Pith. sign in

REVIEW 3 major objections 5 minor 51 references

CA3D: Convolutional-Attentional 3D Nets for Efficient Video Activity Recognition on the Edge

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A 7-million-parameter conv-attention network, trained entirely in float16, reports higher test accuracy than larger baselines on UCF101, HMDB51, and Kinetics400 under a no-pretraining protocol.

desk verdict A well-controlled empirical architecture paper whose quantization method, as written, has a chain-rule error and a range contradiction that make the headline float16 results unreproducible. read the letter →

arxiv 2505.19928 v1 pith:XA7RLCTC submitted 2025-05-26 cs.CV

Gabriele Lagani , Fabrizio Falchi , Claudio Gennaro , Giuseppe Amato This is my paper
classification cs.CV
keywords videoactivityrecognitionconvolutionalneuralnetworksattentionmechanismquantizationfloat16trainingedgecomputingefficientdeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that a small video activity recognition network can combine the strengths of convolutional and attention-based designs: spatial convolutions provide locality and translation invariance, while temporal attention over local windows models long-range frame relationships. The proposed CA3D model, built from CAST blocks, has about 7 million parameters and is trained from raw RGB frames without external pretraining or optical flow. The paper reports test accuracies of 94.8% on UCF101, 63.2% on HMDB51, and 52.1% on Kinetics400 for the larger CA3D model, improving over compared CNN and Transformer baselines under the same single-crop, no-pretraining protocol. It also introduces a float16 quantization scheme based on a pre-parameter mapping, so training itself runs at reduced precision without the numerical instability that normally forces float32 training. A sympathetic reader would care because accuracy at this compute level is what makes on-device video understanding practical for smart home or healthcare settings where data should stay local.

What carries the argument

The load-bearing object is the Convolutional-Attentional Spatio-Temporal (CAST) block: spatial convolutional layers with residual columns extract visual features, while a temporal multi-head self-attention module restricted to local windows centered on each token propagates information across frames with linear complexity, and a second residual column replaces the MLP of a transformer. The quantization mechanism maps network weights w from pre-parameters θ through w = θ/T, with gradients scaled back by T, so optimization happens in a space with an effectively wider representable range and float16 training stays stable. The paper argues this combination gives convolutional inductive bias plus global temporal modeling without quadratic attention or external pretraining.

What would settle it

Retrain the main baselines, for example R2+1D-R18, X3D-XL, STAM-B, and TubeViT-B, under the same hyperparameter search budget, schedule, augmentation, and single-crop evaluation as CA3D, in float32 and with the same float16 mapping; if the accuracy gaps on UCF101, HMDB51, and Kinetics400 shrink below statistical significance or reverse, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that the CA3D architecture, which alternates spatial convolutions with a linear-complexity temporal attention mechanism inside CAST blocks, improves test accuracy over the compared models while keeping compute and memory low. On UCF101, CA3D reaches 94.8% under both QAT and the proposed float16 training; on HMDB51, float16 reaches 63.2%; on Kinetics400, the larger CA3D-L reaches 52.1%. The paper reports 7 million parameters, 6.3 GFLOPs per 16-frame clip, 4.6 GB training memory for batches of 20 clips, and a training throughput of 500 frames per second, figures it interprets as making the model suitable for consumer and edge hardware. The paper claims that attention only along the temporal dimension, with convolutions handling space, is enough to match or beat models that use spatiotemporal attention, and that the proposed quantization does not hurt and sometimes helps generalization.

Load-bearing premise

The load-bearing premise is that the comparison baselines were tuned with the same effort as CA3D under the no-pretraining, single-crop protocol; if some baselines were undertuned for that regime, CA3D's reported accuracy advantage would be inflated.

Editorial extensions

If this is right

  • If CA3D's reported numbers hold, video activity recognition can be trained and run on commodity hardware, including the fine-tuning step, because the whole pipeline operates in float16.
  • Global spatial attention appears unnecessary when convolutions already capture spatial structure; temporal-only local attention is sufficient to match or beat spatiotemporal attention models.
  • The pre-parameter mapping opens a route to low-precision training regimes for other architectures, since it decouples optimization stability from the reduced-precision weight representation.
  • The no-pretraining, single-crop evaluation protocol exposes generalization differences between CNN-based and Transformer-based video models that are hidden when models are pretrained on large image datasets.
  • A 7-million-parameter model at 6.3 GFLOPs per clip provides a concrete compute target for privacy-sensitive edge applications.

Reading between the lines

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

  • The reported accuracy gaps may depend on how thoroughly the baseline models were tuned for the unusual no-pretraining, single-crop protocol; the paper gives no per-model search budgets, so an equal-tuning rerun is the natural stress test.
  • The quantization-induced improvements in some settings suggest that the pre-parameter mapping is not only a numerical fix but could act as a regularizer; one testable extension is applying it intentionally at higher noise levels.
  • The same local-window temporal attention idea could transfer to other spatiotemporal tasks, such as online action segmentation or gesture recognition, where long-range temporal structure matters but full attention is too costly.
  • If float16-only training generalizes, it could reduce the hardware and energy cost of video model development, since float32 pretraining and QAT's float32 master weights would become unnecessary.
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

3 major / 5 minor

Summary. The paper introduces CA3D, a video activity recognition network built from Convolutional-Attentional Spatio-Temporal (CAST) blocks, combining spatial convolutions with temporal local-window attention. A second contribution is a float16 quantization scheme in which network weights w are generated from pre-parameters theta through a mapping w = theta / T, with the goal of running both training and inference without float32 representations. The model is evaluated on UCF101, HMDB51, and Kinetics400 under a common protocol using only raw RGB frames, no external pretraining, and single-crop testing. The authors report accuracy of 94.8% on UCF101, 63.2% on HMDB51, and 52.1% on Kinetics400 with CA3D-L, together with a lower parameter count and GFLOPs than the compared baselines.

Significance. If the technical description is corrected, the paper has clear strengths: a compact ~7M-parameter architecture, a no-external-pretraining protocol that is closer to edge deployment than typical video recognition evaluations, a linear-complexity attention variant, and public code. The reported accuracy-efficiency trade-off on standard benchmarks is a useful data point for resource-constrained video activity recognition. However, the significance is currently limited by an internal inconsistency in the central quantization equations, which makes the headline 'Ours(float16)' results not reproducible from the method description as written, and by insufficient reporting of the baseline tuning effort, which is load-bearing for the comparative accuracy claim.

major comments (3)
  1. [§3.2, Eqs. (1)–(2)] The quantization mapping is internally inconsistent as written. If w = f(theta) = theta / T, then by the chain rule dL/dtheta = (1/T) * dL/dw, so the backward rule grad_theta = T * grad_w in Eq. (2) is not the true gradient; it is the correct rule for the mapping w = theta * T. Moreover, if theta is represented in float16 with maximum 32768 and T = 0.1, the output range stated for w is [1/3276.8, 327680], whose upper endpoint is not representable in float16, so the claim that both phases run entirely in float16 is not reproducible. If the intended mapping is instead w = theta * T, then Eq. (1) and the range-shift sentence are wrong, and the representable range becomes [1/327680, 3276.8]. Since the headline 'Ours(float16)' results in Tables 1 and 2 are produced under this mechanism, the method section must be corrected and the actual mapping and gradient rule used in the released code must be stated unambiguously.
  2. [§4, Tables 1–2] The comparative accuracy claim depends on training all baselines with comparable tuning effort under the same protocol, but the paper reports no hyperparameter search budgets, schedules, or per-model final configurations. The statement that hyperparameters were 'optimized for the float32 quantization regime' and that 'the same training conditions were maintained' is not sufficient to establish equal footing, especially for Transformer-based baselines trained from scratch on small datasets where learning rate and warmup are critical. Please report the tuning budget per model (number of trials, epochs, final hyperparameters) or otherwise provide evidence that the reported accuracy advantage is not a tuning artifact.
  3. [§4, Table 3] Table 3 is used to support the edge-suitability claim, but the measurement protocol is not described: the hardware, batch size, precision (float16 versus float32), and whether the memory and throughput numbers correspond to the quantized regime used in Tables 1–2 are all unspecified. Without this information, the 'training memory footprint' and 'frames/s' columns cannot be meaningfully compared across models, and the reader cannot verify the efficiency advantage attributed to the quantization mechanism.
minor comments (5)
  1. [§1] The dataset name is misspelled as 'HBDM51'; it should be 'HMDB51'.
  2. [Throughout] The text repeatedly refers to the Supplementary Material for hyperparameter details, the experimental protocol, and ablations, but this material is not included in the arXiv version; it must be provided for the reported results to be fully assessable.
  3. [§3.2, Eq. (1)] The paper says 'if, for example, we set T = 0.1', but never states the actual value of T used in the experiments; the exact value should be specified.
  4. [§4, Kinetics400] The single-crop testing protocol should be defined precisely (e.g., a single center crop per clip) to avoid ambiguity alongside the description of the 10-LeftCenterRight strategy that was not used.
  5. [Abstract / Related Work] The linear-complexity claim should be made precise with respect to sequence length and window size; as written, linear complexity is plausible for fixed local windows but the definition of the window and the complexity in terms of T, H, and W should be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the CA3D results are empirical comparisons against external baselines, with the only self-citations appearing in future-work context and not supporting the central claims.

full rationale

The paper's central claims are an architecture (CA3D/CAST) and a quantization mechanism, validated by experiments on UCF101, HMDB51, and Kinetics400 against externally published models (R3D, R2+1D, I3D, X3D, ViViT, Swin3D, TubeViT, etc.). The accuracy results are measured, not derived from the method's assumptions; there is no equation that defines the reported accuracy in terms of the model's own fitted quantities. The quantization mapping in Eq. (1)-(2) is a reparameterization of weights via theta/T; while the gradient formula as written is internally inconsistent (for w=theta/T the chain rule gives grad_theta = grad_w/T, not T*grad_w), this is a mathematical correctness issue, not a circularity. The only self-citations are references [27,28] in the future-work paragraph about neuromorphic devices; they are not load-bearing for any experimental claim or architectural derivation. No fitted parameter is renamed as a prediction, no uniqueness result is imported from the authors' prior work, and no known result is repackaged as new solely by renaming. The comparison protocol is described as common across all methods, with no pre-training and single-crop testing, which makes the empirical comparison potentially under-tuned for baselines but not circular. Therefore, the paper is self-contained against external benchmarks for its main claims, and no circular step can be exhibited from the text.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claims rest on standard deep-learning assumptions about convolutions and attention, plus a hand-chosen scaling constant T for the quantization mechanism. No new physical or conceptual entities are introduced.

free parameters (1)
  • T (quantization scaling constant) = 0.1 (example; exact value not specified)
    Chosen by hand in Eq. (1) to shift the representable range; no tuning procedure or per-layer adaptation is described.
assumptions (3)
  • domain assumption BatchNorm in attention blocks stabilizes training better than LayerNorm, as claimed in Section 3.1
    Reported as an empirical observation without comparative experiments or theoretical justification.
  • domain assumption Local-window attention with one window per token provides sufficient global information propagation over successive layers
    Borrowed from the Swin Transformer rationale; the paper does not provide a separate proof or ablation for this variant.
  • domain assumption Training each baseline with optimized float32 hyperparameters is a fair comparison for the reported tasks
    The fairness of cross-model comparisons depends on equal tuning effort and schedules, which are not detailed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CA3D: Convolutional-Attentional 3D Nets for Efficient Video Activity Recognition on the Edge." pith.science (2026). https://pith.science/paper/XA7RLCTC

@misc{pith2026250519928,
  author       = {Pith},
  title        = {Pith review of: CA3D: Convolutional-Attentional 3D Nets for Efficient Video Activity Recognition on the Edge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XA7RLCTC}},
  note         = {Machine review of arXiv:2505.19928}
}
read the original abstract

In this paper, we introduce a deep learning solution for video activity recognition that leverages an innovative combination of convolutional layers with a linear-complexity attention mechanism. Moreover, we introduce a novel quantization mechanism to further improve the efficiency of our model during both training and inference. Our model maintains a reduced computational cost, while preserving robust learning and generalization capabilities. Our approach addresses the issues related to the high computing requirements of current models, with the goal of achieving competitive accuracy on consumer and edge devices, enabling smart home and smart healthcare applications where efficiency and privacy issues are of concern. We experimentally validate our model on different established and publicly available video activity recognition benchmarks, improving accuracy over alternative models at a competitive computing cost.

Figures

Figures reproduced from arXiv: 2505.19928 by the authors.

Figure 1
Figure 1. Design of the Convolutional-Attentional 3D (CA3D) neural network for video processing, as a series of Convolutional-Attentional Spatio-Temporal (CAST) blocks. In CAST blocks, convolutional layers are alternated with attention layers, to take advan￾tage of both types of processing. Convolutions are applied along the spatial dimensions, while attention aggregates global information from different frames along the temp… view at source ↗
Figure 2
Figure 2. Structure of the CA3D architecture in terms of CAST blocks. Blocks are formed by spatial and temporal processing parts. The spatial part is implemented in terms of convolutional blocks, followed by a column of residually connected layers. The tempo￾ral part is mediated by an attention operator, which is followed again by a column of residual layers. Pooling layers shrink the size of the tensor along the temporal dim… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 25 canonical work pages

  1. [1]

    In: ICCV

    Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., Schmid, C.: Vivit: A video vision transformer. In: ICCV. pp. 6836–6846 (2021)

    2021

  2. [2]

    arXiv preprint arXiv:1607.06450 (2016)

    Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)

    arXiv 2016

  3. [3]

    In: International Conference of the Italian Association for Artificial Intelligence

    Badar, A., Varma, A., Staniec, A., Gamal, M., Magdy, O., Iqbal, H., Arani, E., Zonooz, B.: Highlighting the importance of reducing research bias and carbon emis- sions in cnns. In: International Conference of the Italian Association for Artificial Intelligence. pp. 515–531. Springer (2021)

    work page 2021

  4. [4]

    NeurIPS32(2019)

    Banner, R., Nahshan, Y., Soudry, D.: Post training 4-bit quantization of convolu- tional networks for rapid-deployment. NeurIPS32(2019)

    work page 2019

  5. [5]

    Bertasius, G., Wang, H., Torresani, L.: Is space-time attention all you need for video understanding? In: ICML. vol. 2, p. 4 (2021)

    2021

  6. [6]

    In: ECCV

    Chen, Y., Kalantidis, Y., Li, J., Yan, S., Feng, J.: Multi-fiber networks for video recognition. In: ECCV. pp. 352–367 (2018)

    work page 2018

  7. [7]

    arXiv preprint arXiv:1904.10509 (2019)

    Child, R., Gray, S., Radford, A., Sutskever, I.: Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509 (2019)

    arXiv 2019

  8. [8]

    arXiv preprint arXiv:1805.06085 (2018)

    Choi, J., Wang, Z., Venkataramani, S., Chuang, P.I.J., Srinivasan, V., Gopalakr- ishnan, K.: Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085 (2018)

    arXiv 2018

Show all 51 references
  1. [9]

    arXiv preprint arXiv:2009.14794 (2020)

    Choromanski, K., Likhosherstov, V., Dohan, D., Song, X., Gane, A., Sarlos, T., Hawkins, P., Davis, J., Mohiuddin, A., Kaiser, L., et al.: Rethinking attention with performers. arXiv preprint arXiv:2009.14794 (2020)

    2020 arXiv

  2. [10]

    In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW)

    Choukroun, Y., Kravchik, E., Yang, F., Kisilev, P.: Low-bit quantization of neural networks for efficient inference. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3009–3018. IEEE (2019)

    2019

  3. [11]

    In: CVPR

    Crasto, N., Weinzaepfel, P., Alahari, K., Schmid, C.: Mars: Motion-augmented rgb stream for action recognition. In: CVPR. pp. 7882–7891 (2019)

    2019

  4. [12]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  5. [13]

    In: CVPR

    Fan, L., Huang, W., Gan, C., Ermon, S., Gong, B., Huang, J.: End-to-end learning of motion representation for video understanding. In: CVPR. pp. 6016–6025 (2018)

    2018

  6. [14]

    In: CVPR

    Feichtenhofer, C.: X3d: Expanding architectures for efficient video recognition. In: CVPR. pp. 203–213 (2020)

    2020

  7. [15]

    In: ICCV

    Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recogni- tion. In: ICCV. pp. 6202–6211 (2019)

    2019

  8. [16]

    arXiv preprint arXiv:2009.07485 (2020)

    Gholamalinezhad, H., Khosravi, H.: Pooling methods in deep neural networks, a review. arXiv preprint arXiv:2009.07485 (2020)

    2020 arXiv

  9. [17]

    In: Low-Power Computer Vision, pp

    Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A sur- vey of quantization methods for efficient neural network inference. In: Low-Power Computer Vision, pp. 291–326. Chapman and Hall/CRC (2022)

    2022

  10. [18]

    In: CVPR

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: CVPR. pp. 770–778 (2016)

    2016

  11. [19]

    In: ECCV

    He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. In: ECCV. pp. 630–645. Springer International Publishing, Cham (2016)

    2016

  12. [20]

    arXiv preprint arXiv:1502.03167 (2015) 16 G

    Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167 (2015) 16 G. Lagani et al

    2015 arXiv

  13. [21]

    In: CVPR

    Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., Kalenichenko,D.:Quantizationandtrainingofneuralnetworksforefficientinteger- arithmetic-only inference. In: CVPR. pp. 2704–2713 (2018)

    2018

  14. [22]

    IEEE TPAMI35(1), 221–231 (2012)

    Ji, S., Xu, W., Yang, M., Yu, K.: 3d convolutional neural networks for human action recognition. IEEE TPAMI35(1), 221–231 (2012)

    2012

  15. [23]

    In: CVPR

    Karpathy,A.,Toderici,G.,Shetty,S.,Leung,T.,Sukthankar,R.,Fei-Fei,L.:Large- scale video classification with convolutional neural networks. In: CVPR. pp. 1725– 1732 (2014)

    2014

  16. [24]

    arXiv preprint arXiv:1705.06950 (2017)

    Kay, W., Carreira, J., Simonyan, K., Zhang, B., Hillier, C., Vijayanarasimhan, S., Viola, F., Green, T., Back, T., Natsev, P., et al.: The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017)

    2017 arXiv

  17. [25]

    arXiv preprint arXiv:1806.08342 (2018)

    Krishnamoorthi, R.: Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342 (2018)

    2018 arXiv

  18. [26]

    In: ICCV

    Kuehne, H., Jhuang, H., Garrote, E., Poggio, T., Serre, T.: Hmdb: a large video database for human motion recognition. In: ICCV. pp. 2556–2563. IEEE (2011)

    2011

  19. [27]

    arXiv preprint arXiv:2307.16235 (2023)

    Lagani, G., Falchi, F., Gennaro, C., Amato, G.: Spiking neural networks and bio-inspired supervised deep learning: A survey. arXiv preprint arXiv:2307.16235 (2023)

    2023 arXiv

  20. [28]

    arXiv preprint arXiv:2307.16236 (2023)

    Lagani, G., Falchi, F., Gennaro, C., Amato, G.: Synaptic plasticity models and bio- inspired unsupervised deep learning: A survey. arXiv preprint arXiv:2307.16236 (2023)

    2023 arXiv

  21. [29]

    arXiv preprint arXiv:2112.13492 (2021)

    Lee, S.H., Lee, S., Song, B.C.: Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492 (2021)

    2021 arXiv

  22. [30]

    arXiv preprint arXiv:2201.04676 (2022)

    Li, K., Wang, Y., Gao, P., Song, G., Liu, Y., Li, H., Qiao, Y.: Uniformer: Uni- fied transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676 (2022)

    2022 arXiv

  23. [31]

    arXiv preprint arXiv:2211.09552 (2022)

    Li, K., Wang, Y., He, Y., Li, Y., Wang, Y., Wang, L., Qiao, Y.: Uniformerv2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552 (2022)

    2022 arXiv

  24. [32]

    In: ICCV

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: ICCV. pp. 10012–10022 (2021)

    2021

  25. [33]

    In: CVPR

    Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H.: Video swin trans- former. In: CVPR. pp. 3202–3211 (2022)

    2022

  26. [34]

    In: ICML (2010)

    Nair, V., Hinton, G.E.: Rectified linear units improve restricted boltzmann ma- chines. In: ICML (2010)

    2010

  27. [35]

    In: WACV

    Ng, J.Y.H., Choi, J., Neumann, J., Davis, L.S.: Actionflownet: Learning motion representation for action recognition. In: WACV. pp. 1616–1624. IEEE (2018)

    2018

  28. [36]

    In: CVPR

    Piergiovanni, A., Kuo, W., Angelova, A.: Rethinking video vits: Sparse video tubes for joint image and video learning. In: CVPR. pp. 2214–2224 (2023)

    2023

  29. [37]

    Sharir, G., Noy, A., Zelnik-Manor, L.: An image is worth 16x16 words, what is a video worth? arXiv preprint arXiv:2103.13915 (2021)

    2021 arXiv

  30. [38]

    NeurIPS27(2014)

    Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recog- nition in videos. NeurIPS27(2014)

    2014

  31. [39]

    arXiv preprint arXiv:1212.0402 (2012)

    Soomro, K., Zamir, A.R., Shah, M.: Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012)

    2012 arXiv

  32. [40]

    JMLR15(1), 1929–1958 (2014)

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. JMLR15(1), 1929–1958 (2014)

    2014

  33. [41]

    In: WACV

    Stroud, J., Ross, D., Sun, C., Deng, J., Sukthankar, R.: D3d: Distilled 3d networks for video action recognition. In: WACV. pp. 625–634 (2020) Convolutional-Attentional 3D Nets 17

    2020

  34. [42]

    In: CVPR

    Sun, S., Kuang, Z., Sheng, L., Ouyang, W., Zhang, W.: Optical flow guided feature: A fast and robust motion representation for video action recognition. In: CVPR. pp. 1390–1399 (2018)

    2018

  35. [43]

    In: ICCV

    Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotem- poral features with 3d convolutional networks. In: ICCV. pp. 4489–4497 (2015)

    2015

  36. [44]

    In: CVPR

    Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y., Paluri, M.: A closer look at spatiotemporal convolutions for action recognition. In: CVPR. pp. 6450–6459 (2018)

    2018

  37. [45]

    In: NeurIPS

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. In: NeurIPS. pp. 5998–6008 (2017)

    2017

  38. [46]

    arXiv preprint arXiv:2006.04768 (2020)

    Wang, S., Li, B.Z., Khabsa, M., Fang, H., Ma, H.: Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768 (2020)

    2020 arXiv

  39. [47]

    In: CVPR

    Wang, X., Girshick, R., Gupta, A., He, K.: Non-local neural networks. In: CVPR. pp. 7794–7803 (2018)

    2018

  40. [48]

    In: ECCV

    Xie, S., Sun, C., Huang, J., Tu, Z., Murphy, K.: Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In: ECCV. pp. 305–321 (2018)

    2018

  41. [49]

    In: ECCV

    Zhang, D., Yang, J., Ye, D., Hua, G.: Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In: ECCV. pp. 365–382 (2018)

    2018

  42. [50]

    In: ICCV

    Zhang, Y., Li, X., Liu, C., Shuai, B., Zhu, Y., Brattoli, B., Chen, H., Marsic, I., Tighe, J.: Vidtr: Video transformer without convolutions. In: ICCV. pp. 13577– 13587 (October 2021)

    2021

  43. [51]

    In: ACCV

    Zhu, Y., Lan, Z., Newsam, S., Hauptmann, A.: Hidden two-stream convolutional networks for action recognition. In: ACCV. pp. 363–378. Springer (2019)

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.