Pith. sign in

REVIEW 4 major objections 6 minor 70 references

Hierarchical Multi-Stage Transformer Architecture for Context-Aware Temporal Action Localization

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a three-tier transformer architecture called PCL-Former—one module for proposals, one for classification, one for boundary localization—outperforms prior temporal action localization methods on THUMOS14…

desk verdict A sensible but incremental three-transformer TAL pipeline whose real performance is hidden behind non-comparable tIoU averaging and a few internal number inconsistencies. read the letter →

arxiv 2507.06411 v2 pith:6FK35NNR submitted 2025-07-08 cs.CV

Hayat Ullah , Arslan Munir , Oliver Nina This is my paper · ORCID
classification cs.CV
keywords temporalactionlocalizationtransformerarchitecturevideounderstandinguntrimmedproposalboundaryrefinementoverlaplossPCL-Former
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 dividing temporal action localization among three specialized transformer modules, rather than solving it in one pass, yields the best published accuracy on three standard benchmarks. The proposed architecture, PCL-Former, gives one transformer the job of proposing candidate action segments, a second the job of classifying them, and a third the job of refining their start and end times, with each module trained by a loss tailored to its job. If the reported numbers hold, the paper is evidence that modular task decomposition plus attention-based context modeling is a strong recipe for long, untrimmed video. This matters because temporal action localization is the practical backbone for video search, surveillance, and content understanding, and a robust gain on three major benchmarks would be a step toward usable video understanding.

What carries the argument

The load-bearing mechanism is the three-stage pipeline itself, with each stage a transformer encoder that uses factorized self-attention: spatial and temporal attention are computed separately, reducing the quadratic cost of full self-attention to linear cost in sequence length. The Localization-Former's specialized overlap loss, $$\mathcal{L}_{\text{Overlap}} = \frac{1}{N}\sum_n \left( \frac{1}{2}\left( \frac{P(a_n)$_n^{2}$}{$v_n^{{\alpha}}$} - 1\right)\right),$$ makes the confidence score reflect how much a segment actually overlaps a ground-truth action, so that post-processing keeps the right detections. The same underlying encoder is reused across modules, and the Boundary Refinement Module filters low-overlap predictions before non-maximum suppression is applied.

What would settle it

Recompute the THUMOS14 average mAP using only the tIoU thresholds {0.3, 0.4, 0.5, 0.6, 0.7}, with $\alpha$ and $\lambda$ fixed and variances reported across seeds; if the lead over the prior best method disappears or reverses, the headline improvement is an artifact of including the easy 0.1 and 0.2 thresholds or of hyperparameter choice.

Watch

Extended reading notes

Core claim

On the paper's own terms, PCL-Former is a claim that temporal action localization improves when the task is decomposed into three transformer sub-tasks, each trained with its own loss. The Proposal-Former labels 64-frame video segments as action or background; the Classification-Former assigns an action category to the surviving segments; and the Localization-Former, trained with an overlap loss that rewards segments with high temporal intersection-over-union against ground truth, refines start and end times. The paper reports average mAP (mean average precision) of 78.4% on THUMOS14, 38.6% on ActivityNet-1.3, and 44.5% on HACS, and states that these numbers surpass prior state-of-the-art methods by 2.8, 1.2, and 4.8 percentage points, respectively.

Load-bearing premise

The load-bearing premise is that the reported margins come from the architecture rather than from evaluation choices, because the THUMOS14 average includes tIoU thresholds 0.1 and 0.2 that most compared methods do not report and the values of $\alpha$ and $\lambda$ are not given.

Editorial extensions

If this is right

  • If the reported numbers hold, PCL-Former becomes the best listed method on THUMOS14 with 78.4% average mAP, ahead of the 69.5% of the previous strongest entry.
  • If the reported numbers hold, the ActivityNet-1.3 average mAP of 38.6% beats the 37.8% of the strongest listed comparison.
  • If the reported numbers hold, the HACS average mAP of 44.5% surpasses the 39.7% of the best listed comparison.
  • Ablations show each stage contributes: removing the Proposal-Former, the Localization-Former, or the Boundary Refinement Module lowers average mAP on all three datasets.
  • Inference uses the Proposal-Former to discard background segments first, so the other two modules see only candidate action segments; the paper presents this as a main reason for the improved results.

Reading between the lines

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

  • Beyond the paper, the proposal/classify/localize split could transfer to related video tasks such as temporal proposal generation or language grounding, which face the same tension between labeling a segment and refining its boundaries.
  • Beyond the paper, the largest reported gain is on HACS, yet the HACS comparison table is sparser than the other two; running modern strong baselines on HACS under identical tIoU thresholds is the most direct independent test.
  • Beyond the paper, the overlap loss depends on hyperparameters $\alpha$ and $\lambda$ that are not reported; a sensitivity sweep over these values would reveal whether the margins are robust or tuned.
  • Beyond the paper, the architecture passes each segment through three transformer modules, so measuring inference cost against single-stage baselines would clarify whether the accuracy gain is worth the added computation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes PCL-Former, a three-stage transformer architecture for temporal action localization in which a Proposal-Former filters background segments, a Classification-Former assigns action categories, and a Localization-Former refines temporal boundaries using an overlap loss, followed by a boundary refinement module. The method is evaluated on THUMOS14, ActivityNet-1.3, and HACS, with reported average mAP values of 78.4%, 38.6%, and 44.5%, respectively, and the abstract claims improvements over state-of-the-art approaches of 2.8%, 1.2%, and 4.8%. The paper also presents ablations on NMS thresholds, module contributions, and segment length.

Significance. If the reported numbers were obtained under a common evaluation protocol, PCL-Former would be a strong result: the architecture is described in enough detail to be understood, the ablations in Tables 3–5 provide useful insight into the design choices, and the method is evaluated on held-out test or validation splits rather than derived from training parameters. The overlap loss is properly credited to Shou et al. (2016), and the use of factorized self-attention for long video segments is a sensible efficiency choice. However, the significance is currently undermined by the evaluation-protocol mismatch and internal inconsistencies detailed below; the quantitative claims in the abstract are not reproducible from the tables as presented, and the lack of hyperparameter values and variance estimates prevents the reader from assessing whether the margins are robust.

major comments (4)
  1. [Table 1, Section 4.4] The THUMOS14 'Avg.' column is not computed over a common set of tIoU thresholds across rows. PCL-Former's 78.4% averages thresholds {0.1,...,0.7}, whereas the strongest modern competitors (ActionFormer, ViT-TAD, TriDet, CausalTAD, AdaTAD) report only {0.3,...,0.7}. Restricting PCL-Former to the common range gives (86.8+84.5+77.2+68.6+54.1)/5 ≈ 74.2%, which changes the margin versus ViT-TAD from 8.9 points to about 4.7 points. The abstract's '2.8%' improvement cannot be reproduced from Table 1 under any stated comparison, so the headline quantitative claim is not supported as written.
  2. [Abstract; Section 4.4; Tables 1 and 2] The headline margins in the abstract are internally inconsistent with the tables. Section 4.4 states that the best ActivityNet-1.3 average mAP is 38.2, but Table 1 reports 38.6 for PCL-Former (I3D). The abstract's 1.2% ActivityNet gain matches a comparison to ViT-TAD (37.4), not to the best prior in Table 1, AdaTAD (37.8), which gives a 0.8-point gap. On HACS, the abstract's 4.8% matches a comparison to TCA-NET (39.7) rather than to the best prior TriDet-SlowFast (38.6), which gives 5.9 points. The authors must state explicitly which comparison each percentage refers to and correct the text/table inconsistency.
  3. [Section 4.3; Eqs. (3) and (4)] The values of the overlap-loss exponent α and the loss-balance weight λ are never reported, and no seed count, per-run variance, or code release is indicated. This is load-bearing because the ActivityNet margin over the best prior in Table 1 is only 0.8 points, a difference that could easily be within run-to-run variation or sensitive to these hyperparameters. Without these values and variance estimates, the claimed state-of-the-art superiority cannot be properly assessed.
  4. [Appendix A.1, Figure 4] The classwise comparison reports PCL-Former mAP of 78.4 versus AMNet's 63.3, but the AMNet number is not accompanied by the tIoU threshold set or evaluation protocol used, and it is unlikely to be comparable to PCL-Former's threshold set {0.1,...,0.7}. This comparison should either be removed or documented with identical thresholds; otherwise it is not credible as a same-protocol comparison.
minor comments (6)
  1. [Abstract] The first sentence is a grammatically incomplete fragment ('Inspired by the recent success of transformers...'); it should be merged with the following sentence.
  2. [Section 4.1] The dataset is referred to as 'ActivityNet-1.4' in the text, but the correct name is ActivityNet-1.3.
  3. [Table 3 caption] The caption reads 'ActivitNet-1.3' and should be corrected to 'ActivityNet-1.3'.
  4. [Section 3.2, Eq. (1)] Equation (1) is typeset in a garbled way and should be rewritten; the notation S{s1, s2, ...sn} and the summation over N are unclear.
  5. [Section 3.2, Eq. (3)] The notation P(a_n)_n in the overlap loss is confusing; the probability P and the class index a_n should be defined unambiguously.
  6. [Table 1 and References] The SRF-Net row cites Ou et al. (2022), but the corresponding bibliography entry is a human pose estimation paper, not the temporal action localization method listed in the table; this reference needs to be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed TAL gains are empirical comparisons on held-out benchmarks, and the borrowed S-CNN loss and ViViT-style modules are external, not self-referential inputs.

full rationale

The paper makes no circularity claims of the kinds enumerated. Its central result is an empirical mAP comparison on held-out test/validation splits of THUMOS14, ActivityNet-1.3, and HACS against externally published methods. The architecture is assembled from cited external components (ViViT-like factorized self-attention, the S-CNN overlap loss, and the MMAction2 codebase), none of which is authored by the present authors or conditioned on PCL-Former's result. Equations (2)-(4) define training losses from ground-truth overlap labels v and prediction probabilities P; they do not encode test-set mAP as an input. The 'specialized loss' in Eq. (3) is explicitly attributed to Shou et al. (2016), so presenting it as part of the method is a matter of attribution and novelty, not circular reasoning. The claimed SOTA margins may be weakened by non-comparable tIoU averaging on THUMOS14 (0.1-0.7 vs 0.3-0.7), an internal ActivityNet text/table discrepancy (38.2 vs 38.6), and unreported alpha, lambda, and seed values, but these are evaluation-protocol and reproducibility problems, not cases where a prediction is equivalent to its input by construction. No self-citation chain is load-bearing, no uniqueness theorem is imported, and no fitted parameter is renamed as a prediction. Therefore no circular step meets the evidentiary bar required by the review rules.

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

The pipeline rests on commonplace deep learning components plus several unreported choices (alpha, lambda, per-dataset NMS thresholds, segment length). The alpha and lambda values are not given; the complexity claim is incorrect as stated. No new entities are introduced.

free parameters (5)
  • alpha (overlap loss exponent) = not reported
    Controls confidence score weighting in Eq. 3; no ablations or stated value; likely tuned per dataset.
  • lambda (loss balance weight) = not reported
    Balances softmax and overlap losses in Eq. 4; value not given; affects final loss and thus training.
  • NMS tIoU thresholds = 0.4 (THUMOS14), 0.5 (ActivityNet-1.3, HACS)
    Selected via grid search in Table 3; this is fitting a decision parameter to the test and validation data.
  • Segment length (SGM) = 64 frames
    Chosen per dataset through ablation (Table 5); different lengths yield different mAP, so 64 is a fitted choice.
  • Proposal-Former IoU labels = 0.7 positive / 0.3 negative
    Inherited from S-CNN (Shou et al., 2016); labeling rule is an assumption about what counts as a positive segment.
assumptions (5)
  • domain assumption Pre-extracted TSN/I3D features retain enough temporal boundary information for precise localization without fine-tuning.
    Section 3.2 and 4.3: the backbone is used as a fixed feature extractor; if the features discard boundary cues, the localization modules cannot recover them.
  • domain assumption Factorized self-attention (as in ViViT) reduces attention complexity from quadratic to linear, enabling long-video modeling.
    Section 3.2 states this; standard factorized attention keeps per-factor quadratic cost, so the complexity claim is inaccurate and overstates scalability.
  • domain assumption The S-CNN overlap loss yields confidence scores that improve NMS selection.
    Eq. 3, borrowed from Shou et al. (2016); no analysis or ablation for alpha, and no comparison against alternative localization losses.
  • standard math Standard self-attention, softmax classification, and NMS behave as described in prior literature.
    Background machinery used throughout Section 3; accepted conventions, not contributed here.
  • ad hoc to paper Linear combination of losses with a fixed lambda produces a well-posed training objective.
    Eq. 4; lambda is not reported or ablated, making the unified training protocol underspecified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Multi-Stage Transformer Architecture for Context-Aware Temporal Action Localization." pith.science (2026). https://pith.science/paper/6FK35NNR

@misc{pith2026250706411,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Multi-Stage Transformer Architecture for Context-Aware Temporal Action Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6FK35NNR}},
  note         = {Machine review of arXiv:2507.06411}
}
read the original abstract

Inspired by the recent success of transformers and multi-stage architectures in video recognition and object detection domains. We thoroughly explore the rich spatio-temporal properties of transformers within a multi-stage architecture paradigm for the temporal action localization (TAL) task. This exploration led to the development of a hierarchical multi-stage transformer architecture called PCL-Former, where each subtask is handled by a dedicated transformer module with a specialized loss function. Specifically, the Proposal-Former identifies candidate segments in an untrimmed video that may contain actions, the Classification-Former classifies the action categories within those segments, and the Localization-Former precisely predicts the temporal boundaries (i.e., start and end) of the action instances. To evaluate the performance of our method, we have conducted extensive experiments on three challenging benchmark datasets: THUMOS-14, ActivityNet-1.3, and HACS Segments. We also conducted detailed ablation experiments to assess the impact of each individual module of our PCL-Former. The obtained quantitative results validate the effectiveness of the proposed PCL-Former, outperforming state-of-the-art TAL approaches by 2.8%, 1.2%, and 4.8% on THUMOS14, ActivityNet-1.3, and HACS datasets, respectively.

Figures

Figures reproduced from arXiv: 2507.06411 by the authors.

Figure 1
Figure 1. Visual overview of our proposed PCL-Former. The input untrimmed video is divided into [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The detailed architectural overview of our [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The detailed inference workflow of our PCL￾Former architecture: the Proposal Former identifies candi￾date segment features (SFi), segment filtering discards back￾ground segments, and only positive (action-containing) seg￾ments are passed to the Classification Former and Localiza￾tion Former for action classification and temporal boundary prediction. For a given set of segments S of an untrimmed video V along with it… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Class-wise performance comparison with S-CNN Shou et al. (2016), CDC Shou et al. [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: The prediction results for the Javelin Throw and Hammer Throw action instances from the THUMOS14 test set, evaluated using an overlap threshold of 0.4. For each ground truth instance, two predictions are provided. The NMS process is applied to both, selecting the predi…
Figure 6
Figure 6. Figure 6: The prediction results for the Long Jump and Diving action instances from the THU￾MOS14 test set, evaluated using an overlap threshold of 0.4. For each ground truth instance, two predictions are provided. The NMS process is applied to both, selecting the prediction wit…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 58 canonical work pages

  1. [1]

    Vivit: A video vision transformer

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lu c i \'c , and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 6836--6846, 2021

    work page 2021

  2. [2]

    Boundary content graph neural network for temporal action proposal generation

    Yueran Bai, Yingying Wang, Yunhai Tong, Yang Yang, Qiyue Liu, and Junhui Liu. Boundary content graph neural network for temporal action proposal generation. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XXVIII 16, pp.\ 121--137. Springer, 2020

    work page 2020

  3. [3]

    Is space-time attention all you need for video understanding? In ICML, volume 2, pp.\ 4, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, volume 2, pp.\ 4, 2021

    work page 2021

  4. [4]

    Activitynet: A large-scale video benchmark for human activity understanding

    Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the ieee conference on computer vision and pattern recognition, pp.\ 961--970, 2015

    2015

  5. [5]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pp.\ 213--229. Springer, 2020

    2020

  6. [6]

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

    Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp.\ 6299--6308, 2017

    work page 2017

  7. [7]

    Augmented transformer with adaptive graph for temporal action proposal generation

    Shuning Chang, Pichao Wang, Fan Wang, Hao Li, and Zheng Shou. Augmented transformer with adaptive graph for temporal action proposal generation. In Proceedings of the 3rd International Workshop on Human-Centric Multimedia Analysis, pp.\ 41--50, 2022

    work page 2022

  8. [8]

    Rethinking the faster r-cnn architecture for temporal action localization

    Yu-Wei Chao, Sudheendra Vijayanarasimhan, Bryan Seybold, David A Ross, Jia Deng, and Rahul Sukthankar. Rethinking the faster r-cnn architecture for temporal action localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 1130--1139, 2018

    work page 2018

Show all 70 references
  1. [9]

    Dcan: improving temporal action detection via dual context aggregation

    Guo Chen, Yin-Dong Zheng, Limin Wang, and Tong Lu. Dcan: improving temporal action detection via dual context aggregation. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pp.\ 248--257, 2022

    2022

  2. [10]

    Tallformer: Temporal action localization with a long-memory transformer

    Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with a long-memory transformer. In European Conference on Computer Vision, pp.\ 503--521. Springer, 2022

    2022

  3. [11]

    Openmmlab's next generation video understanding toolbox and benchmark

    MMAction2 Contributors. Openmmlab's next generation video understanding toolbox and benchmark. https://github.com/open-mmlab/mmaction2, 2020

    2020

  4. [12]

    Ms-tct: multi-scale temporal convtransformer for action detection

    Rui Dai, Srijan Das, Kumara Kahatapitiya, Michael S Ryoo, and Fran c ois Br \'e mond. Ms-tct: multi-scale temporal convtransformer for action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 20041--20051, 2022

    2022

  5. [13]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv...

    2010 arXiv

  6. [14]

    Multiscale vision transformers

    Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 6824--6835, 2021

    2021

  7. [15]

    Holistic interaction transformer network for action detection

    Gueter Josmy Faure, Min-Hung Chen, and Shang-Hong Lai. Holistic interaction transformer network for action detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pp.\ 3340--3350, 2023

    2023

  8. [16]

    Turn tap: Temporal unit regression network for temporal action proposals

    Jiyang Gao, Zhenheng Yang, Kan Chen, Chen Sun, and Ram Nevatia. Turn tap: Temporal unit regression network for temporal action proposals. In Proceedings of the IEEE international conference on computer vision, pp.\ 3628--3636, 2017

    2017

  9. [17]

    Temporal action localization with multi-temporal scales

    Zan Gao, Xinglei Cui, Tao Zhuo, Zhiyong Cheng, An-An Liu, Meng Wang, and Shenyong Chen. Temporal action localization with multi-temporal scales. arXiv preprint arXiv:2208.07493, 2022

    2022 arXiv

  10. [18]

    in the wild

    Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos “in the wild”. Computer Vision and Image Understanding, 155: 0 1--23, 2017

    2017

  11. [19]

    Action-aware masking network with group-based attention for temporal action localization

    Tae-Kyung Kang, Gun-Hee Lee, Kyung-Min Jin, and Seong-Whan Lee. Action-aware masking network with group-based attention for temporal action localization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp.\ 6058--6067, 2023

    2023

  12. [20]

    Neighbor-guided pseudo-label generation and refinement for single-frame supervised temporal action localization

    Guozhang Li, De Cheng, Nannan Wang, Jie Li, and Xinbo Gao. Neighbor-guided pseudo-label generation and refinement for single-frame supervised temporal action localization. IEEE Transactions on Image Processing, 2024 a

    2024

  13. [21]

    Detal: open-vocabulary temporal action localization with decoupled networks

    Zhiheng Li, Yujie Zhong, Ran Song, Tianjiao Li, Lin Ma, and Wei Zhang. Detal: open-vocabulary temporal action localization with decoupled networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024 b

    2024

  14. [22]

    Learning salient boundary feature for anchor-free temporal action localization

    Chuming Lin, Chengming Xu, Donghao Luo, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yanwei Fu. Learning salient boundary feature for anchor-free temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni...

    2021

  15. [23]

    Single shot temporal action detection

    Tianwei Lin, Xu Zhao, and Zheng Shou. Single shot temporal action detection. In Proceedings of the 25th ACM international conference on Multimedia, pp.\ 988--996, 2017

    2017

  16. [24]

    Bsn: Boundary sensitive network for temporal action proposal generation

    Tianwei Lin, Xu Zhao, Haisheng Su, Chongjing Wang, and Ming Yang. Bsn: Boundary sensitive network for temporal action proposal generation. In Proceedings of the European conference on computer vision (ECCV), pp.\ 3--19, 2018

    2018

  17. [25]

    Bmn: Boundary-matching network for temporal action proposal generation

    Tianwei Lin, Xiao Liu, Xin Li, Errui Ding, and Shilei Wen. Bmn: Boundary-matching network for temporal action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 3889--3898, 2019

    2019

  18. [26]

    Harnessing temporal causality for advanced temporal action detection

    Shuming Liu, Lin Sui, Chen-Lin Zhang, Fangzhou Mu, Chen Zhao, and Bernard Ghanem. Harnessing temporal causality for advanced temporal action detection. arXiv preprint arXiv:2407.17792, 2024 a

    2024 arXiv

  19. [27]

    End-to-end temporal action detection with 1b parameters across 1000 frames

    Shuming Liu, Chen-Lin Zhang, Chen Zhao, and Bernard Ghanem. End-to-end temporal action detection with 1b parameters across 1000 frames. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 18591--18601, 2024 b

    2024

  20. [28]

    End-to-end temporal action detection with transformer

    Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. IEEE Transactions on Image Processing, 31: 0 5427--5441, 2022 a

    2022

  21. [29]

    Video swin transformer

    Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3202--3211, 2022 b

    2022

  22. [30]

    Gaussian temporal awareness networks for action localization

    Fuchen Long, Ting Yao, Zhaofan Qiu, Xinmei Tian, Jiebo Luo, and Tao Mei. Gaussian temporal awareness networks for action localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 344--353, 2019

    2019

  23. [31]

    Enhancing temporal action localization in an end-to-end network through estimation error incorporation

    Mozhgan Mokari and Khosrow Haj Sadeghi. Enhancing temporal action localization in an end-to-end network through estimation error incorporation. Image and Vision Computing, 145: 0 104994, 2024

    2024

  24. [32]

    Srfnet: selective receptive field network for human pose estimation

    Zhilong Ou, YanMin Luo, Jin Chen, and Geng Chen. Srfnet: selective receptive field network for human pose estimation. The Journal of Supercomputing, 78 0 (1): 0 691--711, 2022

    2022

  25. [33]

    Temporal context aggregation network for temporal action proposal refinement

    Zhiwu Qing, Haisheng Su, Weihao Gan, Dongliang Wang, Wei Wu, Xiang Wang, Yu Qiao, Junjie Yan, Changxin Gao, and Nong Sang. Temporal context aggregation network for temporal action proposal refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern rec...

    2021

  26. [34]

    Action sensitivity learning for temporal action localization

    Jiayi Shao, Xiaohan Wang, Ruijie Quan, Junjun Zheng, Jiang Yang, and Yi Yang. Action sensitivity learning for temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 13457--13469, 2023

    2023

  27. [35]

    Tridet: Temporal action detection with relative boundary modeling

    Dingfeng Shi, Yujie Zhong, Qiong Cao, Lin Ma, Jia Li, and Dacheng Tao. Tridet: Temporal action detection with relative boundary modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 18857--18866, 2023

    2023

  28. [36]

    Temporal action localization in untrimmed videos via multi-stage cnns

    Zheng Shou, Dongang Wang, and Shih-Fu Chang. Temporal action localization in untrimmed videos via multi-stage cnns. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 1049--1058, 2016

    2016

  29. [37]

    Cdc: Convolutional-de-convolutional networks for precise temporal action localization in untrimmed videos

    Zheng Shou, Jonathan Chan, Alireza Zareian, Kazuyuki Miyazawa, and Shih-Fu Chang. Cdc: Convolutional-de-convolutional networks for precise temporal action localization in untrimmed videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 5...

    2017

  30. [38]

    Bsn++: Complementary boundary regressor with scale-balanced relation modeling for temporal action proposal generation

    Haisheng Su, Weihao Gan, Wei Wu, Yu Qiao, and Junjie Yan. Bsn++: Complementary boundary regressor with scale-balanced relation modeling for temporal action proposal generation. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pp.\ 2602--2610, 2021

    2021

  31. [39]

    Pcg-tal: Progressive cross-granularity cooperation for temporal action localization

    Rui Su, Dong Xu, Lu Sheng, and Wanli Ouyang. Pcg-tal: Progressive cross-granularity cooperation for temporal action localization. IEEE Transactions on Image Processing, 30: 0 2103--2113, 2020

    2020

  32. [40]

    Relaxed transformer decoders for direct action proposal generation

    Jing Tan, Jiaqi Tang, Limin Wang, and Gangshan Wu. Relaxed transformer decoders for direct action proposal generation. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 13526--13535, 2021

    2021

  33. [41]

    Learnable feature augmentation framework for temporal action localization

    Yepeng Tang, Weining Wang, Chunjie Zhang, Jing Liu, and Yao Zhao. Learnable feature augmentation framework for temporal action localization. IEEE Transactions on Image Processing, 2024

    2024

  34. [42]

    Learning spatiotemporal features with 3d convolutional networks

    Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pp.\ 4489--4497, 2015

    2015

  35. [43]

    A closer look at spatiotemporal convolutions for action recognition

    Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp.\ 6450--6459, 2018

    2018

  36. [44]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  37. [45]

    Exploring sub-action granularity for weakly supervised temporal action localization

    Binglu Wang, Xun Zhang, and Yongqiang Zhao. Exploring sub-action granularity for weakly supervised temporal action localization. IEEE Transactions on Circuits and Systems for Video Technology, 32 0 (4): 0 2186--2198, 2021 a

    2021

  38. [46]

    Temporal segment networks: Towards good practices for deep action recognition

    Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaoou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pp.\ 20--36. Springer, 2016

    2016

  39. [47]

    Temporal action proposal generation with transformers

    Lining Wang, Haosen Yang, Wenhao Wu, Hongxun Yao, and Hujie Huang. Temporal action proposal generation with transformers. arXiv preprint arXiv:2105.12043, 2021 b

    2021 arXiv

  40. [48]

    Rcl: Recurrent continuous localization for temporal action detection

    Qiang Wang, Yanhao Zhang, Yun Zheng, and Pan Pan. Rcl: Recurrent continuous localization for temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 13566--13575, 2022

    2022

  41. [49]

    Ensemble prototype network for weakly supervised temporal action localization

    Kewei Wu, Wenjie Luo, Zhao Xie, Dan Guo, Zhao Zhang, and Richang Hong. Ensemble prototype network for weakly supervised temporal action localization. IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  42. [50]

    A pursuit of temporal accuracy in general activity detection

    Yuanjun Xiong, Yue Zhao, Limin Wang, Dahua Lin, and Xiaoou Tang. A pursuit of temporal accuracy in general activity detection. arXiv preprint arXiv:1703.02716, 2017

    2017 arXiv

  43. [51]

    R-c3d: Region convolutional 3d network for temporal activity detection

    Huijuan Xu, Abir Das, and Kate Saenko. R-c3d: Region convolutional 3d network for temporal activity detection. In Proceedings of the IEEE international conference on computer vision, pp.\ 5783--5792, 2017

    2017

  44. [52]

    G-tad: Sub-graph localization for temporal action detection

    Mengmeng Xu, Chen Zhao, David S Rojas, Ali Thabet, and Bernard Ghanem. G-tad: Sub-graph localization for temporal action detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10156--10165, 2020

    2020

  45. [53]

    Revisiting anchor mechanisms for temporal action localization

    Le Yang, Houwen Peng, Dingwen Zhang, Jianlong Fu, and Junwei Han. Revisiting anchor mechanisms for temporal action localization. IEEE Transactions on Image Processing, 29: 0 8535--8548, 2020

    2020

  46. [54]

    Adapting short-term transformers for action detection in untrimmed videos

    Min Yang, Huan Gao, Ping Guo, and Limin Wang. Adapting short-term transformers for action detection in untrimmed videos. arXiv preprint arXiv:2312.01897, 2023

    2023 arXiv

  47. [55]

    Weakly-supervised temporal action localization by inferring salient snippet-feature

    Wulian Yun, Mengshi Qi, Chuanming Wang, and Huadong Ma. Weakly-supervised temporal action localization by inferring salient snippet-feature. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pp.\ 6908--6916, 2024

    2024

  48. [56]

    Graph convolutional networks for temporal action localization

    Runhao Zeng, Wenbing Huang, Mingkui Tan, Yu Rong, Peilin Zhao, Junzhou Huang, and Chuang Gan. Graph convolutional networks for temporal action localization. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 7094--7103, 2019

    2019

  49. [57]

    Actionformer: Localizing moments of actions with transformers

    Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pp.\ 492--510. Springer, 2022

    2022

  50. [58]

    Towards efficient use of multi-scale features in transformer-based object detectors

    Gongjie Zhang, Zhipeng Luo, Zichen Tian, Jingyi Zhang, Xiaoqin Zhang, and Shijian Lu. Towards efficient use of multi-scale features in transformer-based object detectors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 6206--6216, 2023

    2023

  51. [59]

    Learning sparse 2d temporal adjacent networks for temporal action localization

    Songyang Zhang, Houwen Peng, Le Yang, Jianlong Fu, and Jiebo Luo. Learning sparse 2d temporal adjacent networks for temporal action localization. arXiv preprint arXiv:1912.03612, 2019

    1912 arXiv

  52. [60]

    Video self-stitching graph network for temporal action localization

    Chen Zhao, Ali K Thabet, and Bernard Ghanem. Video self-stitching graph network for temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 13658--13667, 2021

    2021

  53. [61]

    Re2tal: Rewiring pretrained video backbones for reversible temporal action localization

    Chen Zhao, Shuming Liu, Karttikeya Mangalam, and Bernard Ghanem. Re2tal: Rewiring pretrained video backbones for reversible temporal action localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 10637--10647, 2023

    2023

  54. [62]

    Hacs: Human action clips and segments dataset for recognition and temporal localization

    Hang Zhao, Antonio Torralba, Lorenzo Torresani, and Zhicheng Yan. Hacs: Human action clips and segments dataset for recognition and temporal localization. In Proceedings of the IEEE International Conference on Computer Vision, pp.\ 8668--8678, 2019

    2019

  55. [63]

    Tuber: Tubelet transformer for video action detection

    Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber: Tubelet transformer for video action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, ...

    2022

  56. [64]

    Bottom-up temporal action localization with mutual regularization

    Peisen Zhao, Lingxi Xie, Chen Ju, Ya Zhang, Yanfeng Wang, and Qi Tian. Bottom-up temporal action localization with mutual regularization. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part VIII 16, pp.\ 539--555. Springer, 2020

    2020

  57. [65]

    Temporal action detection with structured segment networks

    Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaoou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision, pp.\ 2914--2923, 2017

    2017

  58. [66]

    Enriching local and global contexts for temporal action localization

    Zixin Zhu, Wei Tang, Le Wang, Nanning Zheng, and Gang Hua. Enriching local and global contexts for temporal action localization. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 13516--13525, 2021

    2021

  59. [67]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  60. [68]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  61. [69]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  62. [70]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

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