representation learn- ing,” in Proceedings of the IEEE/CVF Conference on Computer Visi...

    2021

  20. [28]

    Why can’t I dance in the mall? learn- ing to mitigate scene bias in action recognition,

    J. Choi, C. Gao, J. C. E. Messou, and J.-B. Huang, “Why can’t I dance in the mall? learn- ing to mitigate scene bias in action recognition,” in Advances in Neural Information Processing Sys- tems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d 'Alch´ e-Buc, E. Fox, and R. Gar...

    2019

  21. [29]

    Devias: Learn- ing disentangled video representations of action and scene,

    K. Bae, G. Ahn, Y. Kim, and J. Choi, “Devias: Learn- ing disentangled video representations of action and scene,” in Computer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds. Cham: Springer Nature Switzerland, 2025, pp. 431–448

    2024

  22. [30]

    Mitigating representation bias in action recognition: Algorithms and benchmarks,

    H. Duan, Y. Zhao, K. Chen, Y. Xiong, and D. Lin, “Mitigating representation bias in action recognition: Algorithms and benchmarks,” in Computer Vision – ECCV 2022 Workshops, L. Karlinsky, T. Michaeli, and K. Nishino, Eds. Cham: Springer Nature Switzerland, 2023, pp. 557–575

    2022

  23. [31]

    UCF101: A dataset of 101 human actions classes from videos in the wild,

    K. Soomro, A. R. Zamir, and M. Shah, “UCF101: A dataset of 101 human actions classes from videos in the wild,” CoRR, vol. abs/1212.0402, 2012. [Online]. Available: http://arxiv.org/abs/1212.0402

    2012 arXiv

  24. [32]

    Action recognition using edge trajectories and motion acceleration descriptor,

    X. Wang and C. Qi, “Action recognition using edge trajectories and motion acceleration descriptor,” Ma- chine Vision and Applications, vol. 27, no. 6, pp. 861– 875, 2016

    2016

  25. [33]

    An edge-based ap- proach to motion detection,

    A. D. Sappa and F. Dornaika, “An edge-based ap- proach to motion detection,” inComputational Science – ICCS 2006, V. N. Alexandrov, G. D. van Albada, P. M. A. Sloot, and J. Dongarra, Eds. Berlin, Heidel- berg: Springer Berlin Heidelberg, 2006, pp. 563–570

    2006

  26. [34]

    A sketch-based approach for detecting com- mon human actions,

    E. A. Suma, C. W. Sinclair, J. Babbs, and R. Sou- venir, “A sketch-based approach for detecting com- mon human actions,” in Advances in Visual Comput- ing, G. Bebis, R. Boyle, B. Parvin, D. Koracin, P. Re- magnino, F. Porikli, J. Peters, J. Klosowski, L. Arns, Y. K. Chun, T.-M...

    2008

  27. [35]

    Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,

    S. Ioffe and C. Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,” in Proceedings of the 32nd International Conference on Machine Learning. PMLR, Jun. 2015, pp. 448–456, batch Normalization BN. [Online]. Available: https://...

    2015

  28. [36]

    Instance normalization: The missing ingredient for fast stylization,

    D. Ulyanov, A. Vedaldi, and V. Lempitsky, “Instance normalization: The missing ingredient for fast stylization,” 2017. [Online]. Available: https://arxiv. org/abs/1607.08022

    2017 arXiv

  29. [37]

    Group Normalization,

    Y. Wu and K. He, “Group Normalization,” in Computer Vision – ECCV 2018, V. Ferrari, M. Hebert, C. Sminchisescu, and Y. Weiss, Eds. Cham: Springer International Publishing, 2018, pp. 3–

    2018

  30. [38]

    Arbitrary style transfer in real-time with adaptive instance normalization,

    X. Huang and S. Belongie, “Arbitrary style transfer in real-time with adaptive instance normalization,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017

    2017

  31. [39]

    Available: https://link.springer.com/ chapter/10.1007/978-3-030-01261-8 1

    [Online]. Available: https://link.springer.com/ chapter/10.1007/978-3-030-01261-8 1

    doi

  32. [40]

    mixup: Beyond empirical risk minimization,

    H. Zhang, M. Ciss´ e, Y. N. Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical risk minimization,” in 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings . OpenReview.net, 2018. [Onli...

    2018

  33. [41]

    Po- sitional normalization,

    B. Li, F. Wu, K. Q. Weinberger, and S. Belongie, “Po- sitional normalization,” in Advances in Neural Infor- mation Processing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d 'Alch´ e-Buc, E. Fox, and R. Gar- nett, Eds., vol. 32. Curran Associates, Inc., 2019. [Online]...

    2019

  34. [42]

    Objectmix: Data augmentation by copy-pasting objects in videos for action recognition,

    J. Kimata, T. Nitta, and T. Tamaki, “Objectmix: Data augmentation by copy-pasting objects in videos for action recognition,” in Proceedings of the 4th ACM International Conference on Multimedia in Asia, ser. MMAsia ’22. New York, NY, USA: Association for Computing Machinery, 2...

    2022

  35. [43]

    Videomix: Rethinking data augmentation for video classification,

    S. Yun, S. J. Oh, B. Heo, D. Han, and J. Kim, “Videomix: Rethinking data augmentation for video classification,” 2020. [Online]. Available: https: //arxiv.org/abs/2012.03457

    2020 arXiv

  36. [44]

    The kinetics human action video dataset,

    W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev, M. Suleyman, and A. Zisserman, “The kinetics human action video dataset,” CoRR, vol. abs/1705.06950, 2017. [Online]. Available: http: //arxiv.org/abs/1705.06950

    2017 arXiv

  37. [45]

    Imagenet: A large-scale hierarchical im- age database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical im- age database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255

    2009

  38. [47]

    Mimetics: Towards understanding human actions out of context,

    P. Weinzaepfel and G. Rogez, “Mimetics: Towards understanding human actions out of context,” In- ternational Journal of Computer Vision , vol. 129, no. 5, pp. 1675–1690, 2021. [Online]. Available: https://doi.org/10.1007/s11263-021-01446-y

    2021 doi

  39. [2022]

    Available: https://doi.org/10.1007/ s11263-022-01594-9

    [Online]. Available: https://doi.org/10.1007/ s11263-022-01594-9

Pith tools

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