Prompt-aware of Frame Sampling for Efficient Text-Video Retrieval

Deyu Zhang; Jinrui Zhang; Ju Ren; Ligeng Chen; Tingting Long; Yaoxue Zhang

REVIEW 6 major objections 5 minor 35 references

Prompt-aware of Frame Sampling for Efficient Text-Video Retrieval

T0 review · 6 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 ProCLIP shows that query-aware frame sampling can replace dense uniform sampling, keeping R@1 at 49.0 on MSR-VTT while cutting retrieval latency by 75.3%.

desk verdict ProCLIP is a genuinely query-aware frame sampler that gets most of the way to state-of-the-art accuracy at a fraction of the latency; the overclaim and missing candidate-recall analysis are fixable, but the core idea is worth refereeing. read the letter →

arxiv 2507.15491 v1 pith:KG7B5OC5 submitted 2025-07-21 cs.MM cs.CV

Deyu Zhang , Tingting Long , Jinrui Zhang , Ligeng Chen , Ju Ren , Yaoxue Zhang This is my paper · ORCID

classification cs.MMcs.CV

keywords text-videoretrievalprompt-awareframesamplingcandidatepruningknowledgedistillationCLIPedgedeploymentefficiencymulti-granularattention

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 proposes ProCLIP, a framework for text-video retrieval on edge devices that aims to keep the accuracy of dense uniform frame sampling while drastically reducing computational cost. Its central claim is that both frame selection and video pruning can be conditioned on the user's text prompt rather than on static saliency. The system selects the frames most relevant to the query, uses a lightweight distilled feature extractor to retain only the top 50% of videos for CLIP re-ranking, and reports R@1 of 49.0 on MSR-VTT, nearly identical to the uniform-sampling baseline's 49.2, while cutting first-query latency by about three quarters. The significance would be that user-side, real-time text-video retrieval does not have to sacrifice accuracy for speed.

What carries the argument

The central mechanism is the prompt-aware frame sampling module: word-aware cross-attention and sentence-aware cross-attention, fused by a gated function, produce per-frame relevance scores conditioned on the query, selecting frames via a differentiable Hard Top-K. The second mechanism is the two-stage candidate pruning pipeline, where a lightweight extractor plus a three-layer Transformer distillation module scores all videos, keeps the top k% (k=50), and only then does CLIP's visual encoder process the surviving videos. The key identity is the distillation loss $L_{MSE}$ minimizing the L2 distance between lightweight video features and CLIP video features, which carries the argument that pruning preserves semantically relevant candidates.

What would settle it

Evaluate ProCLIP on a video domain shifted from its training distribution, such as theatrical or sports footage, and measure whether the recall of ground-truth videos within the top-50% candidate set stays high. If candidate-set recall drops measurably below the level needed to preserve end-to-end R@1 and the accuracy gap against uniform-sampling baselines widens, the distillation assumption fails. A direct ablation that replaces the distilled scorer with a random 50% subset should also show a large R@1 drop if the pruning is genuinely carrying the claimed load.

Watch

Extended reading notes

Core claim

In the paper's own terms, ProCLIP establishes that prompt-aware frame sampling plus two-stage candidate pruning yields near-top retrieval accuracy with a 75.3% reduction in latency. The discovery is that the frame-selection bottleneck in efficient retrieval is not how 'informative' a frame is in the abstract, but how relevant it is to the specific query. The paper shows that a lightweight MobileNetV3-based extractor, aligned to CLIP's feature space through MSE distillation, can safely discard half the video corpus before CLIP computes high-cost features, without dropping recall. It further shows that fusing word-level and sentence-level query attention with a learnable gate selects better frames than either level alone.

Load-bearing premise

The load-bearing premise is that the lightweight distilled extractor's similarities with CLIP text embeddings retain almost all truly relevant videos in the top-50% candidate set; if the distillation misaligns on a new video distribution, the pruning would discard good matches before CLIP ever sees them, and accuracy would drop.

Editorial extensions

If this is right

Retrieval latency scales roughly linearly with corpus size, so the 75.3% reduction holds as the number of videos grows.
Reducing the retained candidate ratio from 100% to 50% costs no R@1 on MSR-VTT (49.0), while aggressive pruning below 40% sharply worsens Mean Rank.
Query-specific frame selection improves personalized retrieval: on a 100-video ActivityNet subset with volunteer queries, ProCLIP reaches R@1 of 100.0 while static salient-frame baselines drop to roughly 61-63.
The learned gate between word-level and sentence-level attention outperforms simple addition or concatenation by more than 3 R@1 points.
The full system stays within mid-range GPU limits (5.38 GB memory, 475 J per average query), making it plausible for edge deployment.

Reading between the lines

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

The paper leaves implicit that the same coarse-to-fine pattern may transfer to other cross-modal search settings: a cheap distilled encoder could prune large galleries before an expensive multimodal model re-ranks the survivors.
The k=50% pruning ratio is chosen empirically; a natural extension the paper mentions but does not implement is predicting the pruning ratio per query from the confidence margin between top-ranked candidates.
Because distillation is trained only on three benchmarks, the promised recall preservation of the top-50% candidate set is a distribution-dependent property; out-of-distribution video collections could silently drop correct matches before CLIP sees them.
The user study is small-scale and self-reported, so the real-world satisfaction advantage is suggestive rather than decisive; a larger deployment would be needed to confirm the accuracy-versus-responsiveness balance holds broadly.

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.

Desk Editor's Note

read the letter

The thing to know about arXiv:2507.15491 is that the core idea is real and the paper is mostly honest about it, despite an overclaim in the abstract. ProCLIP's prompt-aware frame sampling — word- and sentence-level cross-attention with a learned gate — is a genuine departure from the static saliency samplers in AdaCLIP and mmSampler. The two-stage pruning with a CLIP-distilled lightweight ranker is a sensible integration of known ingredients, and the ablations actually support it: switching from simple-add fusion to gated fusion buys ~3.1 R@1 on MSR-VTT, and the distillation module is worth ~2.4 R@1 at the 50% pruning ratio. Those are real, attributable gains.

What the paper does well: it reports latency, energy, and memory along with accuracy; it includes a k-sweep on MSR-VTT; and it retrains competing baselines from official code. The reported 75.3% latency reduction over UCoFiA is plausible.

The soft spots are the following. First, the abstract's 'state-of-the-art accuracy' is not supported by Table 1: UCoFiA gets 49.2 R@1 on MSR-VTT versus 49.0, and ProCLIP trails on several metrics. The body mostly says 'competitive', so this is a wording problem, but the abstract is the abstract. Second, there are no error bars anywhere, and one table is internally inconsistent: Table 6 gives UCoFiA a V2T MnR of 191.2 with R@1 of 100.0, which cannot both be true. The personalized evaluation is also small (100 videos, 5 query writers). Third — and this is the one I would want fixed before publication — candidate-set recall at the 50% pruning stage is never reported. The method's final R@1 is upper-bounded by the recall of the lightweight ranker, and the ablation shows that ranker is load-bearing. The k-sweep on MSR-VTT shows R@1 holds at 50%, which is reassuring, but that is one dataset and only one pruning level is validated elsewhere. If the distilled features underperform on a different distribution, the speed-accuracy trade-off silently collapses. That risk is not quantified.

None of this is fatal. The central argument holds up on the evidence presented. This is a useful engineering contribution for edge text-video retrieval, not a new research direction.

My recommendation: deserve a serious referee; the paper should go to review, but the referees should ask for a corrected abstract, error bars or at least multiple seeds, a fix for Table 6, and a candidate-set recall analysis at the pruning stage.

Referee Report

6 major / 5 minor

Summary. The paper presents ProCLIP, a text-video retrieval framework aimed at edge deployment. It combines (i) a prompt-aware frame sampling module that fuses word-level and sentence-level cross-attention via a learned gate, and (ii) a two-stage candidate pruning scheme in which a lightweight MobileNetV3-based encoder with a distilled Transformer module scores all videos and only the top k=50% are processed by CLIP for final ranking. Experiments on MSR-VTT, ActivityNet, and MSVD report R@1=49.0 on MSR-VTT text-to-video retrieval, a 75.3% reduction in average-query latency relative to UCoFiA, and lower latency than most efficient baselines. The paper also includes ablations of the gated fusion and distillation modules, a k-sweep on MSR-VTT, and two small personalized evaluations.

Significance. If the reported numbers are reproducible, the paper would provide a useful practical result: query-conditioned frame selection and lightweight coarse pruning with CLIP distillation can cut retrieval latency by roughly an order of magnitude relative to uniform-sampling methods while incurring only a small R@1 penalty on standard benchmarks. The ablation design is informative, and the efficiency metrics (FQ/AQ latency, energy, memory) are more comprehensive than in most retrieval papers. The strengths are the concrete k-sweep, the distillation/no-distillation comparison, and the attempt to evaluate personalization. However, the significance is moderated by the fact that ProCLIP does not actually beat the best uniform-sampling baselines on several headline metrics (UCoFiA has R@1=49.2 on MSR-VTT), and by unresolved data inconsistencies that cloud the quantitative claims.

major comments (6)

[3.2, Eqs. (3) and (4)] The two reweighting equations are dimensionally inconsistent as written. In Eq. (3), S is defined in R^N, so F·S^T with F in R^{N×D} and S^T in R^{1×N} is not a valid matrix product and cannot yield Wo in R^{N×D}. Similarly, in Eq. (4), softmax(SF^T/√D)^T is in R^{N×1}, so a multiplication by F (N×D) is undefined. The intended operation is presumably elementwise scaling of each frame feature by its relevance score (e.g., Wo = S ⊙ F or Wo = diag(S)F). Please correct the equations and, if the intended operation differs, show the correct dimensions through the gated fusion in Eqs. (5)–(7).
[4.4.1, Table 6] The UCoFiA row reports video-to-text R@1=100.0 with MnR=191.2, which is internally inconsistent: if 100% of queries place the ground truth at rank 1, the mean rank cannot be 191.2. This suggests a data-entry error, and because the personalized table is used to support claims about the behavior of dense-sampling methods, the table must be corrected or the comparison removed. The personalized subset is also small (100 videos, 5 query writers) and no variance is reported, so the stated baseline degradations should be treated as anecdotal unless accompanied by confidence intervals.
[4.2.1, Tables 1 and 3, and Section 4.1.4] There are unresolved numerical and configuration inconsistencies. Table 1 reports ProCLIP on MSR-VTT text-to-video as R@5=75.3, MnR=11.2, but Table 3 at the stated k=50% (the configuration used for Table 1 latency) reports R@5=75.4, MnR=13.7, and Table 4 reports MnR=11.1 for the gated module. Also, Section 4.1.4 says 16/16/64 frames are uniformly sampled for MSR-VTT/MSVD/ActivityNet, whereas Table 1 lists ProCLIP as 12/32/12 frames. Please state precisely which configuration produced each table and reconcile all numbers.
[3.3, Eq. (8), and Tables 3 and 5] The central speed–accuracy claim depends on the lightweight ranker's ability to retain the ground-truth video in the top-50% candidate set, but this candidate recall is never reported. Table 5 shows that removing distillation degrades MSR-VTT R@1 from 49.0 to 46.6, confirming the stage is load-bearing, yet there is no analysis of how MSE loss (Eq. (9)) translates into ranking fidelity or how recall varies across datasets. The k-sweep is provided only for MSR-VTT, so the choice k=50% is not validated on ActivityNet or MSVD. Please report the candidate-set recall (and the resulting R@1 upper bound) of the lightweight stage at k=50% and at other k values on all three benchmarks, with a breakdown by video length or content distribution if possible.
[Abstract and Section 4.2.1, Table 1] The statement that ProCLIP achieves 'state-of-the-art accuracy' is not supported by the paper's own numbers. On MSR-VTT text-to-video, UCoFiA attains R@1=49.2 versus 49.0 for ProCLIP; on MSVD, TEACHCLIP attains R@1=47.3 versus 47.1; and on ActivityNet the T2V R@1 is tied at 45.7. The defensible claim is 'competitive accuracy with substantially lower latency,' and the text should be revised accordingly. The sentence 'ProCLIP outperforms previous SoTA methods on most evaluation metrics' in Section 4.2.1 is also too broad given these R@1 deficits.
[4.2] No standard errors, confidence intervals, or multiple-seed results are reported for any benchmark number. Given that the headline difference from UCoFiA is 0.2 R@1 (49.0 vs 49.2) and several latency-accuracy trade-off points in Table 3 differ by 0.1–0.3 R@1, the 'competitive accuracy' assertion needs at least a few seeds and a simple significance test. Without this, it is unclear whether the observed differences are within training noise.

minor comments (5)

[Abstract] The abstract says 'Code is available at here,' which appears to have a placeholder where the URL should be; include the actual link.
[Title] The title 'Prompt-aware of Frame Sampling for Efficient Text-Video Retrieval' is grammatically awkward; consider 'Prompt-Aware Frame Sampling for Efficient Text-Video Retrieval.'
[4.1.1] The sentence 'conduct two more personalized evaluations' is missing a subject; change to 'we also conduct two more personalized evaluations.'
[Table 1] The 'Frames' column uses slash-separated numbers without explaining the order; add a footnote clarifying per-dataset frame counts, especially since they conflict with the 16/16/64 values in Section 4.1.4.
[4.4.2, Figure 6] The user satisfaction scores are reported as bars without error bars or statistical tests; report the distributions and pairwise significance so the claimed 'significantly outperformed' is supported.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the accuracy and latency claims are empirical measurements, not algebraic consequences of the paper's own definitions.

full rationale

ProCLIP's headline numbers are empirical: R@1=49.0 and the latency reductions are measured under a fixed computation ratio k=50% (Sec. 3.3, Eq. 8; Sec. 4.2.2). The k hyperparameter is selected from the MSR-VTT sweep in Table 3, but that sweep shows R@1 flat at 49.0 from k=100% down to k=50%, so the headline accuracy is not forced by the pruning threshold; the candidate-set quality of the distilled lightweight ranker is an empirically measured property, not an algebraic identity. The distillation loss (Eq. 9) aligns a lightweight module with CLIP features, and the final CLIP re-ranking is a separate stage; final R@1 is not a renamed version of the MSE loss, nor is it equal to the pruning ratio by construction. No load-bearing self-citation chain appears: the only overlap with the authors' prior work is a background citation ([34]) for the growth of user-generated content, and the technical building blocks (Hard Top-k from [12], CLIP from [20]) are external. The absence of candidate-set recall analysis is a robustness concern, not a circularity, because the paper directly measures final retrieval accuracy across k values rather than defining it in terms of the pruning stage. Therefore no circular step can be exhibited from the paper's own equations or citations.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities; it combines existing neural modules. The central claims rest on the tunable pruning ratio k, dataset-specific frame counts, and the empirical validity of distillation-based feature alignment.

free parameters (4)

computation ratio k = 50%
The proportion of videos retained after lightweight pruning; chosen on MSR-VTT as the best latency-accuracy trade-off (Table 3). The headline 75.3% latency reduction and equal accuracy at k=50% depend on this value.
uniform frame count per dataset = 16 (MSR-VTT, MSVD), 64 (ActivityNet)
Number of frames uniformly sampled before prompt-aware selection; set per dataset in implementation details.
HardTop-k temperature schedule = initial 5, decay 0.045
Hyperparameters for the differentiable hard top-k selector, adopted from AdaCLIP.
learning rates = 1e-7 for CLIP backbone, 1e-4 for other modules
Learning rates used in training; standard hyperparameters.

assumptions (3)

domain assumption Lightweight features, after MSE distillation to CLIP features, remain sufficiently aligned with CLIP text features for top-50% pruning to preserve recall across datasets.
This is the key assumption behind the two-stage pruning claim; it is validated empirically only on the three tested datasets, not proven.
domain assumption Query-relevant frames can be identified from lightweight visual features via cross-attention with CLIP text embeddings, and end-to-end training with the retrieval loss drives this alignment.
The prompt-aware sampler's effectiveness rests on this learnability; no formal guarantee is provided.
standard math Standard attention, softmax, and contrastive learning machinery behave as expected.
The methods rely on standard deep learning operations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Prompt-aware of Frame Sampling for Efficient Text-Video Retrieval." pith.science (2026). https://pith.science/paper/KG7B5OC5

@misc{pith2026250715491,
  author       = {Pith},
  title        = {Pith review of: Prompt-aware of Frame Sampling for Efficient Text-Video Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KG7B5OC5}},
  note         = {Machine review of arXiv:2507.15491}
}

read the original abstract

Enabling efficient text-video retrieval on edge-end devices is critical for real-world applications. Yet, existing methods face a critical challenge in balancing accuracy and computational efficiency: uniform frame sampling methods ensure content coverage but incur prohibitive computational costs, while salient-frame sampling methods reduce overhead but suffer from query-agnostic frame selection that biases retrieval results. To address this, we propose ProCLIP, a user-centric framework that achieves state-of-the-art accuracy with significantly improved efficiency. We design a prompt-aware frame sampling strategy that dynamically guides lightweight feature extractors using textual prompts to select semantically relevant frames, overcoming the limitations of existing salient-frame sampling methods which rely on static, query-agnostic selection criteria. Moreover, we adopt a two-stage candidate pruning strategy that combines rapid coarse filtering via a lightweight module with CLIP-powered fine-grained re-ranking, enhancing retrieval efficiency while preserving accuracy. Experiments across benchmarks show ProCLIP achieves 75.3% latency reduction versus baselines while maintaining competitive accuracy, i.e., R@1=49.0 in MSR-VTT dataset. Code is available at https://github.com/tiffylong/ProCLIP.

Figures

Figures reproduced from arXiv: 2507.15491 by the authors.

**Figure 2.** Retrieval Latency-Accuracy Trade-off on MSR-VTT test set. X/Y-axes: retrieval latency (seconds) vs. R@1. Each method is [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Training and inference pipelines of ProCLIP. injection of sinusoidal positional embeddings, the encoded frame representations are propagated through a multilayer temporal dependency modeling module. This module employs a depth-variable Transformer architecture: a 3-layer self-attention network is instantiated for short-duration videos (T ≤ 60 seconds), while a deeper 5- layer configuration is activated for long-for… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Word-aware cross attention and sentence-aware cross attention of prompt-aware frame sampling. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Retrieval latency under various number of video on the MSR-VTT test set (1,000 videos). [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

**Figure 6.** Figure 6: User satisfaction scores. To quantify user perceptions, each participant rated their retrieval experience for every query using a 10-point scale (10 indicating optimal experience, 0 indicating complete failure), jointly evaluating retrieval accuracy and system responsi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 31 canonical work pages

[1]

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

work page 2015
[2]

David Chen and William B Dolan. 2011. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies . 190–200

work page 2011
[3]

Marcos V Conde, Saman Zadtootaghaj, Nabajeet Barman, Radu Timofte, Chenlong He, Qi Zheng, Ruoxi Zhu, Zhengzhong Tu, Haiqiang Wang, Xiangguang Chen, et al. 2024. AIS 2024 challenge on video quality assessment of user-generated content: Methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 5826–5837

work page 2024
[4]

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition . Ieee, 248–255

work page 2009
[5]

Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. 2021. Clip2video: Mastering video-text retrieval via image clip.arXiv preprint arXiv:2106.11097 (2021)

arXiv 2021
[6]

Deepak Ghimire, Dayoung Kil, and Seong-heum Kim. 2022. A Survey on Efficient Convolutional Neural Networks and Hardware Acceleration. Electronics 11, 6 (Jan. 2022), 945. doi:10.3390/electronics11060945 Number: 6 Publisher: Multidisciplinary Digital Publishing Institute

work page doi:10.3390/electronics11060945 2022
[7]

Satya Krishna Gorti, Noël Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. 2022. X-pool: Cross-modal language-video attention for text-video retrieval. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5006–5015

work page 2022
[8]

Ning Han, Jingjing Chen, Hao Zhang, Huanwen Wang, and Hao Chen. 2022. Adversarial multi-grained embedding network for cross-modal text-video retrieval. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) 18, 2 (2022), 1–23

work page 2022

Show all 35 references

[9]

Ning Han, Yawen Zeng, Chuhao Shi, Guangyi Xiao, Hao Chen, and Jingjing Chen. 2023. Bic-net: Learning efficient spatio-temporal relation for text-video retrieval. ACM Transactions on Multimedia Computing, Communications and Applications (TOMM) 20, 3 (2023), 1–21

2023
[10]

Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. 2019. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision . 1314–1324

2019
[11]

Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. 2017. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)

2017 arXiv
[12]

Zhiming Hu, Angela Ning Ye, Salar Hosseini Khorasgani, and Iqbal Mohomed. 2023. AdaCLIP: Towards pragmatic multimodal video retrieval. In Proceedings of the 31st ACM International Conference on Multimedia . 5623–5633

2023
[13]

Zhiming Hu, Ning Ye, and Iqbal Mohomed. 2022. mmSampler: Efficient Frame Sampler for Multimodal Video Retrieval. Proceedings of Machine Learning and Systems 4 (2022), 153–171

2022
[14]

Paul Pu Liang, Amir Zadeh, and Louis-Philippe Morency. 2024. Foundations & trends in multimodal machine learning: Principles, challenges, and open questions. Comput. Surveys 56, 10 (2024), 1–42

2024
[15]

Chengzhi Lin, Ancong Wu, Junwei Liang, Jun Zhang, Wenhang Ge, Wei-Shi Zheng, and Chunhua Shen. 2022. Text-adaptive multiple visual prototype matching for video-text retrieval. Advances in neural information processing systems 35 (2022), 38655–38666

2022
[16]

Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. 2022. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508 (2022), 293–304

2022
[17]

Yiwei Ma, Guohai Xu, Xiaoshuai Sun, Ming Yan, Ji Zhang, and Rongrong Ji. 2022. X-clip: End-to-end multi-grained contrastive learning for video-text retrieval. In Proceedings of the 30th ACM international conference on multimedia . 638–647

2022
[18]

Thao-Nhu Nguyen, Le Minh Quang, Graham Healy, Binh T Nguyen, and Cathal Gurrin. 2024. VideoCLIP 2.0: an interactive clip-based video retrieval system for novice users at vbs2024. In International Conference on Multimedia Modeling . Springer, 394–399

2024
[19]

Guanqiao Qu, Qiyuan Chen, Wei Wei, Zheng Lin, Xianhao Chen, and Kaibin Huang. 2025. Mobile edge intelligence for large language models: A contemporary survey. IEEE Communications Surveys & Tutorials (2025)

2025
[20]

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

2021
[21]

Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. 2018. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition . 4510–4520

2018
[22]

Leqi Shen, Sicheng Zhao, Yifeng Zhang, Pengzhang Liu, Yongjun Bao, and Guiguang Ding. 2025. Spatio-Temporal Attention for Text-Video Retrieval. ACM Transactions on Multimedia Computing, Communications and Applications (TOMM) (2025)

2025
[23]

Kaibin Tian, Yanhua Cheng, Yi Liu, Xinglin Hou, Quan Chen, and Han Li. 2024. Towards efficient and effective text-to-video retrieval with coarse-to-fine visual representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 5207–5214. Manus...

2024
[24]

Kaibin Tian, Ruixiang Zhao, Zijie Xin, Bangxiang Lan, and Xirong Li. 2024. Holistic features are almost sufficient for text-to-video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 17138–17147

2024
[25]

Tianshi Wang, Fengling Li, Lei Zhu, Jingjing Li, Zheng Zhang, and Heng Tao Shen. 2025. Cross-modal retrieval: a systematic review of methods and future directions. Proc. IEEE (2025)

2025
[26]

Yingchao Wang, Chen Yang, Shulin Lan, Liehuang Zhu, and Yan Zhang. 2024. End-edge-cloud collaborative computing for deep learning: A comprehensive survey. IEEE Communications Surveys & Tutorials (2024)

2024
[27]

Ziyang Wang, Yi-Lin Sung, Feng Cheng, Gedas Bertasius, and Mohit Bansal. 2023. Unified coarse-to-fine alignment for video-text retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 2816–2827

2023
[28]

Nina Willis, Abraham Bernstein, and Luca Rossetto. 2025. Effects of Human Cognition-Inspired Task Presentation on Interactive Video Retrieval. ACM Transactions on Multimedia Computing, Communications and Applications (TOMM) 21, 6 (2025), 1–25

2025
[29]

Jun Xu, Tao Mei, Ting Yao, and Yong Rui. 2016. Msr-vtt: A large video description dataset for bridging video and language. In Proceedings of the IEEE conference on computer vision and pattern recognition . 5288–5296

2016
[30]

Chenyang Yu, Xuehu Liu, Yingquan Wang, Pingping Zhang, and Huchuan Lu. 2024. TF-CLIP: Learning text-free CLIP for video-based person re-identification. In Proceedings of the AAAI conference on artificial intelligence , Vol. 38. 6764–6772

2024
[31]

Xuzheng Yu, Chen Jiang, Xingning Dong, Tian Gan, Ming Yang, and Qingpei Guo. 2025. SHE-Net: Syntax-Hierarchy-Enhanced Text-Video Retrieval. IEEE Transactions on Circuits and Systems for Video Technology (2025)

2025
[32]

Yuan Yuan, Zhaojian Li, and Bin Zhao. 2025. A Survey of Multimodal Learning: Methods, Applications, and Future. Comput. Surveys (2025)

2025
[33]

Shunxiang Zhang, Jiajia Liu, Yixuan Jiao, Yulei Zhang, Lei Chen, and Kuanching Li. 2025. A Multimodal Semantic Fusion Network with Cross- Modal Alignment for Multimodal Sentiment Analysis (TOMM).ACM Trans. Multimedia Comput. Commun. Appl.(June 2025). doi:10.1145/3744648 Just Accepted

2025 doi
[34]

Ye Zhang, Jinrui Zhang, Sheng Yue, Wei Lu, Ju Ren, and Xuemin Shen. 2024. Mobile generative ai: Opportunities and challenges. IEEE Wireless Communications 31, 4 (2024), 58–64

2024
[35]

Shuai Zhao, Linchao Zhu, Xiaohan Wang, and Yi Yang. 2022. Centerclip: Token clustering for efficient text-video retrieval. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . 970–981. Manuscript submitted to ACM

2022

Pith tools

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

[1] [1]

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

work page 2015

[2] [2]

David Chen and William B Dolan. 2011. Collecting highly parallel data for paraphrase evaluation. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies . 190–200

work page 2011

[3] [3]

Marcos V Conde, Saman Zadtootaghaj, Nabajeet Barman, Radu Timofte, Chenlong He, Qi Zheng, Ruoxi Zhu, Zhengzhong Tu, Haiqiang Wang, Xiangguang Chen, et al. 2024. AIS 2024 challenge on video quality assessment of user-generated content: Methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 5826–5837

work page 2024

[4] [4]

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition . Ieee, 248–255

work page 2009

[5] [5]

Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. 2021. Clip2video: Mastering video-text retrieval via image clip.arXiv preprint arXiv:2106.11097 (2021)

arXiv 2021

[6] [6]

Deepak Ghimire, Dayoung Kil, and Seong-heum Kim. 2022. A Survey on Efficient Convolutional Neural Networks and Hardware Acceleration. Electronics 11, 6 (Jan. 2022), 945. doi:10.3390/electronics11060945 Number: 6 Publisher: Multidisciplinary Digital Publishing Institute

work page doi:10.3390/electronics11060945 2022

[7] [7]

Satya Krishna Gorti, Noël Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guangwei Yu. 2022. X-pool: Cross-modal language-video attention for text-video retrieval. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5006–5015

work page 2022

[8] [8]

Ning Han, Jingjing Chen, Hao Zhang, Huanwen Wang, and Hao Chen. 2022. Adversarial multi-grained embedding network for cross-modal text-video retrieval. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) 18, 2 (2022), 1–23

work page 2022