Pith. sign in

REVIEW 3 major objections 7 minor 95 references

Discriminative Video Representation Learning Using Support Vector Classifiers

T0 review · 3 major / 7 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 This paper claims that the parameters of an SVM separating a video's features from a bag of noise features form a video descriptor that beats standard average/max pooling on eight benchmarks.

desk verdict Solid empirical extension of the CVPR SVM-pooling work, but the paper's formal story (P1) is not what the experiments actually test—Algorithm 2 is a heuristic, and the claimed equivalence is unproved. read the letter →

arxiv 1909.02856 v1 pith:COSGZ772 submitted 2019-09-05 cs.CV

Jue Wang , Anoop Cherian This is my paper · ORCID
classification cs.CV
keywords videorepresentationactionrecognitiondiscriminativepoolingmultipleinstancelearningsupportvectormachineSVMdeeptemporalordering
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

Most action-recognition models classify short clips and then pool the per-clip predictions, typically by averaging or max pooling, which weights all frames equally even though many frames are uninformative background. This paper proposes SVM pooling: treat the deep features of a video as a positive bag, take a fixed bag of irrelevant features (usually CNN features of white-noise images) as a negative bag, and learn a max-margin SVM hyperplane that separates a fraction of the positive bag from the negatives. The parameters of that hyperplane become the descriptor for the whole video. Because the SVM's support vectors determine the hyperplane, SVMP is effectively a weighted average pooling that assigns zero weight to non-support vectors, filtering out frames the boundary deems irrelevant. On eight benchmarks spanning action recognition, action detection, skeleton-based recognition, image-set verification, and texture recognition, the authors report that SVMP improves over standard pooling by 3–14% and over prior state-of-the-art results by 1–5%.

What carries the argument

The central object is the SVM Pooling (SVMP) descriptor, defined as $\mathrm{SVMP}(X) = [w, b]^\top$ for a sequence $X$, where $(w,b)$ solves the max-margin MIL objective in Eq. (1): separate all negative-bag features from at least a fraction $\eta$ of the positive-bag features, allowing slacks $\zeta$ with regularization $C_1$. The paper's practical inference is Algorithm 2, which increases $C$ until the desired fraction of positive features lies on the positive side of the hyperplane; the resulting $w$ is a sparse linear combination of support vectors, giving the descriptor its weighted-pooling interpretation. Temporal extension adds constraints $w^\top x_j + \delta \le w^\top x_k$ for $j<k$; the end-to-end CNN version differentiates through the argmin using the implicit function theorem, yielding a gradient of the form $-\left(I+\lambda \sum_j (\theta_j z_j)(\theta_j z_j)^\top\right)^{-1}[\cdots]$ over the active hinge terms.

What would settle it

On a dataset whose frames carry per-frame action/background labels, test whether the positive support vectors selected by SVMP match the labeled action frames well above chance; if they do not, the claimed mechanism of filtering uninformative frames is not what drives the accuracy gains. A second test: swap the white-noise negative bag for features of background frames from the same videos and see whether accuracy improves, stays flat, or drops.

Watch

Extended reading notes

Core claim

The central claim is that the decision boundary of a binary SVM, trained to separate a video's own frame-level features from a fixed bag of irrelevant features in a multiple-instance-learning setup, is itself a strong video-level descriptor. The objective requires at least a fraction $\eta$ of the positive bag to be classified on the positive side, and the descriptor is the tuple $[w,b]$ defining the separating hyperplane (Definition 1). Around this, the paper builds three practical solvers—exhaustive enumeration, alternating optimization, and a parameter-tuning heuristic that raises the SVM cost $C$ until the $\eta$ constraint is met—plus nonlinear (homogeneous-kernel) and temporally ordered variants, and an end-to-end trainable version whose gradient is computed via the implicit function theorem. The authors report that this SVM-pooled descriptor, with temporal ordering enabled, outperforms average and max pooling and beats previous state-of-the-art results on HMDB-51, UCF-101, Kinetics-600, Charades, MSR Action3D, NTU-RGBD, PubFig, and YUP++ by 1–5%. They further observe that the gains over strong baselines grow with sequence length, being negligible on very short clips and largest on videos with at least 180–260 frames.

Load-bearing premise

The method assumes that features from random white-noise images (or whichever negative bag is chosen) resemble the uninformative frames inside real action videos closely enough that separating a video from them selects the truly informative frames.

Editorial extensions

If this is right

  • Replacing average or max pooling with SVMP improves accuracy on all eight benchmarks, by 3–14% over standard pooling and by 1–5% over prior state-of-the-art methods, as reported in the paper.
  • The descriptor is feature-agnostic: it works on CNN activations, hand-crafted trajectory features, and skeleton Lie-algebra encodings, so it can be dropped into existing pipelines without retraining the feature extractor.
  • Temporal-ordering constraints (ordered-SVMP) and homogeneous-kernel nonlinear maps (NSVMP) consistently improve on the plain SVMP, and the combination of linear and nonlinear descriptors gives a further boost.
  • The end-to-end trainable SVMP layer performs slightly below the offline SVMP on two-stream VGG/ResNet models, because of the need to approximate the Hessian, yet still beats average pooling.
  • The gains over strong baselines such as I3D+ depend on sequence length: the paper reports negligible improvement on short 64-frame chunks and growing improvements on longer sequences, with the largest gains on subsets of videos with at least 180–260 frames.

Reading between the lines

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

  • One extension the authors leave implicit is that the negative bag is the lever for the whole mechanism: choosing negatives that better match the background distribution of a target video could sharpen the descriptor, and the white-noise choice may not be optimal across datasets or feature spaces.
  • The sparse support-vector selection can be read as a hard attention over frames, suggesting SVMP could seed weakly supervised frame localization or keyframe selection for long-form video summarization.
  • Because the gradient of the argmin is derived, the pooling layer could be inserted into modern transformer-based video models in place of CLS-token or average pooling; the short-sequence weakness observed for I3D+ suggests the per-sequence regularization constant, not the pooling idea itself, may need recalibrating for such models.
  • The alternating MIL solver (Algorithm 1) carries no convergence guarantee, and the deployed parameter-tuning solver (Algorithm 2) is a heuristic; a more principled mixed-integer or differentiable MIL solver might change which frames are kept, potentially closing the gap to end-to-end training.
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

3 major / 7 minor

Summary. The paper proposes SVM pooling (SVMP): given the CNN features of frames or clips in a video as a positive bag and a fixed negative bag (typically CNN features from random white-noise images), it learns an SVM hyperplane that separates at least a fraction η of the positive features from all negatives, and uses the hyperplane parameters (w,b) as a video descriptor. The manuscript presents linear, kernelized, temporally ordered, and end-to-end trainable variants of the idea, and reports experiments on eight benchmarks spanning action recognition, action detection, skeleton-based recognition, image-set verification, and dynamic texture recognition. The central claim is that this discriminative, max-margin pooling filters out uninformative frames and yields state-of-the-art or competitive accuracy, improving over standard average/max pooling by 3–14% and over prior published results by 1–5%.

Significance. If the claims hold, SVMP is a simple, feature-extractor-agnostic pooling layer that could be plugged into many CNN-based video systems; its compact fixed-length descriptor and the breadth of the evaluation are genuine strengths. The paper also makes a useful connection between multiple instance learning and standard SVM solvers, and the temporal and end-to-end variants broaden the potential impact. However, the formal bridge between the paper's Definition 1 and the algorithm actually used in the experiments is not established, and several empirical comparisons lack statistical support. The central idea is credible, but the manuscript needs substantial clarification and additional validation before the claimed results can be accepted at face value.

major comments (3)
  1. [3.3.3 (Eq. (1)–(5) vs. Algorithm 2)] The claim that Algorithm 2 is equivalent to solving P1 is not established. P1 involves a non-convex subset-selection constraint |X̂+| ≥ η|X+| with unknown labels θ, whereas Algorithm 2 trains a soft-margin SVM with every positive-bag feature labeled +1 and raises C until at least the fraction η of positives lie on the positive side; the points on the negative side are margin errors, not decision variables. Convexity of each fixed-label SVM subproblem does not imply that the two optimizers coincide, and the paper itself calls Algorithm 2 a heuristic. Because all subsequent experiments use Algorithm 2 (and its ordered variant), the reported results validate Algorithm 2, not the Definition 1 SVMP descriptor. The authors should either prove the claimed equivalence under precise assumptions, or redefine SVMP as the output of Algorithm 2 and adjust the conceptual narrative accordingly; a small-scale comparison of Algorithms 1 and 2 against an exact or enumerated P1 solution would clarify the actual gap.
  2. [5.2 (Figure 5, negative-bag selection)] The choice of negative bag is the defining modeling assumption of the SVMP descriptor, but the comparison across the four negative-bag candidates (ActivityNet, UCF-101, Thumos, white noise) is reported without error bars or repeated trials, and the accompanying TSNE discussion is qualitative. The claim that white-noise CNN features are a good proxy for uninformative frames in real videos needs direct quantitative support: please report mean and variance over several random negative-bag draws on at least a subset of the datasets, and ideally show how validation-selected hyperparameters (η, C, bag sizes) transfer across datasets. Without this, it is difficult to tell how much of the reported gain depends on the single chosen negative-bag distribution.
  3. [4.2 (Eq. (9))] The end-to-end gradient is derived for a squared-hinge SVM objective with a fixed regularization parameter λ, but the SVMP output used in experiments is obtained by Algorithm 2, which iteratively increases C until the η constraint is satisfied. This iterative search is not the argmin of a fixed objective, so Lemma 3.3 of [80] does not apply directly as written. The end-to-end experiments are therefore not supported by the derivation as stated; the authors should either derive gradients for the actual Algorithm 2 procedure, or present a controlled experiment showing that the approximate gradient behaves as intended (e.g., by comparing E2E training curves against a finite-difference or alternative approximation).
minor comments (7)
  1. [3.5] The temporally-ordered variant is not fully specified: Eq. (8) lists ordering constraints, but the actual optimization problem minimized by the 'custom solver' is not stated, and reference [78] is Manopt, a manifold optimization toolbox, not an SVM solver. Please provide the complete objective and implementation details, since ordered-SVMP is the variant used in most video experiments.
  2. [5.2] The text and Figure 5 panels are inconsistent: the text says Figure 5(b) shows accuracy versus C, but the caption labels that panel as a bag-size analysis, and the panel labels for negative-bag selection and η are also unclear. Please renumber the panels and fix all references.
  3. [Tables 6–9] All experimental numbers are point estimates with no error bars, number of repeats, or significance tests. For claims of 1–3% improvements over state-of-the-art, some measure of variance is needed, especially on the three-fold HMDB/UCF protocols where splits are known to have nontrivial variance.
  4. [5.7, Table 8] Please clarify whether the row labeled 'I3D' is the I3D+ model with average pooling and whether the 'SVMP' row uses exactly the same input features, training data, and evaluation protocol; otherwise the comparison is not controlled.
  5. [5.6 (MSR Action3D)] For non-CNN skeleton features, the paper says white noise with the dataset mean and deviation is used directly as the negative bag; please explain why raw white noise in the feature space is a meaningful proxy for uninformative skeleton configurations, since this is a different construction from the CNN-feature noise negatives used elsewhere.
  6. [Throughout] Please correct typos and spacing errors, including 'Howeer', 'we ues', 'V emulapalli', and 'bgs'.
  7. [Abstract] The phrase 'nonlinear hyperplane' is imprecise; a kernelized SVM still yields a hyperplane in the induced feature space. Please rephrase accordingly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SVMP descriptor is an empirical pooling method validated on held-out benchmarks; the unproved Algorithm 2 to P1 equivalence is a correctness risk, not a circular reduction.

full rationale

This is an empirical method paper, not a derivation of a constant. The SVMP descriptor is defined as the solution of P1 (Definition 1), but the experiments use Algorithm 2, which the paper explicitly labels 'a heuristic scheme' in Section 3.3.3. The bridge between the definition and the implementation is the claim that 'due to the convexity of SVM, it can be shown that the optimizing objective of P1 will be the same in both cases.' Even if that equivalence is unproved or false, it is not circularity: Algorithm 2 is a proposed optimization shortcut for the same descriptor, and no experimental accuracy is forced by construction. Hyperparameters C, eta, and bag sizes are selected on validation data, and the reported numbers are measured on held-out test splits against external benchmark protocols. The negative-bag choice (white noise) is a model-selection step validated experimentally, not a fitted parameter renamed as a prediction. Self-citations appear for provenance (the conference version [73]) and for standard technical tools such as implicit differentiation of argmin problems ([80]); none supplies the central empirical claim, and the cited results are external mathematical facts rather than assertions of the paper's own conclusion. The main substantive risk is the correctness of the Algorithm 2 versus P1 equivalence, which belongs in a correctness review, not in a circularity finding. Accordingly, no circular step is identified and the score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method rests on two domain assumptions: the MIL assumption that a fraction eta of each video's features are discriminative, and the negative-bag assumption that white-noise CNN features stand in for irrelevant or background features. Standard SVM and implicit-differentiation results are used as background mathematics. No new physical entities are introduced.

free parameters (5)
  • eta (minimum positive fraction) = 0.9
    Set via parameter analysis in Figure 5(a); controls how many positive features must be classified on the positive side.
  • SVM regularization C = 10
    Tuned in Figure 5(c); larger C increases accuracy up to a point, then overfits.
  • Positive/negative bag sizes = 50/50 frames
    Figure 5(b) suggests about 30 frames per bag is sufficient; authors use 50.
  • Negative bag source = White noise CNN features
    Chosen as best among ActivityNet, UCF-101, THUMOS backgrounds, and white noise (Figure 5(a)); affects all subsequent results.
  • Temporal margin delta = not reported
    Introduced in Eq. 8 for the ordered-SVMP variant; no value or tuning procedure is given.
assumptions (4)
  • ad hoc to paper At least a fraction eta of the features in each positive video bag are classifiable and action-discriminative.
    Equation (5) in Section 3.2; the entire pooling formulation depends on this assumption.
  • ad hoc to paper CNN features from random white-noise images are linearly separable from and representative of irrelevant or background features in action videos.
    Section 5.2, 'Selecting Negative Bags'; the choice of negative bag drives the descriptor.
  • standard math SVM max-margin classification and the Representer theorem provide valid representations for this pooling problem.
    Used in Sections 3.2 and 3.4; background from [76], [77].
  • standard math The implicit function theorem applies to the bilevel optimization for end-to-end learning.
    Section 4.2, following [79], [80].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Discriminative Video Representation Learning Using Support Vector Classifiers." pith.science (2026). https://pith.science/paper/COSGZ772

@misc{pith2026190902856,
  author       = {Pith},
  title        = {Pith review of: Discriminative Video Representation Learning Using Support Vector Classifiers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/COSGZ772}},
  note         = {Machine review of arXiv:1909.02856}
}
read the original abstract

Most popular deep models for action recognition in videos generate independent predictions for short clips, which are then pooled heuristically to assign an action label to the full video segment. As not all frames may characterize the underlying action---many are common across multiple actions---pooling schemes that impose equal importance on all frames might be unfavorable. In an attempt to tackle this problem, we propose discriminative pooling, based on the notion that among the deep features generated on all short clips, there is at least one that characterizes the action. To identify these useful features, we resort to a negative bag consisting of features that are known to be irrelevant, for example, they are sampled either from datasets that are unrelated to our actions of interest or are CNN features produced via random noise as input. With the features from the video as a positive bag and the irrelevant features as the negative bag, we cast an objective to learn a (nonlinear) hyperplane that separates the unknown useful features from the rest in a multiple instance learning formulation within a support vector machine setup. We use the parameters of this separating hyperplane as a descriptor for the full video segment. Since these parameters are directly related to the support vectors in a max-margin framework, they can be treated as a weighted average pooling of the features from the bags, with zero weights given to non-support vectors. Our pooling scheme is end-to-end trainable within a deep learning framework. We report results from experiments on eight computer vision benchmark datasets spanning a variety of video-related tasks and demonstrate state-of-the-art performance across these tasks.

Figures

Figures reproduced from arXiv: 1909.02856 by the authors.

Figure 1
Figure 1. A illustration of our discriminative pooling scheme [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. Our SVMP scheme/descriptor shares several properties of standard pooled descriptors, however also showcases several im￾portant advantages. For example, similar to other pooling schemes, SVM pooling results in a compact and fixed length representation of videos of arbitrary length. However differently, our pooling gives different weights to different features, and thus may be seen as a type of weighted average poolin… view at source ↗
Figure 2
Figure 2. Illustration of our SVM Pooling scheme. (i) Extracti [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (6 more)
Figure 3
Figure 3. Figure 3: An illustration of our overall idea. (a) the input dat [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 4
Figure 4. Figure 4: Two possible ways to insert SVM pooling layer within a [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Analysis of the parameters used in our scheme. All exp [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: T-SNE plots of positive (blue) and negative bags (red [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: T-SNE visualizations of SVMP and other pooling metho [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Visualizations of various pooled descriptors. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

95 extracted references · 75 canonical work pages

  1. [80]

    On differentiating parameterized argmin and argmax probl ems with application to bi-level optimization,

    S. Gould, B. Fernando, A. Cherian, P . Anderson, R. S. Cru z, and E. Guo, “On differentiating parameterized argmin and argmax probl ems with application to bi-level optimization,” 2016

    work page 2016

  2. [1]

    Deep residual learnin g for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learnin g for image recognition,” in CVPR, 2016, pp. 770–778

    2016

  3. [2]

    Mask R-C NN,

    K. He, G. Gkioxari, P . Doll´ ar, and R. Girshick, “Mask R-C NN,” in ICCV. IEEE, 2017

    2017

  4. [3]

    Quo vadis, action recogni tion? a new model and the kinetics dataset,

    J. Carreira and A. Zisserman, “Quo vadis, action recogni tion? a new model and the kinetics dataset,” in CVPR, July 2017

    2017

  5. [4]

    Spatiotempor al multiplier networks for video action recognition,

    C. Feichtenhofer, A. Pinz, and R. Wildes, “Spatiotempor al multiplier networks for video action recognition,” in CVPR, 2017

    2017

  6. [5]

    Deep b ilinear learning for rgb-d action recognition,

    J.-F. Hu, W.-S. Zheng, J. Pan, J. Lai, and J. Zhang, “Deep b ilinear learning for rgb-d action recognition,” in ECCV, September 2018

    2018

  7. [6]

    Tempora l relational reasoning in videos,

    B. Zhou, A. Andonian, A. Oliva, and A. Torralba, “Tempora l relational reasoning in videos,” ECCV, 2018

    2018

  8. [7]

    Two-stream convolutiona l networks for action recognition in videos,

    K. Simonyan and A. Zisserman, “Two-stream convolutiona l networks for action recognition in videos,” in NIPS, 2014

    2014

Show all 95 references
  1. [8]

    3D convolutional neural n etworks for human action recognition,

    S. Ji, W. Xu, M. Y ang, and K. Y u, “3D convolutional neural n etworks for human action recognition,” PAMI, vol. 35, no. 1, pp. 221–231, 2013

    2013

  2. [9]

    Moments in time dataset: one million videos for event understanding,

    M. Monfort, B. Zhou, S. A. Bargal, A. Andonian, T. Y an, K. R amakrish- nan, L. Brown, Q. Fan, D. Gutfruend, C. V ondrick et al. , “Moments in time dataset: one million videos for event understanding,” 2018

    2018

  3. [10]

    Convolut ional two-stream network fusion for video action recognition,

    C. Feichtenhofer, A. Pinz, and A. Zisserman, “Convolut ional two-stream network fusion for video action recognition,” in CVPR, 2016

    2016

  4. [11]

    V ery deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “V ery deep convolutional networks for large-scale image recognition,” arXiv preprint arXiv:1409.1556 , 2014

    2014 arXiv

  5. [12]

    Action recognition with t rajectory- pooled deep-convolutional descriptors,

    L. Wang, Y . Qiao, and X. Tang, “Action recognition with t rajectory- pooled deep-convolutional descriptors,” in CVPR, 2015

    2015

  6. [13]

    T wo-stream sr-cnns for action recognition in videos,

    Y . Wang, J. Song, L. Wang, L. V an Gool, and O. Hilliges, “T wo-stream sr-cnns for action recognition in videos,” in BMVC, 2016

    2016

  7. [14]

    Dynamic image networks for action recognition,

    H. Bilen, B. Fernando, E. Gavves, A. V edaldi, and S. Goul d, “Dynamic image networks for action recognition,” in CVPR, 2016

    2016

  8. [15]

    Modeling video evolution for action recognition,

    B. Fernando, E. Gavves, J. M. Oramas, A. Ghodrati, and T. Tuytelaars, “Modeling video evolution for action recognition,” in CVPR, 2015

    2015

  9. [16]

    Non-linea r temporal subspace representations for activity recognition,

    A. Cherian, S. Sra, S. Gould, and R. Hartley, “Non-linea r temporal subspace representations for activity recognition,” in CVPR, 2018

    2018

  10. [17]

    Gene ralized rank pooling for activity recognition,

    A. Cherian, B. Fernando, M. Harandi, and S. Gould, “Gene ralized rank pooling for activity recognition,” in CVPR, 2017

    2017

  11. [18]

    Second-order temporal poolin g for action recognition,

    A. Cherian and S. Gould, “Second-order temporal poolin g for action recognition,” arXiv preprint arXiv:1704.06925, 2017

    2017 arXiv

  12. [19]

    Higher-order poo ling of cnn features via kernel linearization for action recognition,

    A. Cherian, P . Koniusz, and S. Gould, “Higher-order poo ling of cnn features via kernel linearization for action recognition, ” in WACV, 2017

    2017

  13. [20]

    Temporal segment networks: Towards good practices for dee p action recognition,

    L. Wang, Y . Xiong, Z. Wang, Y . Qiao, D. Lin, X. Tang, and L. V an Gool, “Temporal segment networks: Towards good practices for dee p action recognition,” in ECCV, 2016

    2016

  14. [21]

    Action snippets: How many frames does human action recognition require?

    K. Schindler and L. V an Gool, “Action snippets: How many frames does human action recognition require?” in CVPR, 2008

    2008

  15. [22]

    Hmdb: a large video database for human motion recognition,

    H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre , “Hmdb: a large video database for human motion recognition,” in ICCV, 2011

    2011

  16. [23]

    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,” 2012

    2012

  17. [24]

    The kinetics human action video dataset,

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

    2017 arXiv

  18. [25]

    Hollywood in homes: Crowdsourcing data collecti on for activity understanding,

    G. A. Sigurdsson, G. V arol, X. Wang, A. Farhadi, I. Lapte v, and A. Gupta, “Hollywood in homes: Crowdsourcing data collecti on for activity understanding,” in ECCV, 2016

    2016

  19. [26]

    Action recognition based on a bag of 3d points,

    W. Li, Z. Zhang, and Z. Liu, “Action recognition based on a bag of 3d points,” in CVPRW, 2010

    2010

  20. [27]

    Ntu rgb+ d: A large scale dataset for 3d human activity analysis,

    A. Shahroudy, J. Liu, T.-T. Ng, and G. Wang, “Ntu rgb+ d: A large scale dataset for 3d human activity analysis,” in CVPR, 2016

    2016

  21. [28]

    At tribute and simile classifiers for face verification,

    N. Kumar, A. C. Berg, P . N. Belhumeur, and S. K. Nayar, “At tribute and simile classifiers for face verification,” in ICCV, 2009

    2009

  22. [29]

    Temporal res idual networks for dynamic scene recognition,

    C. Feichtenhofer, A. Pinz, and R. Wildes, “Temporal res idual networks for dynamic scene recognition,” in CVPR, 2017

    2017

  23. [30]

    Going deeper int o action recog- nition: A survey,

    S. Herath, M. Harandi, and F. Porikli, “Going deeper int o action recog- nition: A survey,” Image and vision computing , vol. 60, pp. 4–21, 2017. TRANSACTIONS ON P A TTERN ANAL YSIS AND MACHINE INTELLIGENCE 12

    2017

  24. [31]

    A survey on vision-based human action recogn ition,

    R. Poppe, “A survey on vision-based human action recogn ition,” Image and vision computing , vol. 28, no. 6, pp. 976–990, 2010

    2010

  25. [32]

    Human activity analysis: A review,

    J. K. Aggarwal and M. S. Ryoo, “Human activity analysis: A review,” ACM Computing Surveys (CSUR) , vol. 43, no. 3, p. 16, 2011

    2011

  26. [33]

    Action re cognition by dense trajectories,

    H. Wang, A. Kl¨ aser, C. Schmid, and C.-L. Liu, “Action re cognition by dense trajectories,” in CVPR, 2011

    2011

  27. [34]

    Action recognition with improve d trajectories,

    H. Wang and C. Schmid, “Action recognition with improve d trajectories,” in ICCV, 2013

    2013

  28. [35]

    Action bank: A high-level r epresentation of activity in video,

    S. Sadanand and J. J. Corso, “Action bank: A high-level r epresentation of activity in video,” in CVPR, 2012

    2012

  29. [36]

    Video Google: A text retriev al approach to object matching in videos,

    J. Sivic and A. Zisserman, “Video Google: A text retriev al approach to object matching in videos,” in ICCV, 2003, p. 1470

    2003

  30. [37]

    Improving t he fisher kernel for large-scale image classification,

    F. Perronnin, J. S´ anchez, and T. Mensink, “Improving t he fisher kernel for large-scale image classification,” in ECCV, 2010

    2010

  31. [38]

    Aggregating local image descriptors into compact codes,

    H. Jegou, F. Perronnin, M. Douze, J. S´ anchez, P . Perez, and C. Schmid, “Aggregating local image descriptors into compact codes,” TPAMI, vol. 34, no. 9, pp. 1704–1716, 2012

    2012

  32. [39]

    Bag of visual word s and fusion methods for action recognition: Comprehensive stud y and good practice,

    X. Peng, L. Wang, X. Wang, and Y . Qiao, “Bag of visual word s and fusion methods for action recognition: Comprehensive stud y and good practice,” Computer Vision and Image Understanding, vol. 150, pp. 109– 125, 2016

    2016

  33. [40]

    Imagene t classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagene t classification with deep convolutional neural networks,” in NIPS, 2012

    2012

  34. [41]

    Deep reconstruction models for image set classification,

    M. Hayat, M. Bennamoun, and S. An, “Deep reconstruction models for image set classification,” PAMI, vol. 37, no. 4, pp. 713–727, 2015

    2015

  35. [42]

    Interpretable 3D human action a nalysis with temporal convolutional networks,

    T. S. Kim and A. Reiter, “Interpretable 3D human action a nalysis with temporal convolutional networks,” 2017

    2017

  36. [43]

    Sequen- tial deep learning for human action recognition,

    M. Baccouche, F. Mamalet, C. Wolf, C. Garcia, and A. Bask urt, “Sequen- tial deep learning for human action recognition,” in Human Behavior Understanding, 2011, pp. 29–39

    2011

  37. [44]

    Long-term recurrent co nvolutional networks for visual recognition and description,

    J. Donahue, L. Anne Hendricks, S. Guadarrama, M. Rohrba ch, S. V enu- gopalan, K. Saenko, and T. Darrell, “Long-term recurrent co nvolutional networks for visual recognition and description,” in CVPR, 2015

    2015

  38. [45]

    Hierarchical recurrent neu ral network for skeleton based action recognition,

    Y . Du, W. Wang, and L. Wang, “Hierarchical recurrent neu ral network for skeleton based action recognition,” in CVPR, 2015

    2015

  39. [46]

    Action re cognition by learning deep multi-granular spatio-temporal video rep resentation,

    Q. Li, Z. Qiu, T. Y ao, T. Mei, Y . Rui, and J. Luo, “Action re cognition by learning deep multi-granular spatio-temporal video rep resentation,” in ICMR, 2016

    2016

  40. [47]

    Uns upervised learn- ing of video representations using lstms

    N. Srivastava, E. Mansimov, and R. Salakhutdinov, “Uns upervised learn- ing of video representations using lstms.” in ICML, 2015, pp. 843–852

    2015

  41. [48]

    Beyond short snippets: Deep netw orks for video classification,

    J. Y ue-Hei Ng, M. Hausknecht, S. Vijayanarasimhan, O. V inyals, R. Monga, and G. Toderici, “Beyond short snippets: Deep netw orks for video classification,” in CVPR, 2015

    2015

  42. [49]

    On the difficulty of training recurrent neural networks,

    R. Pascanu, T. Mikolov, and Y . Bengio, “On the difficulty of training recurrent neural networks,” in ICML, 2013

    2013

  43. [50]

    Learning spatiotemporal features with 3D convolutional networks,

    D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Palu ri, “Learning spatiotemporal features with 3D convolutional networks,” in ICCV, 2015

    2015

  44. [51]

    Non-local neu ral networks,

    X. Wang, R. Girshick, A. Gupta, and K. He, “Non-local neu ral networks,” in CVPR, 2018

    2018

  45. [52]

    Mict: Mixed 3d/2 d convolu- tional tube for human action recognition,

    Y . Zhou, X. Sun, Z.-J. Zha, and W. Zeng, “Mict: Mixed 3d/2 d convolu- tional tube for human action recognition,” in CVPR, 2018

    2018

  46. [53]

    Appearance-and-r elation networks for video classification,

    L. Wang, W. Li, W. Li, and L. V an Gool, “Appearance-and-r elation networks for video classification,” 2017

    2017

  47. [54]

    Second-order temporal poolin g for action recognition,

    A. Cherian and S. Gould, “Second-order temporal poolin g for action recognition,” IJCV, 2018

    2018

  48. [55]

    Action- VLAD: Learning spatio-temporal aggregation for action cla ssification,

    R. Girdhar, D. Ramanan, A. Gupta, J. Sivic, and B. Russel l, “Action- VLAD: Learning spatio-temporal aggregation for action cla ssification,” in CVPR, 2017

    2017

  49. [56]

    Dynamic pooling fo r complex event recognition,

    J. Wang, A. Cherian, and F. Porikli, “Dynamic pooling fo r complex event recognition,” in WACV, 2017

    2017

  50. [57]

    Ensemble of e xemplar-svms for object detection and beyond,

    T. Malisiewicz, A. Gupta, and A. A. Efros, “Ensemble of e xemplar-svms for object detection and beyond,” in ICCV, 2011

    2011

  51. [58]

    Exemplar- based action recognition in video

    G. Willems, J. H. Becker, T. Tuytelaars, and L. J. V an Goo l, “Exemplar- based action recognition in video.” in BMVC, 2009

    2009

  52. [59]

    Exemplar svms as visual feature encoders,

    J. Zepeda and P . Perez, “Exemplar svms as visual feature encoders,” in CVPR, 2015

    2015

  53. [60]

    Weakly supervi sed object localization with multi-fold multiple instance learning,

    R. G. Cinbis, J. V erbeek, and C. Schmid, “Weakly supervi sed object localization with multi-fold multiple instance learning, ” PAMI, vol. 39, no. 1, pp. 189–203, 2017

    2017

  54. [61]

    Multiple instance learning f or soft bags via top instances,

    W. Li and N. V asconcelos, “Multiple instance learning f or soft bags via top instances,” in CVPR, 2015

    2015

  55. [62]

    Deep multiple instance learning for image classification and auto-annotation,

    J. Wu, Y . Y u, C. Huang, and K. Y u, “Deep multiple instance learning for image classification and auto-annotation,” in CVPR, 2015

    2015

  56. [63]

    Human action recognition with graph-b ased multiple- instance learning,

    Y . Yi and M. Lin, “Human action recognition with graph-b ased multiple- instance learning,” Pattern Recognition, vol. 53, pp. 148–162, 2016

    2016

  57. [64]

    A s elf- paced multiple-instance learning framework for co-salien cy detection,

    D. Zhang, D. Meng, C. Li, L. Jiang, Q. Zhao, and J. Han, “A s elf- paced multiple-instance learning framework for co-salien cy detection,” in ICCV, 2015

    2015

  58. [65]

    Modeling the temporal extent o f actions,

    S. Satkin and M. Hebert, “Modeling the temporal extent o f actions,” in ECCV, 2010

    2010

  59. [66]

    Discriminative sub sequence mining for action classification,

    S. Nowozin, G. Bakir, and K. Tsuda, “Discriminative sub sequence mining for action classification,” in ICCV, 2007

    2007

  60. [67]

    Dynamic pooling for complex event recognition,

    W. Li, Q. Y u, A. Divakaran, and N. V asconcelos, “Dynamic pooling for complex event recognition,” in ICCV, 2013

    2013

  61. [68]

    Discover: Discovering importan t segments for classification of video events and recounting,

    C. Sun and R. Nevatia, “Discover: Discovering importan t segments for classification of video events and recounting,” in CVPR, 2014

    2014

  62. [69]

    Compos itional models for video event detection: A multiple kernel learnin g latent variable approach,

    A. V ahdat, K. Cannons, G. Mori, S. Oh, and I. Kim, “Compos itional models for video event detection: A multiple kernel learnin g latent variable approach,” in ICCV, 2013

    2013

  63. [70]

    Multi-instance kernels

    T. G¨ artner, P . A. Flach, A. Kowalczyk, and A. J. Smola, “ Multi-instance kernels.” in ICML, 2002

    2002

  64. [71]

    Video e vent detection by inferring temporal instance labels,

    K.-T. Lai, F. X. Y u, M.-S. Chen, and S.-F. Chang, “Video e vent detection by inferring temporal instance labels,” in CVPR, 2014

    2014

  65. [72]

    propto svm for learning with label proportions,

    F. X. Y u, D. Liu, S. Kumar, T. Jebara, and S.-F. Chang, “ propto svm for learning with label proportions,” arXiv preprint arXiv:1306.0886, 2013

    2013 arXiv

  66. [73]

    Video rep resentation learning using discriminative pooling,

    J. Wang, A. Cherian, F. Porikli, and S. Gould, “Video rep resentation learning using discriminative pooling,” in CVPR, 2018

    2018

  67. [74]

    Multiple instance learn ing for sparse positive bags,

    R. C. Bunescu and R. J. Mooney, “Multiple instance learn ing for sparse positive bags,” in ICML, 2007

    2007

  68. [75]

    Mixed-integer quadratic programming,

    R. Lazimy, “Mixed-integer quadratic programming,” Mathematical Pro- gramming, vol. 22, no. 1, pp. 332–349, 1982

    1982

  69. [76]

    A. J. Smola and B. Sch¨ olkopf, Learning with kernels . Citeseer, 1998

    1998

  70. [77]

    Efficient additive kernel s via explicit feature maps,

    A. V edaldi and A. Zisserman, “Efficient additive kernel s via explicit feature maps,” PAMI, vol. 34, no. 3, pp. 480–492, 2012

    2012

  71. [78]

    Ma nopt, a Matlab toolbox for optimization on manifolds,

    N. Boumal, B. Mishra, P .-A. Absil, and R. Sepulchre, “Ma nopt, a Matlab toolbox for optimization on manifolds,” Journal of Machine Learning Research, vol. 15, pp. 1455–1459, 2014

    2014

  72. [79]

    Implicit functio ns and solution mappings,

    A. L. Dontchev and R. T. Rockafellar, “Implicit functio ns and solution mappings,” Springer Monogr . Math., 2009

    2009

  73. [81]

    Asynchronous temporal fields for action recognition,

    G. A. Sigurdsson, S. Divvala, A. Farhadi, and A. Gupta, “ Asynchronous temporal fields for action recognition,” in CVPR, 2017

    2017

  74. [82]

    Human act ion recognition by representing 3d skeletons as points in a lie group,

    R. V emulapalli, F. Arrate, and R. Chellappa, “Human act ion recognition by representing 3d skeletons as points in a lie group,” in CVPR, 2014

    2014

  75. [83]

    Do We Really Need to Collect Millions of Faces for Effective Face Recogni tion?

    I. Masi, A. Tran, T. Hassner, J. T. Leksut, and G. Medioni , “Do We Really Need to Collect Millions of Faces for Effective Face Recogni tion?” in ECCV, 2016

    2016

  76. [84]

    Inc eption-v4, inception-resnet and the impact of residual connections on learning

    C. Szegedy, S. Ioffe, V . V anhoucke, and A. A. Alemi, “Inc eption-v4, inception-resnet and the impact of residual connections on learning.” in AAAI, 2017

    2017

  77. [85]

    Activ- itynet: A large-scale video benchmark for human activity un derstanding,

    F. Caba Heilbron, V . Escorcia, B. Ghanem, and J. Carlos N iebles, “Activ- itynet: A large-scale video benchmark for human activity un derstanding,” in CVPR, 2015

    2015

  78. [86]

    Spatiotempo ral residual networks for video action recognition,

    C. Feichtenhofer, A. Pinz, and R. Wildes, “Spatiotempo ral residual networks for video action recognition,” in NIPS, 2016

    2016

  79. [87]

    Caffe: Convolutional archi tecture for fast feature embedding

    Y . Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell, “Caffe: Convolutional archi tecture for fast feature embedding.” ACM, 2014

    2014

  80. [88]

    A short note about kinetics-600,

    J. Carreira, E. Noland, A. Banki-Horvath, C. Hillier, a nd A. Zisserman, “A short note about kinetics-600,” 2018

    2018

  81. [89]

    Adascan: Adapti ve scan pool- ing in deep convolutional neural networks for human action r ecognition in videos,

    A. Kar, N. Rai, K. Sikka, and G. Sharma, “Adascan: Adapti ve scan pool- ing in deep convolutional neural networks for human action r ecognition in videos,” in CVPR, 2017

    2017

  82. [90]

    Disc riminative hierarchical rank pooling for activity recognition,

    B. Fernando, P . Anderson, M. Hutter, and S. Gould, “Disc riminative hierarchical rank pooling for activity recognition,” in CVPR, 2016

    2016

  83. [91]

    Deep inside convolutional networks: Visualising image classification models and sali ency maps,

    K. Simonyan, A. V edaldi, and A. Zisserman, “Deep inside convolutional networks: Visualising image classification models and sali ency maps,” 2013

    2013

  84. [92]

    Actionvlad: Learning spatio-temporal aggregation for action classific ation,

    R. Girdhar, D. Ramanan, A. Gupta, J. Sivic, and B. Russel l, “Actionvlad: Learning spatio-temporal aggregation for action classific ation,” in CVPR, 2017

    2017

  85. [93]

    Skele ton-based action recognition using spatio-temporal lstm network wit h trust gates,

    J. Liu, A. Shahroudy, D. Xu, A. C. Kot, and G. Wang, “Skele ton-based action recognition using spatio-temporal lstm network wit h trust gates,” 2017

    2017

  86. [94]

    Empowering simp le binary classifiers for image set based face recognition,

    M. Hayat, S. H. Khan, and M. Bennamoun, “Empowering simp le binary classifiers for image set based face recognition,” IJCV, pp. 1–20, 2017

    2017

  87. [95]

    Dense tra jectories and motion boundary descriptors for action recognition,

    H. Wang, A. Kl¨ aser, C. Schmid, and C.-L. Liu, “Dense tra jectories and motion boundary descriptors for action recognition,” IJCV, vol. 103, no. 1, pp. 60–79, 2013. TRANSACTIONS ON P A TTERN ANAL YSIS AND MACHINE INTELLIGENCE 13 Jue Wang is a PhD student with the Research Sch...

    2013

Pith tools

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