Pith. sign in

REVIEW 4 major objections 6 minor 44 references

Two-Stream Video Classification with Cross-Modality Attention

T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Cross-modality attention, which lets one video branch query the other's feature maps, beats late score fusion while using fewer parameters.

desk verdict The CMA gain is plausibly real but not demonstrated: the +0.96% is confounded with a different training protocol. read the letter →

arxiv 1908.00497 v1 pith:ISZBEAE7 submitted 2019-08-01 cs.CV

Lu Chi , Guiyu Tian , Yadong Mu , Qi Tian This is my paper · ORCID
classification cs.CV
keywords videoclassificationcross-modalityattentiontwo-streamnetworksopticalflowactionrecognitionmechanismnon-localtemporalsegment
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 argues that the usual way of combining RGB and optical-flow streams in video classification—averaging their prediction scores at the very end—leaves useful cross-modal information on the table. It proposes a cross-modality attention (CMA) block that lets features in one branch query the other branch's feature maps at intermediate layers, so each modality can selectively pull in complementary information from the other. On Kinetics, the RGB branch of the CMA model reaches 72.17% top-1 accuracy with 1.8 times a ResNet-50's parameters, beating the 71.21% of a two-stream ResNet-50 that uses 2 times the parameters; using both branches gives 72.62%. The block is designed as a drop-in component for existing two-stream and 3D architectures, which is why the result would matter beyond this specific setting.

What carries the argument

The central object is the CMA block, a wrapper around a cross-modality attention operation $z_i = \frac{1}{C(x,y)} \sum_j f(x_i,y_j) v(y_j)$ with $f(x_i,y_j) = \exp(q(x_i) k(y_j)^T / \sqrt{d_k})$, where $x$ and $y$ are feature maps from different branches. Queries come from one modality while keys and values come from the other, so each spatial position can absorb information from every position of the other modality. A $1\times 1$ convolution $W_{\text{out}}$ and a residual connection $+x_i$ wrap the operation, and the final batch-normalization scale is initialized to zero, making the block an identity mapping at initialization. That identity initialization is what lets it be inserted into any network without disrupting pretrained weights.

What would settle it

Train the two-stream ResNet-50 baseline under the exact CMA schedule—flow branch pretrained on Kinetics and frozen, RGB branch trained in isolation—but with no CMA blocks; if it also reaches roughly 72% top-1 on Kinetics, the attention block is not what drives the gain. A second check is to verify that zeroing $W_{\text{out}}$ in every CMA block recovers the baseline accuracy exactly, as the identity-initialization claim predicts.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-modality attention, computed globally as a softmax-weighted sum over positions in the other modality, is a more effective fusion mechanism than late score averaging. The CMA operation follows the query-key-value form: a position in the RGB stream queries all positions in the flow stream, and the weighted flow features are added back into the RGB stream through a residual connection. The same operation applied in reverse lets the flow stream attend to RGB. Inserting five such blocks into a ResNet-50 at stages res3 and res4 yields a single RGB stream that outperforms the full two-stream baseline (72.17% versus 71.21% top-1 on Kinetics) at lower parameter cost, and the fusion of both streams reaches 72.62%. The paper also shows the block works when both inputs come from the same modality, in which case it becomes the non-local operation, and when inserted into a 3D P3D network.

Load-bearing premise

The load-bearing premise is that the accuracy gain comes from the CMA attention operation itself, yet the CMA model is trained with a different procedure than the two-stream baseline—pretraining and freezing the flow branch and training branches iteratively—so the training schedule alone could explain the improvement.

Editorial extensions

If this is right

  • Hierarchical cross-modal attention can replace late score averaging without increasing parameter count: the RGB stream alone exceeds the two-stream fusion accuracy.
  • Because CMA reduces to the non-local operation when both inputs come from the same modality, one block type can supply both cross-modal and self-attention, making separate non-local blocks unnecessary.
  • CMA blocks transfer to 3D convolutional backbones: adding one block to P3D improves its accuracy whether the network is fed 12 or 16 frames.
  • On UCF-101 after Kinetics pretraining, fine-tuning only the final layer of the CMA model reaches 96.5% with a 2D backbone, approaching models that use 3D convolutions.

Reading between the lines

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

  • Because the CMA model is trained with a different schedule—flow branch pretrained on Kinetics and frozen, branches updated iteratively—than the jointly trained two-stream baseline, the reported gain cannot be cleanly attributed to attention; a same-schedule baseline without CMA blocks would isolate the effect.
  • The attention maps' focus on moving, task-relevant objects suggests the learned attention could double as a motion-saliency prior for weakly supervised action localization, a use the paper does not develop.
  • The query-key-value formulation is modality-agnostic, so a natural extension is fusing RGB with audio, depth, or pose streams wherever their features can be spatially or temporally aligned.
  • The paper's reported failure mode—queries on background attending to background—points to a concrete refinement: a confidence or saliency gate on query positions could reduce wasted attention and is testable on Kinetics.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 a cross-modality attention (CMA) block that lets one modality branch (e.g., RGB) attend over positions in another modality branch (e.g., optical flow), in contrast to self-attention non-local blocks. The block is designed as a residual wrapper and can be inserted into existing two-stream or 3D CNN architectures. The authors report experiments on Kinetics and UCF-101, claiming that CMA blocks improve accuracy over two-stream baselines and non-local networks. The central quantitative claim is that CMA iter1-R reaches 72.17% top-1 on Kinetics compared with the two-stream baseline's 71.21%, while using fewer parameters (Table 2). The paper also includes attention visualizations and a transfer-learning experiment to UCF-101.

Significance. If the empirical claim were cleanly established, the cross-modality attention idea would be a useful and well-motivated contribution to video classification: it generalizes non-local self-attention to cross-modal attention, is architecture-agnostic, and the visualizations suggest interpretable behavior. The paper also demonstrates compatibility with both 2D and 3D backbones, and the residual design with zero-initialized BN is a sensible implementation choice. However, the evidence as presented does not isolate the effect of the CMA block from the accompanying change in training strategy, so the significance of the claimed gains is currently not established.

major comments (4)
  1. [§4.3, §5.3, Table 2] The main claim that the CMA block improves accuracy is confounded by a difference in training protocol. The baseline two-stream (iter0) is trained with both branches trained independently (per the footnote in §5.3), whereas CMA iter1-R is obtained after training the RGB branch while the flow branch is frozen and pretrained on Kinetics. No no-CMA baseline is run under the same iterative pretrained-frozen-flow protocol. Therefore the reported +0.96% (72.17 vs. 71.21) cannot be causally attributed to the cross-modality attention block; it may be an artifact of the pretraining/freezing scheme. A controlled ablation (e.g., two-stream with the same frozen flow branch and iterative training but without CMA blocks) is required to support the paper's central claim.
  2. [§5.5, Table 4] The comparison with non-local networks is also uncontrolled. The non-local RGB+Flow model is trained without the iterative frozen-flow scheme, while the CMA models use that scheme, so the accuracy gap (71.67 vs. 72.17 or 72.62) cannot be assigned to the attention mechanism. In addition, the non-local blocks are added only to the RGB branch, whereas the CMA model includes blocks in both branches (even if the flow branch is frozen during iter1), making the architectural comparison asymmetric. The conclusion that CMA outperforms non-local is therefore not justified by the presented experiments.
  3. [§5.6, Table 5] The 3D-CMA experiments suffer from the same missing control. In Table 5, the Flow column is identical to the P3D baseline, and only the RGB branch is augmented with a single CMA block and then partially fine-tuned. There is no no-CMA P3D model fine-tuned with the same layer-wise protocol (training only the last block and following layers), so the improvement in the RGB column (e.g., 74.41 vs. 70.98 for 12 frames) could stem from the fine-tuning scheme rather than the CMA block. A matched baseline is needed to attribute the gain.
  4. [§5.2–§5.4] All reported accuracy differences are small (roughly 0.5–1%) and no error bars or multiple-seed results are provided. Given the known variance in training deep video models on Kinetics and the sensitivity to hyperparameters (including the fusion weights selected on the validation set in §5.2), the reported gaps are not shown to be statistically meaningful even apart from the confounds above. The paper should report mean and standard deviation over multiple runs or otherwise justify that the differences exceed training noise.
minor comments (6)
  1. [§2] The text contains a typo: "litter has been done" should be "little has been done."
  2. [§5.3 footnote] The footnote explaining that iter0 does not initialize the CMA model is helpful but confusingly worded; it should explicitly state that the baseline is trained with the standard two-stream protocol (both branches trained from scratch or with standard initialization) and not with the iterative scheme described in §4.3.
  3. [§5.2, Figure 3] The x-axis of Figure 3 is labeled "Fusion Weight Ratio (RGB:Flow)" but the actual tick values are not readable in the text version; please ensure the figure is legible and that the selected weights (5:1 and 1:5) are clearly marked.
  4. [§5.4] The sentence "Figure 4 showed the top-20 most improved categories" mixes tenses; also, the figure itself is not included in the text, so the reader cannot verify the claimed comparison. Please include the figure or refer to the appendix.
  5. [Table 4] The meaning of the "modality" column for CMA rows is ambiguous: the rows labeled "RGB + Flow" refer to the two-stream fusion output, but it would be clearer to indicate which branch output is being evaluated (e.g., "CMA iter1-R" is the RGB branch output).
  6. [§4.4] There is a typo: "spatial aligned" should be "spatially aligned."

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the CMA results are empirical, and the missing controlled baseline is an experimental confound rather than a circular reduction.

full rationale

This paper proposes a cross-modality attention block and evaluates it empirically on Kinetics and UCF-101. There is no theoretical derivation in which a predicted quantity is defined in terms of the result, no fitted parameter is renamed as a prediction, and no load-bearing self-citation chain. The CMA operation (Eq. 1) and block (Eq. 5) are defined directly from the input feature maps and a residual connection; the residual connection guarantees an identity initialization but does not force the reported accuracy gain. The fusion weights in Section 5.2 are selected by validation performance, which is standard hyperparameter tuning rather than a circular prediction. The closest concern is that Table 2 attributes CMA iter1-R's 72.17% top-1 accuracy to the CMA block, while Section 4.3 describes an iterative training scheme in which the flow branch is pretrained on Kinetics and frozen while the RGB branch is trained, and the iter0 two-stream baseline is trained independently. This missing controlled baseline means the observed gain may be confounded by the training protocol, but that is a soundness or attribution issue, not circularity: the paper does not derive its conclusion from the same data by construction. No specific equation or citation is shown to reduce the claim to its inputs, so under the hard rules no circular step can be flagged.

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

The paper's central claim relies on the validity of the experimental comparison. The main free parameter is the fusion weight tuned on validation. The axioms reflect the untested assumptions about alignment, the training procedure, and the adequacy of the baseline.

free parameters (1)
  • Fusion weight ratio (RGB:Flow) = 5:1 at odd iterations, 1:5 at even iterations
    The fusion weights are tuned on the Kinetics validation set, as shown in Figure 3 and described in Section 5.2. The paper chooses the ratio that maximizes validation accuracy, and this choice directly affects the reported fusion accuracy numbers.
assumptions (3)
  • domain assumption Spatial alignment between RGB frames and optical flow fields is preserved through the network, so that cross-position attention between the two modalities is meaningful.
    Stated in Section 4.4: 'all RGB frame / optical flows are spatially aligned'. If the feature maps are not aligned, the attention operation may attend to irrelevant locations, yet the paper does not ablate this assumption.
  • ad hoc to paper The iterative training strategy with one branch frozen at a time is a valid optimization approach that does not itself change accuracy.
    The paper introduces this training procedure in Section 4.3 without comparing against a jointly trained CMA model. The procedure is a distinctive element of the method, and its effect is not controlled.
  • ad hoc to paper The reported two-stream baseline (71.21% on Kinetics) is a fair and strong baseline.
    The baseline is trained jointly without the iterative freezing scheme. The paper does not provide a baseline using the same training strategy but without CMA blocks, which is essential to attribute the improvement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Two-Stream Video Classification with Cross-Modality Attention." pith.science (2026). https://pith.science/paper/ISZBEAE7

@misc{pith2026190800497,
  author       = {Pith},
  title        = {Pith review of: Two-Stream Video Classification with Cross-Modality Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ISZBEAE7}},
  note         = {Machine review of arXiv:1908.00497}
}
read the original abstract

Fusing multi-modality information is known to be able to effectively bring significant improvement in video classification. However, the most popular method up to now is still simply fusing each stream's prediction scores at the last stage. A valid question is whether there exists a more effective method to fuse information cross modality. With the development of attention mechanism in natural language processing, there emerge many successful applications of attention in the field of computer vision. In this paper, we propose a cross-modality attention operation, which can obtain information from other modality in a more effective way than two-stream. Correspondingly we implement a compatible block named CMA block, which is a wrapper of our proposed attention operation. CMA can be plugged into many existing architectures. In the experiments, we comprehensively compare our method with two-stream and non-local models widely used in video classification. All experiments clearly demonstrate strong performance superiority by our proposed method. We also analyze the advantages of the CMA block by visualizing the attention map, which intuitively shows how the block helps the final prediction.

Figures

Figures reproduced from arXiv: 1908.00497 by the authors.

Figure 1
Figure 1. An example of CMA block. We show the shape of fea￾ture maps at each stage, such as H ×W ×1024, where 1024 is the number of channels. Let X be the feature maps of the RGB branch and Y be the feature maps of flow branch. The number of chan￾nels is halved via 1 × 1 convolutions. Reshaping or transposing is performed whenever needed. “N” denotes matrix multiplication, and “L” denotes element-wise sum. 3.2. CMA Block A C… view at source ↗
Figure 2
Figure 2. An overview of the video classification model. This model contains both RGB branch and Flow branch. In each branch, we insert n CMA blocks, which play an important role in transmitting information between different branches. There are three outputs in this model. Operations in dotted box are not essential since practically we can only use the output of the RGB branch for prediction. convolution in space-time, denote… view at source ↗
Figure 3
Figure 3. The top-1 accuracies with different fusion weights. The CMA models perform better than the two-stream when we give higher weight to the more reliable branch. models, at the first training iteration, we just train the RGB branch with the Flow branch fixed, so the RGB branch per￾forms better than the Flow branch. In other words, the RGB branch is more reliable. Giving the RGB branch more weight will make the final acc… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparing top-20 most improved categories between the proposed CMA model and two-stream. also list the top-10 confusing categories in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Examples of the attention maps. We train CMA iter1-R on Kinetics and visualize the attention maps of the last CMA block in res4 since the last block is the most related to the final classification. These samples are taken from Kinetics randomly. Each set contains three…
Figure 6
Figure 6. Figure 6: Examples of attention maps in CMA. The columns from left to right are RGB frame, optical flow field and attention maps of three different interested positions. Examples highlighted by red bounding boxes are failure cases [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Examples of Non-local and CMA’s attention maps [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 32 canonical work pages

  1. [1]

    Abu-El-Haija, N

    S. Abu-El-Haija, N. Kothari, J. Lee, P. Natsev, G. Toderici, B. Varadarajan, and S. Vijayanarasimhan. Youtube-8m: A large-scale video classification bench- mark. CoRR, abs/1609.08675, 2016

    arXiv 2016

  2. [2]

    Bahdanau, K

    D. Bahdanau, K. Cho, and Y . Bengio. Neural machine translation by jointly learning to align and translate. CoRR, abs/1409.0473, 2014

    arXiv 2014

  3. [3]

    Carreira and A

    J. Carreira and A. Zisserman. Quo vadis, action recog- nition? A new model and the kinetics dataset. In CVPR, pages 4724–4733, 2017

    work page 2017

  4. [4]

    Y . Chen, Z. Wang, Y . Peng, Z. Zhang, G. Yu, and J. Sun. Cascaded pyramid network for multi-person pose estimation. CoRR, abs/1711.07319, 2017

    arXiv 2017

  5. [5]

    Devlin, M

    J. Devlin, M. Chang, K. Lee, and K. Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018

    arXiv 2018

  6. [6]

    W. Du, Y . Wang, and Y . Qiao. Rpan: An end-to-end recurrent pose-attention network for action recogni- tion in videos. In ICCV, volume 2, 2017

    work page 2017

  7. [7]

    L. Fan, W. Huang, S. E. Chuang Gan, B. Gong, and J. Huang. End-to-end learning of motion representa- tion for video understanding. In CVPR, pages 6016– 6025, 2018

    work page 2018

  8. [8]

    J. Fu, J. Liu, H. Tian, Z. Fang, and H. Lu. Dual attention network for scene segmentation. CoRR, abs/1809.02983, 2018

    arXiv 2018

Show all 44 references
  1. [9]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770– 778, 2016

    2016

  2. [10]

    Huang, Z

    G. Huang, Z. Liu, L. van der Maaten, and K. Q. Wein- berger. Densely connected convolutional networks. In CVPR, pages 2261–2269, 2017

    2017

  3. [11]

    E. Ilg, N. Mayer, T. Saikia, M. Keuper, A. Dosovit- skiy, and T. Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, pages 1647–1655, 2017

    2017

  4. [12]

    S. Ji, W. Xu, M. Yang, and K. Yu. 3d convolu- tional neural networks for human action recognition. IEEE Trans. Pattern Anal. Mach. Intell. , 35(1):221– 231, 2013

    2013

  5. [13]

    Karpathy, G

    A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Suk- thankar, and F. Li. Large-scale video classification with convolutional neural networks. In CVPR, pages 1725–1732, 2014

    2014

  6. [14]

    W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev, M. Suleyman, and A. Zisser- man. The kinetics human action video dataset. CoRR, abs/1705.06950, 2017

    2017 arXiv

  7. [15]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. Im- agenet classification with deep convolutional neural networks. In NIPS, pages 1097–1105, 2012

    2012

  8. [16]

    Liu and J

    M. Liu and J. Yuan. Recognizing human actions as the evolution of pose estimation maps. In CVPR, pages 1159–1168, 2018

    2018

  9. [17]

    X. Long, C. Gan, G. de Melo, J. Wu, X. Liu, and S. Wen. Attention clusters: Purely attention based local feature integration for video classification. In CVPR, pages 7834–7843, 2018

    2018

  10. [18]

    Mayer, E

    N. Mayer, E. Ilg, P. H ¨ausser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, pages 4040–4048, 2016

    2016

  11. [19]

    J. Y . Ng, J. Choi, J. Neumann, and L. S. Davis. Ac- tionflownet: Learning motion representation for ac- tion recognition. In WACV, pages 1616–1624, 2018

    2018

  12. [20]

    J. Y . Ng, M. J. Hausknecht, S. Vijayanarasimhan, O. Vinyals, R. Monga, and G. Toderici. Beyond short snippets: Deep networks for video classification. In CVPR, pages 4694–4702, 2015

    2015

  13. [21]

    J. C. Niebles, C. Chen, and F. Li. Modeling temporal structure of decomposable motion segments for activ- ity classification. In ECCV, pages 392–405, 2010

    2010

  14. [22]

    C. Peng, T. Xiao, Z. Li, Y . Jiang, X. Zhang, K. Jia, G. Yu, and J. Sun. Megdet: A large mini-batch object detector. CoRR, abs/1711.07240, 2017

    2017 arXiv

  15. [23]

    Z. Qiu, T. Yao, and T. Mei. Learning spatio-temporal representation with pseudo-3d residual networks. In ICCV, pages 5534–5542, 2017

    2017

  16. [24]

    S. Ren, K. He, R. B. Girshick, and J. Sun. Faster R- CNN: towards real-time object detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell., 39(6):1137–1149, 2017

    2017

  17. [25]

    Russakovsky, J

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, and M. Bernstein. Imagenet large scale visual recog- nition challenge. IJCV, 115(3):211–252, 2015

    2015

  18. [26]

    Sevilla-Lara, Y

    L. Sevilla-Lara, Y . Liao, F. G ¨uney, V . Jampani, A. Geiger, and M. J. Black. On the integra- tion of optical flow and action recognition. CoRR, abs/1712.08416, 2017

    2017 arXiv

  19. [27]

    Sharma, R

    S. Sharma, R. Kiros, and R. Salakhutdinov. Ac- tion recognition using visual attention. CoRR, abs/1511.04119, 2015

    2015 arXiv

  20. [28]

    Simonyan and A

    K. Simonyan and A. Zisserman. Two-stream convo- lutional networks for action recognition in videos. In NIPS, pages 568–576, 2014

    2014

  21. [29]

    Simonyan and A

    K. Simonyan and A. Zisserman. Very deep convo- lutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014

    2014 arXiv

  22. [30]

    Soomro, A

    K. Soomro, A. R. Zamir, and M. Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012

    2012 arXiv

  23. [31]

    S. Sun, Z. Kuang, L. Sheng, W. Ouyang, and W. Zhang. Optical flow guided feature: A fast and robust motion representation for video action recogni- tion. In CVPR, 2018

    2018

  24. [32]

    D. Tran, L. D. Bourdev, R. Fergus, L. Torresani, and M. Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, pages 4489–4497, 2015

    2015

  25. [33]

    D. Tran, H. Wang, L. Torresani, J. Ray, Y . LeCun, and M. Paluri. A closer look at spatiotemporal convolu- tions for action recognition. In CVPR, pages 6450– 6459, 2018

    2018

  26. [34]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In NIPS, pages 6000–6010, 2017

    2017

  27. [35]

    H. Wang, M. M. Ullah, A. Kl ¨aser, I. Laptev, and C. Schmid. Evaluation of local spatio-temporal fea- tures for action recognition. In BMVC, pages 1–11, 2009

    2009

  28. [36]

    L. Wang, W. Li, W. Li, and L. V . Gool. Appearance- and-relation networks for video classification. In CVPR, pages 1430–1439, 2018

    2018

  29. [37]

    L. Wang, Y . Xiong, Z. Wang, Y . Qiao, D. Lin, X. Tang, and L. V . Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, pages 20–36, 2016

    2016

  30. [38]

    X. Wang, R. Girshick, A. Gupta, and K. He. Non-local neural networks. CVPR, 2018

    2018

  31. [39]

    Z. Wu, X. Wang, Y . Jiang, H. Ye, and X. Xue. Mod- eling spatial-temporal clues in a hybrid deep learning framework for video classification. In ACM Multime- dia, pages 461–470, 2015

    2015

  32. [40]

    S. Xie, C. Sun, J. Huang, Z. Tu, and K. Murphy. Re- thinking spatiotemporal feature learning for video un- derstanding. CoRR, abs/1712.04851, 2017

    2017 arXiv

  33. [41]

    C. Zach, T. Pock, and H. Bischof. A duality based approach for realtime tv- L1 optical flow. In DAGM, pages 214–223, 2007

    2007

  34. [42]

    Zhang, I

    H. Zhang, I. J. Goodfellow, D. N. Metaxas, and A. Odena. Self-attention generative adversarial net- works. CoRR, abs/1805.08318, 2018

    2018 arXiv

  35. [43]

    Z. Zhao, H. Ma, and S. You. Single image ac- tion recognition using semantic body part actions. In ICCV, pages 3411–3419, 2017

    2017

  36. [44]

    Zolfaghari, K

    M. Zolfaghari, K. Singh, and T. Brox. ECO: efficient convolutional network for online video understanding. In ECCV, 2018. Appendices A. More attention maps To more intuitively illustrate the effect of CMA blocks, we show more attention maps of the last CMA block of CMA iter1-R ...

    2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.