Pith. sign in

REVIEW 4 major objections 4 minor 60 references

Hierarchical Event Memory for Accurate and Low-latency Online Video Temporal Grounding

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Online video grounding works better when the model remembers hierarchical events instead of raw frames.

desk verdict Solid method for online VTG with large gains, but missing hyperparameters (δ, positive-IoU threshold) and absent supplementary make the long-term memory claim under-specified. read the letter →

arxiv 2508.04546 v1 pith:AZPYYBFI submitted 2025-08-06 cs.CV

Minghang Zheng , Yuxin Peng , Benyuan Sun , Yi Yang , Yang Liu This is my paper
classification cs.CV
keywords onlinevideotemporalgroundinghierarchicaleventmemoryproposalsfuturepredictionsegmenttreelong-termunderstandingvideo-languagestart-timelatency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that online video temporal grounding should be done by classifying and refining event proposals, not by scoring each frame as a start or end boundary. To keep long-term history in a stream that never ends, it stores proposals in a fixed-size hierarchical event memory: small scales keep recent fine-grained events, large scales keep coarse-grained long-ago events, and a dynamic update rule merges or evicts redundant entries. It adds a future-prediction branch that detects whether the queried event will start soon and regresses the start time, which removes most of the latency of waiting for a complete event. The reported result is higher accuracy on TACoS, ActivityNet Captions, and MAD, while the start-time delay becomes negative on TACoS and ActivityNet Captions, meaning the model fires before the event's ground-truth start. The tradeoff is explicit: using the future branch sacrifices some recall for much lower latency, and the choice can be tuned with the future window.

What carries the argument

The central object is the hierarchical event memory: a fixed-size bank of event proposals at $L=8$ temporal scales, generated by a segment tree in which each scale merges two adjacent proposals from the scale below. It is what lets the model represent both recent fine-grained events and long-term coarse-grained events within a hard memory limit of $K=64$. The dynamic size rule allocates more slots to scales with higher positive-sample frequency, and the adaptive update rule compresses redundancy by average-pooling adjacent proposals whose cosine similarity exceeds $\delta$, falling back to first-in-first-out eviction. The future prediction branch is the second load-bearing mechanism: it cast

What would settle it

Construct a synthetic stream in which a query can only be grounded by one brief early event, then fill the stream with many high-similarity neighbors so the adaptive rule merges or evicts that event before the query time. If recall collapses while a model with unlimited history stays accurate, the claim that a 64-slot hierarchical event memory preserves long-term grounding evidence is refuted.

Watch

Extended reading notes

Core claim

The central claim is that a bounded, multi-scale memory of event proposals can replace frame-level memory for online video temporal grounding and improve both accuracy and start-time latency. Event proposals are built with a segment tree: each larger scale merges two adjacent smaller proposals, so durations grow by powers of two and extend beyond the short-term window by pulling stored events from the matching memory scale. Per-scale memory is allocated by positive-proposal frequency; when a scale overflows, similar adjacent proposals are merged, otherwise the oldest are evicted. Current proposals are refined by attention over stored history before classification and boundary regression. A f

Load-bearing premise

The method assumes that a fixed-size memory of 64 event proposals, updated by merging adjacent proposals whose cosine similarity exceeds an unspecified threshold and otherwise dropping the oldest, keeps enough long-term historical detail for any query; if a merged or evicted event was the only cue, the model cannot recover it.

Editorial extensions

If this is right

  • Without the future branch, event-proposal grounding improves R1@0.5 by 14.45 points over the existing online grounding method on TACoS (44.19 vs 29.74), so event-level memory alone is the main accuracy gain.
  • With the future branch, start delay becomes -1.28s on TACoS and -1.58s on ActivityNet Captions, meaning the model emits a start prediction before the ground-truth start; end delay is negative on all three datasets.
  • A fixed memory of 64 event proposals is enough across videos of very different lengths (MAD averages 110.8 minutes), which makes the approach practical for indefinite streams; inference remains real-time at about 614 FPS on MAD with pre-extracted features.
  • The future window (a,b) gives a tunable accuracy-latency knob: shifting it lets an application trade recall for earlier alerts, as shown in the reported accuracy-latency curve.

Reading between the lines

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

  • The same proposal-bank memory could be grafted onto online action localization, where the comparative baselines still use frame-level history; if the memory is what drives the gains, those methods should improve without changing their classifiers.
  • Because the update rule merges adjacent events whose similarity is high, two distinct occurrences of the same kind of event (the saxophone example in the paper) may collapse into one stored proposal; a stress test with repeated near-identical events would show whether the memory preserves 'again' distinctions.
  • The future branch could be exploiting event-duration priors learned from the dataset rather than genuinely anticipating from current context; varying the statistics of event durations or shuffling temporal order would separate these explanations.
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 / 4 minor

Summary. The paper proposes an online video temporal grounding (OnVTG) framework built on event proposals generated by a segment tree over short-term windows, enriched by a hierarchical event memory with dynamic per-scale allocation and adaptive merging/eviction. A future prediction branch predicts whether the target event will start in a near future window and regresses its start offset, enabling low-latency start predictions. The method is evaluated on TACoS, ActivityNet Captions, and MAD against online action detection/localization baselines and one prior OnVTG method, reporting substantial improvements in R@n,IoU=m and lower start delay. Ablations quantify the contributions of event memory, dynamic size, adaptive updating, and future prediction.

Significance. If the reported results are robust, this is a useful step for online video grounding: it moves from frame-level memory to hierarchical event-level memory, directly addresses long-term retention, and provides a practical accuracy-latency tradeoff through the future-prediction branch. The method is evaluated on three datasets of different regimes, code is promised, and the ablations are informative. However, the validation of the central long-term-retention mechanism is incomplete: key memory-update hyperparameters are not reported, no analysis shows that large-scale memory actually retains long-term information, and all results are single-run numbers. The contribution is significant but needs stronger empirical verification.

major comments (4)
  1. [Section 3.4, Eq. (5) and adaptive memory update] The merge threshold delta and the IoU threshold defining a 'positive' proposal are never specified. The adaptive update rule merges adjacent events only when cosine similarity exceeds delta; without delta the eviction/merge behavior is undefined. Eq. (5) allocates memory based on w_i, the frequency of positive proposals, which depends on the unspecified positive IoU threshold. The reference to Supplementary Materials for SD/ED does not cover these values, and no supplementary text is present. Please report both thresholds and include sensitivity analyses; otherwise the claimed long-term retention mechanism is unreproducible.
  2. [Section 3.4 and Section 3.6] The claimed advantage over frame-level memory is long-term retention via large-scale memory. With K=64 and L=8, Eq. (5) yields K_i = 1 + 56 * w_i / sum(w_j); if w_i is concentrated in small scales, large-scale buffers can have size 1, effectively discarding history. The paper does not report w_i values, per-scale memory sizes, or how often merge/evict operations remove query-relevant proposals. Please add analysis or sensitivity experiments, e.g., varying K and reporting per-scale effective retention, or ablating against fixed large-scale allocations.
  3. [Tables 1-4] All quantitative claims rest on single numbers with no error bars or significance tests. Headline gains (e.g., +14.45 on R1 0.5 in Table 1) are large, but without variance across seeds or runs it is impossible to judge whether differences are meaningful or whether ablation rows differ within noise. Report mean ± std over at least three seeds, or perform significance tests on the main comparisons.
  4. [Section 4.4 and SD/ED definitions] The modifications to OAT, HAT, OadTR, LSTR, GateHUB, TeSTra, and MiniROAD for OnVTG are not described. The reader cannot tell whether the comparison is favorable or whether these baselines were tuned. Moreover, SD/ED are claimed to be defined in the Supplementary Materials, but no supplementary text is included. Define SD/ED in the main text and document baseline adaptation and training details.
minor comments (4)
  1. [Section 3.3, Eq. (1)] The sentence 'At the j-th scale, the duration of proposal P^{j+1} is 2^j' appears to have an indexing mismatch: the text says proposals at scale j have duration 2^j but the equation defines P^{j+1} from merging two P^j proposals. Clarify the indexing.
  2. [Table 4] The first row of the ablation table is not explicitly described in the caption. State clearly which components are absent in each row (e.g., first row uses frame-level memory and no future prediction).
  3. [Section 4.4, performance comparisons] The abstract and conclusions state 'state-of-the-art performance', but the comparison set for online video grounding is limited to one prior method and several adapted baselines. Consider softening the claim or explicitly noting that offline methods are not directly comparable because they observe the full video.
  4. [Section 4.5, Table 5] Inference speed is reported as FPS, but the 'real-time' claim would be clearer if the typical video frame rate used in the datasets were stated (e.g., C3D/CLIP feature extraction rate).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper trains on labeled data and evaluates on held-out benchmarks; the equations are standard supervised objectives and no prediction reduces to a fitted input by construction.

full rationale

This is an empirical CV paper, not a formal derivation. The central claims—hierarchical event memory improves grounding accuracy and the future-prediction branch reduces start latency—are supported by training on labeled data and evaluating on held-out TACoS/ActivityNet Captions/MAD test sets. I checked the construction chain: Eq. (1) defines multi-scale proposals by merging adjacent features; Eqs. (2)-(3) are focal/DIoU losses supervised by ground-truth IoU labels; Eq. (4) supervises the future branch with annotated future-start labels (cf_t=1 iff t+a<=s<=t+b); Eq. (5) allocates per-scale memory from wi, the training-time frequency of positive proposals, not from test outputs. None of these quantities is fitted to the reported test metric and then called a prediction. The memory update rule (Section 3.4) is a heuristic with unspecified δ and a positive-IoU threshold (Section 3.3: 'an event proposal is regarded as positive if the IoU between the proposal and ground truth is larger than a certain threshold'), and SD/ED details are deferred to 'the Supplementary Materials' (Section 4.2). These are reproducibility/robustness gaps, not circular reductions: the unreported values are hyperparameters of a trained model, not terms that force the test numbers. Self-citations appear in related work and for standard components (MLP, transformer), but no load-bearing premise invokes the authors' prior theorem or uniqueness result. The ablation tables transparently compare variants (w/o FP, FIFO, equal memory) and report a latency-accuracy trade-off, which is the opposite of a forced or definitionally determined outcome. Verdict: no significant circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim depends on a set of hand-set hyperparameters (K, L, window, a, b), a data-derived allocation vector w_i, and an unspecified merge threshold delta. The method also assumes that fixed visual features and recursive merging preserve event semantics. No new physical entities are introduced.

free parameters (7)
  • memory size K = 64
    Total number of event proposals in the hierarchical memory, set by hand for all datasets.
  • scale count L = 8
    Number of proposal scales in the segment tree, set by hand.
  • short-term window length L_s = 8
    Number of feature steps in the current window, set by hand.
  • future prediction window (a,b) = -4, 4
    Hyperparameters defining when an event is considered 'about to start'; negative a allows predictions after event onset.
  • adaptive merge threshold delta = not specified
    Cosine similarity threshold for merging adjacent memory events; value not given in the paper.
  • per-scale positive frequency w_i = data-derived
    Frequency of positive proposals at each scale, used to allocate memory sizes; source (training vs. test) not clearly stated.
  • positive proposal IoU threshold = not specified
    Threshold for labeling a proposal as positive during training; value not given.
assumptions (5)
  • domain assumption Fixed frame-level features (C3D or CLIP) preserve the information needed for temporal grounding.
    Section 4.3 uses pre-extracted features from C3D on ActivityNet/TACoS and CLIP on MAD; all methods share these features.
  • domain assumption The segment tree merge in Eq. 1 yields proposal features that faithfully represent events of duration 2^j.
    Section 3.3 constructs proposals by recursively merging adjacent ones; no evidence that this preserves all event semantics.
  • domain assumption An event proposal is positive iff its IoU with ground truth exceeds a threshold; this assignment is a valid training signal.
    Section 3.3 defines positive proposals via IoU but does not state the threshold.
  • ad hoc to paper The future prediction window [t-4, t+4] correctly captures the notion of 'about to start' for the target event.
    Section 3.3 sets a=-4, b=4; this defines the training label for the future branch and biases the latency-accuracy trade-off.
  • ad hoc to paper Average pooling of adjacent memory events with cosine similarity above delta does not destroy information needed for later queries.
    Section 3.4's adaptive memory update merges redundant events; no analysis of information loss.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Event Memory for Accurate and Low-latency Online Video Temporal Grounding." pith.science (2026). https://pith.science/paper/AZPYYBFI

@misc{pith2026250804546,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Event Memory for Accurate and Low-latency Online Video Temporal Grounding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AZPYYBFI}},
  note         = {Machine review of arXiv:2508.04546}
}
read the original abstract

In this paper, we tackle the task of online video temporal grounding (OnVTG), which requires the model to locate events related to a given text query within a video stream. Unlike regular video temporal grounding, OnVTG requires the model to make predictions without observing future frames. As online videos are streaming inputs and can go on indefinitely, it is impractical and inefficient to store all historical inputs. The existing OnVTG models employ memory to store recent historical video frame features and predict scores indicating whether the current frame corresponds to the start or end time of the target event. However, these methods lack effective event modeling and cannot retain long-term historical information, leading to low performance. To tackle these challenges, we propose a hierarchical event memory for OnVTG. We propose an event-based OnVTG framework that makes predictions based on event proposals that model event-level information with various durations. To preserve historically valuable event information, we introduce a hierarchical event memory that retains historical events, allowing the model to access both recent and long-term information. To enable the real-time prediction, we further propose a future prediction branch that predicts whether the target event will occur shortly and further regresses the start time of the event. We achieve state-of-the-art performance on the TACoS, ActivityNet Captions, and MAD datasets. Code is available at https://github.com/minghangz/OnVTG.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 44 canonical work pages

  1. [1]

    A new surveil- lance and security alert system based on real-time motion de- tection.Journal of Smart Systems Research, 4:31–47, 2023

    Ahmed Al-Slemani and Ahmet ZENG ˙IN. A new surveil- lance and security alert system based on real-time motion de- tection.Journal of Smart Systems Research, 4:31–47, 2023. 1

    work page 2023

  2. [2]

    Miniroad: Minimal rnn framework for online action detection

    Joungbin An, Hyolim Kang, Su Ho Han, Ming-Hsuan Yang, and Seon Joo Kim. Miniroad: Minimal rnn framework for online action detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 10341– 10350, 2023. 2, 7

    work page 2023

  3. [3]

    E2e-load: end-to-end long-form online action de- tection

    Shuqiang Cao, Weixin Luo, Bairui Wang, Wei Zhang, and Lin Ma. E2e-load: end-to-end long-form online action de- tection. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 10422–10432, 2023

    work page 2023

  4. [4]

    Gatehub: Gated history unit with background sup- pression for online action detection

    Junwen Chen, Gaurav Mittal, Ye Yu, Yu Kong, and Mei Chen. Gatehub: Gated history unit with background sup- pression for online action detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19925–19934, 2022. 2, 7

    work page 2022

  5. [5]

    Enhancing long video understanding via hierarchical event-based memory

    Dingxin Cheng, Mingda Li, Jingyu Liu, Yongxin Guo, Bin Jiang, Qingbin Liu, Xi Chen, and Bo Zhao. Enhancing long video understanding via hierarchical event-based memory. arXiv preprint arXiv:2409.06299, 2024. 3

    arXiv 2024

  6. [6]

    Empirical evaluation of gated recurrent neural networks on sequence modeling.arXiv preprint arXiv:1412.3555, 2014

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling.arXiv preprint arXiv:1412.3555, 2014. 3

    arXiv 2014

  7. [7]

    Moment detection in long tutorial videos

    Ioana Croitoru, Simion-Vlad Bogolin, Samuel Albanie, Yang Liu, Zhaowen Wang, Seunghyun Yoon, Franck Der- noncourt, Hailin Jin, and Trung Bui. Moment detection in long tutorial videos. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pages 2594–2604, 2023. 1

    work page 2023

  8. [8]

    Online ac- tion detection

    Roeland De Geest, Efstratios Gavves, Amir Ghodrati, Zhenyang Li, Cees Snoek, and Tinne Tuytelaars. Online ac- tion detection. InComputer Vision–ECCV 2016: 14th Euro- pean Conference, Amsterdam, The Netherlands, October 11- 14, 2016, Proceedings, Part V 14, pages 269–284. Springer,

    work page 2016

Show all 60 references
  1. [9]

    Learning to discriminate information for online action detection

    Hyunjun Eun, Jinyoung Moon, Jongyoul Park, Chanho Jung, and Changick Kim. Learning to discriminate information for online action detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 809–818, 2020. 2

    2020

  2. [10]

    Temporal sentence grounding in streaming videos

    Tian Gan, Xiao Wang, Yan Sun, Jianlong Wu, Qingpei Guo, and Liqiang Nie. Temporal sentence grounding in streaming videos. InProceedings of the 31st ACM International Con- ference on Multimedia, pages 4637–4646, 2023. 1, 2, 3, 6, 7, 8

    2023

  3. [11]

    Tall: Temporal activity localization via language query

    Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on com- puter vision, pages 5267–5275, 2017. 1, 2

    2017

  4. [12]

    Prentice Hall PTR, 1994

    Simon Haykin.Neural networks: a comprehensive founda- tion. Prentice Hall PTR, 1994. 5

    1994

  5. [13]

    Ma-lmm: Memory-augmented large multimodal model for long-term video understanding

    Bo He, Hengduo Li, Young Kyun Jang, Menglin Jia, Xue- fei Cao, Ashish Shah, Abhinav Shrivastava, and Ser-Nam Lim. Ma-lmm: Memory-augmented large multimodal model for long-term video understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2024

  6. [14]

    Video activity localisation with uncertainties in temporal boundary

    Jiabo Huang, Hailin Jin, Shaogang Gong, and Yang Liu. Video activity localisation with uncertainties in temporal boundary. InProceedings of the European Conference on Computer Vision (ECCV), 2022. 2

    2022

  7. [15]

    Cag-qil: Context-aware actionness grouping via q im- itation learning for online temporal action localization

    Hyolim Kang, Kyungmin Kim, Yumin Ko, and Seon Joo Kim. Cag-qil: Context-aware actionness grouping via q im- itation learning for online temporal action localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13729–13738, 2021. 3

    2021

  8. [16]

    A sliding window scheme for online temporal action localization

    Young Hwi Kim, Hyolim Kang, and Seon Joo Kim. A sliding window scheme for online temporal action localization. In European Conference on Computer Vision, pages 653–669. Springer, 2022. 2, 3, 6, 7

    2022

  9. [17]

    Dense-captioning events in videos

    Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In International Conference on Computer Vision (ICCV), 2017. 2, 6

    2017

  10. [18]

    Efficient adaptive human-object inter- action detection with concept-guided memory, 2023

    Ting Lei, Fabian Caba, Qingchao Chen, Hailin Ji, Yuxin Peng, and Yang Liu. Efficient adaptive human-object inter- action detection with concept-guided memory, 2023. 2

    2023

  11. [19]

    Cross modal adaptive few-shot learning based on task depen- dence.Chinese Journal of Electronics, 32(1):85–96, 2023

    DAI Leichao, FENG Lin, SHANG Xinglin, and SU Han. Cross modal adaptive few-shot learning based on task depen- dence.Chinese Journal of Electronics, 32(1):85–96, 2023. 1

    2023

  12. [20]

    G2l: Semantically aligned and uniform video grounding via geodesic and game theory

    Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhi- hong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12032–12042, 2023. 7

    2023

  13. [21]

    Mo- mentdiff: Generative video moment retrieval from random to real.Advances in neural information processing systems, 36, 2024

    Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Mo- mentdiff: Generative video moment retrieval from random to real.Advances in neural information processing systems, 36, 2024. 1, 2

    2024

  14. [22]

    Focal loss for dense object detection.arXiv preprint arXiv:1708.02002, 2017

    T Lin. Focal loss for dense object detection.arXiv preprint arXiv:1708.02002, 2017. 5

    2017 arXiv

  15. [23]

    Towards balanced alignment: Modal-enhanced semantic modeling for video moment re- trieval

    Zhihang Liu, Jun Li, Hongtao Xie, Pandeng Li, Jiannan Ge, Sun-Ao Liu, and Guoqing Jin. Towards balanced alignment: Modal-enhanced semantic modeling for video moment re- trieval. InProceedings of the AAAI Conference on Artificial Intelligence, pages 3855–3863, 2024. 1, 2

    2024

  16. [24]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 6

    2017 arXiv

  17. [25]

    Towards generalisable video moment retrieval: Visual-dynamic injection to image-text pre-training

    Dezhao Luo, Jiabo Huang, Shaogang Gong, Hailin Jin, and Yang Liu. Towards generalisable video moment retrieval: Visual-dynamic injection to image-text pre-training. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2

    2023

  18. [26]

    Zero-shot video moment retrieval from frozen vision-language models

    Dezhao Luo, Jiabo Huang, Shaogang Gong, Hailin Jin, and Yang Liu. Zero-shot video moment retrieval from frozen vision-language models. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5464–5473, 2024. 2

    2024

  19. [27]

    Snag: Scalable and accurate video grounding

    Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18930–18940, 2024. 7

    2024

  20. [28]

    Local- global video-text interactions for temporal grounding

    Jonghwan Mun, Minsu Cho, and Bohyung Han. Local- global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 10810–10819, 2020. 1, 2

    2020

  21. [29]

    Scanning only once: An end-to-end framework for fast temporal grounding in long videos.arXiv preprint arXiv:2303.08345, 2023

    Yulin Pan, Xiangteng He, Biao Gong, Yiliang Lv, Yujun Shen, Yuxin Peng, and Deli Zhao. Scanning only once: An end-to-end framework for fast temporal grounding in long videos.arXiv preprint arXiv:2303.08345, 2023. 7

    2023 arXiv

  22. [30]

    An overview of cross-media retrieval: Concepts, methodologies, bench- marks, and challenges.IEEE Transactions on Circuits and Systems for Video Technology, 28:2372–2385, 2017

    Yuxin Peng, Xin Huang, and Yunzhen Zhao. An overview of cross-media retrieval: Concepts, methodologies, bench- marks, and challenges.IEEE Transactions on Circuits and Systems for Video Technology, 28:2372–2385, 2017. 1

    2017

  23. [31]

    Streaming long video understanding with large language models.Advances in Neu- ral Information Processing Systems, 37:119336–119360,

    Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuan- grui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models.Advances in Neu- ral Information Processing Systems, 37:119336–119360,

  24. [32]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, p...

    2021

  25. [33]

    Hat: History-augmented anchor transformer for on- line temporal action localization

    Sakib Reza, Yuexi Zhang, Mohsen Moghaddam, and Octavia Camps. Hat: History-augmented anchor transformer for on- line temporal action localization. InEuropean Conference on Computer Vision, pages 205–222. Springer, 2025. 2, 3, 6, 7

    2025

  26. [34]

    Coherent multi-sentence video description with variable level of detail

    Anna Rohrbach, Marcus Rohrbach, Wei Qiu, Annemarie Friedrich, Manfred Pinkal, and Bernt Schiele. Coherent multi-sentence video description with variable level of detail. InPattern Recognition: 36th German Conference, GCPR 2014, M¨unster, Germany, September 2-5, 2014, Proceedin...

    2014

  27. [35]

    Online action detection in untrimmed, streaming videos-modeling and evaluation

    Zheng Shou, Junting Pan, Jonathan Chan, Kazuyuki Miyazawa, Hassan Mansour, Anthony Vetro, Xavier Giroi Nieto, and Shih-Fu Chang. Online action detection in untrimmed, streaming videos-modeling and evaluation. In ECCV, page 5, 2018. 3

    2018

  28. [36]

    Mad: A scalable dataset for language grounding in videos from movie audio descriptions

    Mattia Soldan, Alejandro Pardo, Juan Le ´on Alc´azar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Re...

    2022

  29. [37]

    Moviechat: From dense token to sparse memory for long video understanding

    Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pat...

    2024

  30. [38]

    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. InProceedings of the IEEE inter- national conference on computer vision, pages 4489–4497,

  31. [39]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 3, 5

    2017

  32. [40]

    Oadtr: Online action detection with transformers

    Xiang Wang, Shiwei Zhang, Zhiwu Qing, Yuanjie Shao, Zhengrong Zuo, Changxin Gao, and Nong Sang. Oadtr: Online action detection with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 7565–7575, 2021. 3, 7

    2021

  33. [41]

    Efficient temporal extrapolation of multimodal large language models with temporal grounding bridge

    Yuxuan Wang, Yueqian Wang, Pengfei Wu, Jianxin Liang, Dongyan Zhao, Yang Liu, and Zilong Zheng. Efficient temporal extrapolation of multimodal large language models with temporal grounding bridge. InProceedings of the Em- pirical Methods in Natural Language Processing (EMNLP),

  34. [42]

    Videollamb: Long-context video understanding with recur- rent memory bridges.arXiv preprint arXiv:2409.01071,

    Yuxuan Wang, Cihang Xie, Yang Liu, and Zilong Zheng. Videollamb: Long-context video understanding with recur- rent memory bridges.arXiv preprint arXiv:2409.01071,

    arXiv

  35. [43]

    Negative sample matters: A renaissance of met- ric learning for temporal grounding

    Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gang- shan Wu. Negative sample matters: A renaissance of met- ric learning for temporal grounding. InProceedings of the AAAI Conference on Artificial Intelligence, pages 2613– 2623, 2022. 1, 2

    2022

  36. [44]

    Longvlm: Efficient long video understand- ing via large language models

    Yuetian Weng, Mingfei Han, Haoyu He, Xiaojun Chang, and Bohan Zhuang. Longvlm: Efficient long video understand- ing via large language models. InEuropean Conference on Computer Vision, pages 453–470. Springer, 2024. 3

    2024

  37. [45]

    Bridging the gap: A unified video comprehension framework for mo- ment retrieval and highlight detection

    Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Heng- wei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for mo- ment retrieval and highlight detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Reco...

    2024

  38. [46]

    Temporal recurrent networks for online action detection

    Mingze Xu, Mingfei Gao, Yi-Ting Chen, Larry S Davis, and David J Crandall. Temporal recurrent networks for online action detection. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 5532–5541,

  39. [47]

    Long short-term trans- former for online action detection.Advances in Neural In- formation Processing Systems, 34:1086–1099, 2021

    Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Xia, Zhuowen Tu, and Stefano Soatto. Long short-term trans- former for online action detection.Advances in Neural In- formation Processing Systems, 34:1086–1099, 2021. 3, 7

    2021

  40. [48]

    Active object detection with knowledge aggregation and distillation from large models

    Dejie Yang and Yang Liu. Active object detection with knowledge aggregation and distillation from large models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3

    2024

  41. [49]

    Flowgananomaly: Flow-based anomaly network intrusion detection with adver- sarial learning.Chinese Journal of Electronics, 33(1):58–71,

    LI Zeyi, W ANG Pan, and W ANG Zixuan. Flowgananomaly: Flow-based anomaly network intrusion detection with adver- sarial learning.Chinese Journal of Electronics, 33(1):58–71,

  42. [50]

    Flash-vstream: Memory- based real-time understanding for long video streams.arXiv preprint arXiv:2406.08085, 2024

    Haoji Zhang, Yiqin Wang, Yansong Tang, Yong Liu, Jiashi Feng, Jifeng Dai, and Xiaojie Jin. Flash-vstream: Memory- based real-time understanding for long video streams.arXiv preprint arXiv:2406.08085, 2024. 3

    2024 arXiv

  43. [51]

    Learning 2d temporal adjacent networks for moment local- ization with natural language

    Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment local- ization with natural language. InProceedings of the AAAI Conference on Artificial Intelligence, pages 12870–12877,

  44. [52]

    Progressive privileged knowledge distillation for on- line action detection.Pattern Recognition, 129:108741,

    Peisen Zhao, Lingxi Xie, Jiajie Wang, Ya Zhang, and Qi Tian. Progressive privileged knowledge distillation for on- line action detection.Pattern Recognition, 129:108741,

  45. [53]

    Real-time online video detection with temporal smoothing transformers

    Yue Zhao and Philipp Kr ¨ahenb¨uhl. Real-time online video detection with temporal smoothing transformers. InEuro- pean Conference on Computer Vision (ECCV), 2022. 7

    2022

  46. [54]

    Unsupervised cross-media hashing learning via knowledge graph.Chinese Journal of Electronics, 31(6):1081–1091, 2022

    YE Zhaoda, HE Xiangteng, and PENG Yuxin. Unsupervised cross-media hashing learning via knowledge graph.Chinese Journal of Electronics, 31(6):1081–1091, 2022. 1

    2022

  47. [55]

    Weakly supervised video moment localization with con- trastive negative sample mining

    Minghang Zheng, Yanjie Huang, Qingchao Chen, and Yang Liu. Weakly supervised video moment localization with con- trastive negative sample mining. InProceedings of the AAAI Conference on Artificial Intelligence, 2022. 2

    2022

  48. [56]

    Weakly supervised temporal sentence grounding with gaussian-based contrastive proposal learn- ing

    Minghang Zheng, Yanjie Huang, Qingchao Chen, Yuxin Peng, and Yang Liu. Weakly supervised temporal sentence grounding with gaussian-based contrastive proposal learn- ing. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2022. 2

    2022

  49. [57]

    Generating structured pseudo labels for noise- resistant zero-shot video sentence localization

    Minghang Zheng, Shaogang Gong, Hailin Jin, Yuxin Peng, and Yang Liu. Generating structured pseudo labels for noise- resistant zero-shot video sentence localization. InAnnual Meeting of the Association for Computational Linguistics,

  50. [58]

    Phrase-level temporal relationship mining for temporal sentence localization

    Minghang Zheng, Sizhe Li, Qingchao Chen, Yuxin Peng, and Yang Liu. Phrase-level temporal relationship mining for temporal sentence localization. InProceedings of the AAAI Conference on Artificial Intelligence, 2023. 2

    2023

  51. [59]

    Training-free video temporal grounding usinglarge-scale pre-trained models

    Minghang Zheng, Xinhao Cai, Qingchao Chen, Yuxin Peng, and Yang Liu. Training-free video temporal grounding usinglarge-scale pre-trained models. InProceedings of the European Conference on Computer Vision (ECCV), 2024. 2

    2024

  52. [60]

    Faster and better learning for bounding box regres- sion., 2020, 34.DOI: https://doi

    Z Zheng, P Wang, W Liu, J Li, R Ye, and D Ren Distance- IoU Loss. Faster and better learning for bounding box regres- sion., 2020, 34.DOI: https://doi. org/10.1609/aaai. v34i07, 6999:12993–13000, 2020. 5

    2020 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.