Pith. sign in

REVIEW 3 major objections 5 minor 57 references

Dual Guidance Semi-Supervised Action Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper argues that semi-supervised spatio-temporal action localization improves when a global frame-level classifier filters pseudo-bounding boxes by requiring class agreement between frames and boxes.

desk verdict A plausible SSL extension for spatio-temporal action detection, but the core dual-guidance filter is never isolated from auxiliary losses and thresholds are tuned on test, so the central claim remains unproven. read the letter →

arxiv 2507.21247 v1 pith:LV2GQGB2 submitted 2025-07-28 cs.CV

Ankit Singh , Efstratios Gavves , Cees G. M. Snoek , Hilde Kuehne This is my paper
classification cs.CV
keywords semi-supervisedlearningspatio-temporalactionlocalizationdetectionpseudo-labelingframe-levelclassificationbounding-boxteacher-studentvideounderstanding
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that semi-supervised learning can be extended from image tasks to spatio-temporal action localization by adding a global frame-level classifier that gates the pseudo-bounding boxes produced by a box head. The proposed dual guidance network accepts a candidate pseudo-box only when its predicted action class matches the frame-level prediction, after both heads pass confidence thresholds. Evaluated with 5-30% labeled data on UCF101-24, J-HMDB-21, and AVA, the approach reports consistent gains over supervised training and over video-domain extensions of the image SSL baselines FixMatch and Mean Teacher. The paper also reports that unlabeled background frames and a temporal consistency loss each contribute to the improvement. If the claim holds, global context is a useful filter for noisy local pseudo-labels in video detection.

What carries the argument

The dual guidance network: one video backbone $G$ feeding two heads, a frame-level classification head $F$ (global) and a YOLO-style bounding-box head $P$ (local). In the semi-supervised stage, weakly augmented clips go to a teacher model and strongly augmented clips to a student model; $F$ produces a pseudo-label when $\max(F(G(U^i))) > F_{th}$, and $P$ produces candidate boxes via NMS with confidence above $O_{th}$. Only boxes whose class matches the pseudo-label become pseudo-bounding boxes, and the student's box head is trained against them with the unsupervised box loss $L^u_{bou}$. The temporal consistency term $L_{tmp} = |F(G(V^i_s)) - F(G(V^i_{s+1}))|$ encourages adjacent frames to agree at the frame level. The mechanism's job is to filter noisy local proposals using global context that the box head cannot see on its own.

What would settle it

Run the same teacher-student training with the class-matching filter disabled, keeping only the box confidence threshold $O_{th}$, and compare video-mAP. The central claim is falsified if removing the filter does not reduce performance, or if the frame classifier's pseudo-labels agree with ground-truth classes on a small labeled validation subset at near-chance rates.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a frame-level action classifier, trained jointly with a bounding-box head, can act as a reliable class-consistency gate for pseudo-labels in semi-supervised spatio-temporal action localization. In the teacher-student stage, the teacher's weakly augmented frame is passed to both heads: the classifier's argmax becomes the pseudo action label when its maximum probability exceeds $F_{th}$, and candidate boxes whose confidence exceeds $O_{th}$ are kept as pseudo-bounding boxes only when their class matches. The student is trained on strongly augmented inputs to match these targets, and a temporal loss enforces agreement between adjacent frames' frame-level predictions. The reported results include 9.4 video-mAP@0.5 on UCF101-24 with 10% labeled data and no pretraining, versus 5.4 for supervised training and 5.1 for the Mean Teacher extension, and 34.8 on J-HMDB-21 with 30% labeled data versus 24.6 for the best baseline on that setting.

Load-bearing premise

The load-bearing premise is that the frame-level classifier, trained on only a small labeled set, is accurate enough on unlabeled frames that requiring a candidate box's class to match the frame prediction improves pseudo-box selection; if the classifier is confidently wrong, it can reject correct boxes and amplify its own errors.

Editorial extensions

If this is right

  • Semi-supervised video action detection can draw on global context to clean box-level pseudo-labels, reducing error accumulation without any extra human annotation.
  • The dual-guidance recipe should transfer to other single-stage or two-stage video detectors that have a box head and a shared-feature global classifier.
  • Threshold behavior reported in the paper implies that practitioners can tune the frame-level gate $F_{th}$ as a new lever for SSL quality, with $F_{th}=0.8$ performing best and saturating beyond it.
  • Unlabeled background frames are a useful training resource: trimming them drops video-mAP@0.5 from 9.4 to 6.1 at 10% labeled UCF101-24, so negative sampling matters.
  • Temporal consistency of frame-level predictions adds a small but consistent gain, making frame smoothness a cheap extra supervision signal for video SSL.

Reading between the lines

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

  • A direct ablation the paper does not report: remove only the class-matching filter while keeping both heads trained. If performance stays flat, the improvement comes from the extra classification task, not from the gating mechanism itself.
  • The frame/box agreement rule is most natural when a clip has one dominant action; the smaller gains on the multi-label AVA dataset are consistent with that reading, and a per-class thresholded gate might close the gap.
  • The global classifier could double as a pseudo-label quality monitor: tracking its confidence on unlabeled frames may signal when the student is drifting, since the same head gates the pseudo-boxes.
  • Because the gate compares class labels rather than IoU quality, the mechanism likely helps precision more than recall; that could be tested by decomposing accepted pseudo-boxes into true and false positives against ground truth on a held-out labeled subset.
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 a semi-supervised framework for spatial-temporal action detection. A YOWO-style detector is trained on a small labeled set and a large unlabeled set; on unlabeled clips, a frame-level classification head and a bounding-box head jointly decide which candidate boxes become pseudo-ground-truth boxes. The objective combines supervised losses, an unsupervised frame-classification loss, unsupervised box losses, and a temporal consistency term. Experiments on UCF101-24, J-HMDB-21, and AVA report improvements over extended FixMatch and Mean-Teacher baselines at low label fractions, plus ablations of negative sampling and temporal consistency.

Significance. If the reported gains hold, the paper would be a useful step for semi-supervised video action detection, a setting that is less explored than image SSL. The authors make a reasonable design choice to add a global frame-level classifier and use agreement with local box predictions to filter pseudo-boxes, and they provide a classwise analysis that helps understand where gains occur. The negative-sampling ablation and temporal-consistency ablation are useful diagnostics. However, the central mechanism is not isolated: the paper never ablates the frame-level filter while keeping the additional frame-classification loss and temporal loss, and the threshold analysis in Fig. 4 selects hyperparameters on test video-mAP. As a result, the contribution of the dual-guidance mechanism itself is not yet established. The paper also reports no error bars or multiple-seed results, and one J-HMDB-21 no-pretrain result is non-monotonic in label fraction, which suggests variance that a single run cannot adjudicate.

major comments (3)
  1. [§3.3 and §4.5] The central contribution is the dual-guidance selection rule in Sec. 3.3: a candidate pseudo-box is retained only when the box-head class prediction argmax(P(G(U^i))_c) matches the frame-level pseudo-label argmax(F(G(U^i))) and both confidence thresholds are exceeded. However, the ablations in Sec. 4.5 do not isolate this rule. Table 5 removes negative sampling and Table 6 removes the temporal loss, but no experiment removes the frame-level agreement filter while keeping L^u_frame and the temporal term. Since the full objective (Eq. 11) differs from the FixMatch/Mean-Teacher baselines in several ways at once, the reported gains cannot be attributed specifically to dual guidance. I recommend adding an ablation that disables only the agreement filter (e.g., using all boxes above Oth as pseudo-boxes) and reporting that variant.
  2. [§4.5.2 and Fig. 4] The thresholds Fth and Oth are selected by maximizing test video-mAP, as shown in Fig. 4a and 4b. This leaks test-set information into the reported numbers, so part of the margin over baselines may reflect fitting to the evaluation set rather than the method itself. Please report either fixed thresholds, thresholds selected on a validation split, or a sensitivity analysis that does not use the test set. In addition, the lack of multiple-seed runs is problematic: Table 3 shows J-HMDB-21 no-pretrain results that decrease when the label fraction increases from 20% to 30% (V@0.1 14.1 vs. 10.9, V@0.2 12.6 vs. 9.8). This non-monotonic behavior suggests high run-to-run variance, and single runs cannot determine whether the reported improvements are significant.
  3. [§3.3 and §4.1 (AVA)] The AVA dataset is multi-action per frame and the frame-level head uses sigmoid with weighted binary cross-entropy (Sec. 4.2), yet the pseudo-labeling rule in Sec. 3.3 uses argmax over classes to form yhat_u and to match the box class. Argmax is not a well-defined pseudo-label for a multi-label per-frame setting. The paper should clarify how the dual-guidance selection is adapted to AVA, or the AVA results should be interpreted with caution. As written, the mechanism described in Sec. 3.3 does not directly support Table 4.
minor comments (5)
  1. [Eq. (7)] The summation index in Eq. (7) is written as b=1 over B, but the terms use i and B; the notation should be made consistent (e.g., sum over i=1..B) and the indicator function should be written with proper argument parentheses.
  2. [Eq. (5)] The frame-level loss uses g(V^i) without defining g; it should be F(G(V^i)) to match the surrounding notation.
  3. [Fig. 1 and Fig. 2] The figures include a label "CFAM" and "3D Model/2D Model" that are never explained in the text; please define these elements in the caption or in Sec. 3.
  4. [Table 3] The row "Fully-Supervised [17]" is formatted as "- 87.8 85.7", which is ambiguous. Please clarify which columns these values refer to.
  5. [Sec. 4.2] The implementation details state that a cosine classifier with two linear layers is used for the frame-level head; please specify dimensions, normalization, and whether this differs from the standard linear head used in YOWO.

Circularity Check

1 steps flagged · score 4.0 of 10

Test-set threshold selection makes the reported margins partially fitted, but the dual-guidance method itself is not derived from its inputs.

  1. fitted input called prediction [Section 4.5.2 (hyper-parameter analysis) and Figure 4 caption]
    "Experimental results show the increase in video-mAP@0.5 values from 6.2 to 9.4 as the frame-level classification threshold increases from0.6 to 0.8. The performance starts to saturate on further increase of the threshold. We used the bounding-box threshold of 0.4 for our approach, which achieves maximum value as depicted in the figure 4(b). Figure 4 caption: (a) Effect of varying the frame-level classification threshold on test video-mAP values (b) Effect of varying Bounding Box threshold on test video-mAP values."

    The thresholds Fth and Oth are selected by sweeping the very metric that is later reported as the paper's headline result (test video-mAP), and Oth=0.4 is explicitly chosen because it 'achieves maximum value' on that test metric. Consequently, the final 'Ours' numbers in Tables 2-3 are not independent predictions but the result of a selection procedure on the evaluation set. The dual-guidance rule itself is not defined in terms of the test metric, so this is a partial fit-to-evaluation rather than a fully self-definitional derivation.

full rationale

The paper's core derivation chain (Secs. 3.2-3.5) is largely self-contained: the frame-level head F and box head P are trained with standard supervised losses; pseudo-labels come from a teacher/student FixMatch-style consistency procedure; and the dual-guidance filter gates pseudo-boxes by requiring class agreement between argmax(F(G(U^i))) and the box head's class scores. None of these losses is constructed from test video-mAP, and the method is not justified by a load-bearing self-citation chain. The one concrete fit-to-evaluation step is in Sec. 4.5.2 / Fig. 4, where Fth and Oth are swept on test video-mAP and Oth=0.4 is selected as the maximum of that curve; the reported Ours results then re-use the same evaluation metric. This means part of the reported margin may be an artifact of test-set hyperparameter selection rather than an independent prediction, but the central claim that frame-level guidance helps select pseudo-boxes is not forced by construction. The absence of an ablation isolating the global filter is an attribution gap, not circularity.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The method relies on two hand-selected confidence thresholds for pseudo-labeling, tuned on the test set, plus standard loss weights. No new entities are postulated.

free parameters (7)
  • Fth = 0.8
    Frame-level classification threshold for pseudo-label generation; selected by maximizing test video-mAP (Fig 4a).
  • Oth = 0.4
    Bounding-box confidence threshold for pseudo-box selection; selected by maximizing test video-mAP (Fig 4b).
  • alpha = 0.5
    Weight for coordinate and confidence losses in Eq (4) and Eq (8); set by hand.
  • beta = 1
    Weight for frame classification loss in Eq (6); set by hand.
  • gamma = 2 (0.5 for AVA)
    Focal loss modulating factor; chosen per dataset.
  • eta = 0.5
    Huber loss slope parameter in Eq (1); set by hand.
  • delta = 1
    Weight for unsupervised frame loss in Eq (9); set by hand.
assumptions (3)
  • domain assumption The action category of a video clip is constant across adjacent frames; used in temporal loss Eq (10).
    Assumed global action category remains the same across consecutive frames; likely true for single-action clips but not for multi-action AVA or fast transitions.
  • domain assumption A global frame-level classifier can validate the class of a local bounding-box proposal; used in pseudo-box selection in Sec 3.3.
    The paper assumes agreement between frame-level and box-level class predictions indicates a correct pseudo-box; no evidence is given for reliability at low label ratios.
  • domain assumption Teacher and student models can share weights with different augmentations (no EMA required); used in Sec 3.3 and Algorithm 1.
    The paper cites Mean-Teacher but does not state whether the teacher is an EMA copy; the algorithm suggests weights are shared.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual Guidance Semi-Supervised Action Detection." pith.science (2026). https://pith.science/paper/LV2GQGB2

@misc{pith2026250721247,
  author       = {Pith},
  title        = {Pith review of: Dual Guidance Semi-Supervised Action Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LV2GQGB2}},
  note         = {Machine review of arXiv:2507.21247}
}
read the original abstract

Semi-Supervised Learning (SSL) has shown tremendous potential to improve the predictive performance of deep learning models when annotations are hard to obtain. However, the application of SSL has so far been mainly studied in the context of image classification. In this work, we present a semi-supervised approach for spatial-temporal action localization. We introduce a dual guidance network to select better pseudo-bounding boxes. It combines a frame-level classification with a bounding-box prediction to enforce action class consistency across frames and boxes. Our evaluation across well-known spatial-temporal action localization datasets, namely UCF101-24 , J-HMDB-21 and AVA shows that the proposed module considerably enhances the model's performance in limited labeled data settings. Our framework achieves superior results compared to extended image-based semi-supervised baselines.

Figures

Figures reproduced from arXiv: 2507.21247 by the authors.

Figure 1
Figure 1. Overview of the Supervised Training: The input clip is passed through G to extract features, which are then processed through F and P where F is the frame-level classification head and P is the bounding box head. The prediction from F is trained against the ground-truth action class while the bounding boxes predicted through P are trained against the labeled ground truth bounding box. problem and notation used in ou… view at source ↗
Figure 2
Figure 2. Outline of our Framework: Our framework consists of aligning the outputs of the network at two levels. On the global level, if the frame-level prediction is greater than Fth, then one hot pseudo-label is generated, and prediction on strongly augmented input is ensured to match the one-hot pseudo-label using cross-entropy loss. On the local level, the features from G(.) are passed to P(.) generate the bounding boxes … view at source ↗
Figure 3
Figure 3. Change in Video-mAP of our framework over su￾pervised model on UCF101-24 10% labeled data. Blue bars show the change in video-mAP values on different action cate￾gories (sorted according to the number of training videos) of the UCF101-24 dataset under 10% labeled scenario. Compared to Su￾pervised model, our framework improves the Video-mAP perfor￾mance of most classes in the UCF101-24 dataset. (Best viewed in color.… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Qualitative Examples of our approach on UCF101-24 test dataset Blue colored box denotes the ground-truth while the green box denotes the bounding-box predicted by our framework at different frames of the video. Last row shows the failure case of our models. Our model i…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 52 canonical work pages

  1. [1]

    Remixmatch: Semi-Supervised Learning with Distribution Matching and Augmentation Anchoring

    David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-Supervised Learning with Distribution Matching and Augmentation Anchoring. In International Conference on Learning Representa- tions, 2019. 1, 2

    work page 2019

  2. [2]

    Mixmatch: A Holistic Approach to Semi-Supervised Learning

    David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A Holistic Approach to Semi-Supervised Learning. In Neural Information Processing Systems, pages 5050–5060, 2019. 1, 2

    work page 2019

  3. [3]

    St-hoi: A spatial-temporal baseline for human-object interaction detection in videos

    Meng-Jiun Chiou, Chun-Yu Liao, Li-Wei Wang, Roger Zimmermann, and Jiashi Feng. St-hoi: A spatial-temporal baseline for human-object interaction detection in videos. Proceedings of the 2021 Work- shop on Intelligent Cross-Data Analysis and Retrieval,

    work page 2021

  4. [4]

    Victor Escorcia, Cuong Duc Dao, Mihir Jain, Bernard Ghanem, and Cees G. M. Snoek. Guess where? actor-supervision for spatiotemporal action localiza- tion. ArXiv, abs/1804.01824, 2020. 2

    work page Pith review arXiv 2020

  5. [5]

    SlowFast Networks for Video Recognition

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Ma- lik, and Kaiming He. SlowFast Networks for Video Recognition. In IEEE International Conference on Computer Vision, pages 6202–6211, 2019. 1

    work page 2019

  6. [6]

    Instant-teaching: An end-to-end semi-supervised object detection framework

    Qiang feng Zhou, Chaohui Yu, Zhibin Wang, Qi Qian, and Hao Li. Instant-teaching: An end-to-end semi-supervised object detection framework. 2021 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 4079–4088, 2021. 1, 2, 6

    work page 2021

  7. [7]

    Devnet: A deep event net- work for multimedia event detection and evidence re- counting

    Chuang Gan, Naiyan Wang, Yi Yang, Dit-Yan Yeung, and Alexander Hauptmann. Devnet: A deep event net- work for multimedia event detection and evidence re- counting. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 2568–2577,

    work page 2015

  8. [8]

    Kirill Gavrilyuk, Mihir Jain, Ilia Karmanov, and Cees G. M. Snoek. Motion-augmented self-training for video recognition at smaller scale. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10409–10418, 2021. 1

    work page 2021

Show all 57 references
  1. [9]

    Video action transformer network

    Rohit Girdhar, Jo ˜ao Carreira, Carl Doersch, and An- drew Zisserman. Video action transformer network. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 244–253, 2019. 2

    2019

  2. [10]

    Fast r-cnn

    Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision , pages 1440–1448, 2015. 3

    2015

  3. [11]

    Finding action tubes

    Georgia Gkioxari and Jitendra Malik. Finding action tubes. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 759–768, 2015. 2, 6

    2015

  4. [12]

    Ross, George Toderici, Yeqing Li, Susanna Ricco, Rahul Sukthankar, Cordelia Schmid, and Jitendra Malik

    Chunhui Gu, Chen Sun, Sudheendra Vijaya- narasimhan, Caroline Pantofaru, David A. Ross, George Toderici, Yeqing Li, Susanna Ricco, Rahul Sukthankar, Cordelia Schmid, and Jitendra Malik. Ava: A video dataset of spatio-temporally localized atomic visual actions. 2018 IEEE/CVF Co...

    2018

  5. [13]

    Tube convo- lutional neural network (t-cnn) for action detection in videos

    Rui Hou, Chen Chen, and Mubarak Shah. Tube convo- lutional neural network (t-cnn) for action detection in videos. 2017 IEEE International Conference on Com- puter Vision (ICCV), pages 5823–5832, 2017. 2

    2017

  6. [14]

    Consistency-based semi-supervised learning for object detection

    Jisoo Jeong, Seungeui Lee, Jeesoo Kim, and Nojun Kwak. Consistency-based semi-supervised learning for object detection. In NeurIPS, 2019. 2

    2019

  7. [15]

    Towards understanding action recognition

    Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE inter- national conference on computer vision, pages 3192– 3199, 2013. 5

    2013

  8. [16]

    Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid

    Vicky S. Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid. Action tubelet detector for spatio-temporal action localization. 2017 IEEE In- ternational Conference on Computer Vision (ICCV) , pages 4415–4423, 2017. 1, 2

    2017

  9. [17]

    You only watch once: A unified cnn architecture for real-time spatiotemporal action localization

    Okan K ¨op¨ukl¨u, Xiangyu Wei, and Gerhard Rigoll. You only watch once: A unified cnn architecture for real-time spatiotemporal action localization. ArXiv, abs/1911.06644, 2019. 1, 3, 5, 6, 7

    1911 arXiv

  10. [18]

    Kuehne, H

    H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. HMDB: a large video database for human mo- tion recognition. In Proceedings of the International Conference on Computer Vision (ICCV), 2011. 5

    2011

  11. [19]

    Recurrent tubelet proposal and recognition networks for action detection

    Dong Li, Zhaofan Qiu, Qi Dai, Ting Yao, and Tao Mei. Recurrent tubelet proposal and recognition networks for action detection. In ECCV, 2018. 2

    2018

  12. [20]

    Actions as moving points

    Yixuan Li, Zixu Wang, Limin Wang, and Gangshan Wu. Actions as moving points. In ECCV, 2020. 1, 2

    2020

  13. [21]

    Zhenyang Li, Kirill Gavrilyuk, Efstratios Gavves, Mi- hir Jain, and Cees G. M. Snoek. Videolstm convolves, attends and flows for action recognition. ArXiv, abs/1607.01794, 2018. 2

    2018 arXiv

  14. [22]

    TSM: Temporal Shift Module for Efficient Video Understanding

    Ji Lin, Chuang Gan, and Song Han. TSM: Temporal Shift Module for Efficient Video Understanding. In IEEE International Conference on Computer Vision , pages 7083–7093, 2019. 1

    2019

  15. [23]

    Girshick, Kaim- ing He, Bharath Hariharan, and Serge J

    Tsung-Yi Lin, Piotr Doll ´ar, Ross B. Girshick, Kaim- ing He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 936–944, 2017. 1 9

    2017

  16. [24]

    Focal loss for dense object de- tection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll ´ar. Focal loss for dense object de- tection. In Proceedings of the IEEE international con- ference on computer vision , pages 2980–2988, 2017. 1, 3

    2017

  17. [25]

    Unbiased teacher for semi- supervised object detection

    Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and P ´eter Vajda. Unbiased teacher for semi- supervised object detection. ArXiv, abs/2102.09480,

    arXiv

  18. [26]

    Pascal Mettes, Cees G. M. Snoek, and Shih-Fu Chang. Localizing actions from video labels and pseudo- annotations. ArXiv, abs/1707.09143, 2017. 1, 2

    2017 arXiv

  19. [27]

    Pascal Mettes, William Thong, and Cees G. M. Snoek. Object priors for classifying and localizing unseen ac- tions. ArXiv, abs/2104.04715, 2021. 1, 7

    2021 arXiv

  20. [28]

    Multi-region two-stream r-cnn for action detection

    Xiaojiang Peng and Cordelia Schmid. Multi-region two-stream r-cnn for action detection. InECCV, 2016. 2, 6

    2016

  21. [29]

    Girshick, Georgia Gkioxari, and Kaiming He

    Ilija Radosavovic, Piotr Doll ´ar, Ross B. Girshick, Georgia Gkioxari, and Kaiming He. Data distillation: Towards omni-supervised learning. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages 4119–4128, 2018. 2

    2018

  22. [30]

    Yolo9000: better, faster, stronger

    Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 7263–7271, 2017. 1, 3, 5

    2017

  23. [31]

    Girshick, and Jian Sun

    Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object de- tection with region proposal networks. IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 39:1137–1149, 2015. 1

    2015

  24. [32]

    Suman Saha, Gurkirt Singh, Michael Sapienza, Philip H. S. Torr, and Fabio Cuzzolin. Deep learning for detecting multiple space-time action tubes in videos. ArXiv, abs/1608.01529, 2016. 2

    2016 arXiv

  25. [33]

    Two-Stream Convolutional Networks for Action Recognition in Videos

    Karen Simonyan and Andrew Zisserman. Two-Stream Convolutional Networks for Action Recognition in Videos. In Neural Information Processing Systems ,

  26. [34]

    Gurkirt Singh, Suman Saha, Michael Sapienza, Philip H. S. Torr, and Fabio Cuzzolin. Online real-time mul- tiple spatiotemporal action localisation and prediction. 2017 IEEE International Conference on Computer Vi- sion (ICCV), pages 3657–3666, 2017. 2

    2017

  27. [35]

    Fixmatch: Simplifying Semi-Supervised Learning with Consis- tency and Confidence

    Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: Simplifying Semi-Supervised Learning with Consis- tency and Confidence. In Neural Information Process- ing Systems, 2020. 1, 2, 3, 6

    2020

  28. [36]

    A simple semi-supervised learning framework for object detec- tion

    Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detec- tion. ArXiv, abs/2005.04757, 2020. 2

    2005 arXiv

  29. [37]

    Tacnet: Transition-aware context network for spatio- temporal action detection

    Lin Song, Shiwei Zhang, Gang Yu, and Hongbin Sun. Tacnet: Transition-aware context network for spatio- temporal action detection. 2019 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 11979–11987, 2019. 2

    2019

  30. [38]

    Ucf101: A dataset of 101 human actions classes from videos in the wild

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5

    2012 arXiv

  31. [39]

    Murphy, Rahul Sukthankar, and Cordelia Schmid

    Chen Sun, Abhinav Shrivastava, Carl V ondrick, Kevin P. Murphy, Rahul Sukthankar, and Cordelia Schmid. Actor-centric relation network. ArXiv, abs/1807.10982, 2018. 2

    2018 arXiv

  32. [40]

    Proposal learning for semi-supervised object detection

    Peng Tang, Chetan Ramaiah, Ran Xu, and Caiming Xiong. Proposal learning for semi-supervised object detection. 2021 IEEE Winter Conference on Applica- tions of Computer Vision (WACV), pages 2290–2300,

    2021

  33. [41]

    Humble teachers teach better students for semi-supervised object detection

    Yihe Tang, Weifeng Chen, Yijun Luo, and Yuting Zhang. Humble teachers teach better students for semi-supervised object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR), pages 3131–3140, 2021. 1, 2, 6

    2021

  34. [42]

    Mean Teachers are Better Role Models: Weight-Averaged Consis- tency Targets Improve Semi-Supervised Deep Learn- ing Results

    Antti Tarvainen and Harri Valpola. Mean Teachers are Better Role Models: Weight-Averaged Consis- tency Targets Improve Semi-Supervised Deep Learn- ing Results. In Neural Information Processing Sys- tems, pages 1195–1204, 2017. 1, 2, 3, 6

    2017

  35. [43]

    Stage: Spatio- temporal attention on graph entities for video action detection

    Matteo Tomei, Lorenzo Baraldi, Simone Calderara, Simone Bronzin, and Rita Cucchiara. Stage: Spatio- temporal attention on graph entities for video action detection. ArXiv, abs/1912.04316, 2019. 2

    1912 arXiv

  36. [44]

    Video action detection by learning graph-based spatio-temporal in- teractions

    Matteo Tomei, Lorenzo Baraldi, Simone Calderara, Simone Bronzin, and Rita Cucchiara. Video action detection by learning graph-based spatio-temporal in- teractions. Comput. Vis. Image Underst., 206:103187,

  37. [45]

    Actionness estimation using hybrid fully con- volutional networks

    Limin Wang, Yu Qiao, Xiaoou Tang, and Luc Van Gool. Actionness estimation using hybrid fully con- volutional networks. 2016 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) , pages 2708–2717, 2016. 2

    2016

  38. [46]

    Tem- poral Segment Networks for Action Recognition in Videos

    Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaoou Tang, and Luc Van Gool. Tem- poral Segment Networks for Action Recognition in Videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:2740–2755, 2019. 1

    2019

  39. [47]

    Data-uncertainty guided multi-phase learning 10 for semi-supervised object detection

    Zhenyu Wang, Yali Li, Ye Guo, Lu Fang, and Shengjin Wang. Data-uncertainty guided multi-phase learning 10 for semi-supervised object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 4566–4575, 2021. 1, 2

    2021

  40. [48]

    Learning to track for spatio-temporal action localization

    Philippe Weinzaepfel, Za ¨ıd Harchaoui, and Cordelia Schmid. Learning to track for spatio-temporal action localization. 2015 IEEE International Conference on Computer Vision (ICCV), pages 3164–3172, 2015. 2

    2015

  41. [49]

    End-to-end semi-supervised object detection with soft teacher

    Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. ArXiv, abs/2106.09018, 2021. 1

    2021 arXiv

  42. [50]

    Interactive self-training with mean teachers for semi-supervised object detection

    Qize Yang, Xihan Wei, Biao Wang, Xia Hua, and Lei Zhang. Interactive self-training with mean teachers for semi-supervised object detection. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 5937–5946, 2021. 1

    2021

  43. [51]

    Davis, and Jan Kautz

    Xitong Yang, Xiaodong Yang, Ming-Yu Liu, Fanyi Xiao, Larry S. Davis, and Jan Kautz. Step: Spatio- temporal progressive learning for video action detec- tion. 2019 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 264–272,

    2019

  44. [52]

    S4L: Self-Supervised Semi- Supervised Learning

    Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4L: Self-Supervised Semi- Supervised Learning. In IEEE International Confer- ence on Computer Vision, pages 1476–1485, 2019. 2

    2019

  45. [53]

    A structured model for action detec- tion

    Yubo Zhang, Pavel Tokmakov, Cordelia Schmid, and Martial Hebert. A structured model for action detec- tion. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 9967–9976,

    2019

  46. [54]

    Jiaojiao Zhao and Cees G. M. Snoek. Dance with flow: Two-in-one stream action detection. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 9927–9936, 2019. 2

    2019

  47. [55]

    Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Shuai Bing, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, Ivan Marsic, Cees G. M. Snoek, and Joseph Tighe. Tuber: Tubelet transformer for video action detection. In IEEE Con- ference on Computer Vision and ...

  48. [56]

    Temporal action de- tection with structured segment networks

    Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaoou Tang, and Dahua Lin. Temporal action de- tection with structured segment networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2933–2942, 2017. 1

    2017

  49. [57]

    Oliveira, Nima Sedaghat, and Thomas Brox

    Mohammadreza Zolfaghari, Gabriel L. Oliveira, Nima Sedaghat, and Thomas Brox. Chained multi-stream networks exploiting pose, motion, and appearance for action classification and detection. 2017 IEEE In- ternational Conference on Computer Vision (ICCV) , pages 2923–2932, 2017. 1 11

    2017

Pith tools

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