Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

TCMA: Text-Conditioned Multi-granularity Alignment for Drone Cross-Modal Text-Video Retrieval

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read This paper tries to establish that drone text-video retrieval is best served by fine-grained captions and a model that aligns text with video at three levels—whole video, relevant frames, and individual patches—backed by a new benchmark dat

desk verdict Useful drone retrieval benchmark with an unvalidated LLM annotation pipeline; the method is a modest, incremental gain that is overstated in the abstract. read the letter →

arxiv 2510.10180 v2 pith:3TRGZ7OK submitted 2025-10-11 cs.CV

classification cs.CV
keywords dronetext-videoretrievalaerialvideounderstandingmulti-granularityalignmentvideo-textdatasetcross-modalvision-languagemodelsadaptiveattentiontemperatureUAV
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 existing aerial video datasets have coarse, redundant captions that blunt text-video retrieval, and that both a better dataset and a better alignment model are needed. To prove this, it introduces DVTMD, a drone video-text dataset with 2,864 videos and five detailed captions per video, then proposes TCMA, a model that aligns sentences to whole videos, uses sentence-guided attention to weight frames, and uses word-guided attention to select informative image patches. On DVTMD, TCMA reports 45.5% R@1 for text-to-video retrieval and 42.8% R@1 for video-to-text retrieval, outperforming prior text-video retrieval methods. If correct, DVTMD becomes the standard benchmark for drone text-video retrieval and TCMA demonstrates that multi-granularity alignment, token filtering, and adaptive attention sharpness are the right ingredients for aerial video search.

What carries the argument

The central mechanism is the Text-Conditioned Multi-granularity Aggregation module, which progressively refines a video representation under textual guidance. It produces three video embeddings: a text-agnostic mean-pooled video feature, a sentence-guided frame-level feature computed by softmax attention with a text-predicted temperature, and a word-guided patch-level feature computed by attending from selected words to selected patches. The Word and Patch Selection module uses small scoring networks to keep the top-K words and top-K patches per frame, reducing noise from cluttered aerial scenes. The Text-Adaptive Dynamic Temperature Mechanism is a lightweight network that outputs a positive

What would settle it

Take a random sample of DVTMD captions, show each caption alongside its source video to human annotators, and ask whether every detailed claim in the caption is visibly present in the frames; if a substantial fraction of captions contain hallucinated objects, colors, or actions, the benchmark's ground truth is unreliable and the reported R@1 gains may reflect matching caption style rather than visual understanding.

Watch

Extended reading notes

Core claim

The paper's central claim is that drone text-video retrieval fails when captions are generic and when models align at a single granularity. Its proposed TCMA framework integrates three complementary alignment levels: global video-sentence alignment captures overall scene semantics; sentence-guided frame aggregation weights frames by relevance to the caption; and word-guided patch alignment matches selected words to selected visual patches. A Word and Patch Selection module filters out irrelevant background patches and uninformative words, and a Text-Adaptive Dynamic Temperature Mechanism predicts a softmax temperature from the text so that action-oriented captions focus attention sharply on

Load-bearing premise

The DVTMD captions are generated by a vision-language model and summarized by another language model, with no human verification, so all retrieval accuracy numbers are only as trustworthy as those machine-generated captions' fidelity to what is actually visible in the videos.

Editorial extensions

If this is right

  • DVTMD provides a reusable benchmark with 14,320 fine-grained captions, enabling fair comparison of future drone text-video retrieval methods.
  • Ablations indicate that no single alignment level is sufficient: combining global, frame-level, and patch-level alignment yields the best retrieval accuracy.
  • Top-K word and patch selection improves over aligning all tokens, showing that filtering irrelevant content matters in wide-field aerial video.
  • Adapting attention temperature to the text type gives consistent recall improvements over a fixed temperature.
  • Qualitative results suggest the model can capture object counts, colors, and spatial layout, not just coarse scene categories.

Reading between the lines

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

  • If DVTMD is adopted widely, the fact that its captions are machine-generated without human verification could bias models toward matching caption-writing style rather than true visual content; a human audit of a caption sample would clarify this.
  • The dynamic temperature idea is not drone-specific and could transfer to other cluttered video domains such as surveillance or egocentric video, where some queries target brief actions and others target persistent scenes.
  • A natural testable extension is to compute retrieval accuracy per event category; if gains concentrate in action-heavy categories, that would corroborate the dynamic-temperature mechanism's intended behavior.
  • The one-to-many caption design (five captions per video) could be reused for training generative video models, not just retrieval, since it exposes multiple linguistically distinct descriptions of the same visual content.
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 addresses text-video retrieval for drone videos. It introduces DVTMD, a dataset of 2,864 ERA videos each paired with five captions generated by Qwen2.5-VL frame descriptions and Qwen3 summarization. It then proposes TCMA, a CLIP-based framework with mean-pooled video embedding, sentence-guided frame aggregation with a text-adaptive dynamic temperature, word/patch top-k selection, word-guided patch aggregation, and a hierarchical contrastive/Pearson loss. Experiments on DVTMD and CapERA report R@1/R@5/R@10/MdR/MnR; the authors claim state-of-the-art retrieval and the first complete benchmark for drone text-video retrieval.

Significance. If the dataset annotations are accurate and the comparisons are fair, the paper would provide a useful benchmark and a systematic evaluation of existing text-video retrieval methods in the drone domain. The method itself is mostly a composition of known components, but the multi-granularity combination and the top-k selection with dynamic temperature are interesting to test on aerial video. However, the central claims are currently undermined by the lack of human validation of the LLM-generated captions, by an overbroad SOTA claim on CapERA, and by ablation results that do not consistently support one of the three proposed contributions. The dataset and code are not released at submission, so reproducibility cannot be verified. With careful revision — human-validated annotation statistics, narrowed claims, and corrected ablation interpretation — this could become a solid contribution.

major comments (4)
  1. [Sec. 3.1, Steps 2–3; Figs. 2–4] The benchmark's ground truth is entirely model-generated: frame descriptions come from Qwen2.5-VL-7B and video-level captions from Qwen3. No human verification, no sample audit, and no inter-annotator agreement is reported. The prompt instructing the VLM to be 'objective' and 'not infer' is not evidence of compliance; VLMs are known to hallucinate objects, colors, and actions, and the six-frame sampling can omit transient events. Because every DVTMD retrieval number in Tables 2 and 3 and the 'first complete benchmark' claim depends on caption accuracy, the central contribution is not yet supported. Please provide a human-validated subset, report hallucination/omission rates, and, ideally, release a corrected annotation set.
  2. [Abstract; Table 2] The abstract claims state-of-the-art performance for drone text-video retrieval, but on CapERA TCMA does not beat existing methods. In Table 2, TCMA obtains ViT-B/32 R@1 14.9 vs UniAdapter 15.7, and ViT-B/16 R@1 16.1 vs 16.5 for Clip4Clip/EERCF/TempMe. Thus the SOTA claim is unsupported on one of the two benchmarks used. Please restrict the claim to DVTMD, or provide a fair comparison with statistical significance and a description of the baseline training/evaluation protocol.
  3. [Table 7; Sec. 5.3] The dynamic temperature ablation does not show 'consistent improvements across recall metrics' as stated. T2V R@1 decreases from 45.7 without tau_dyn to 45.5 with tau_dyn, while T2V R@10 improves from 84.0 to 84.6 and V2T improves from 42.0 to 42.8. This is a mixed, partially negative effect. Since the dynamic temperature is one of the three named contributions, this should be explained, tested for significance, or the claim should be revised.
  4. [Sec. 5.1; Sec. 5.3] Hyperparameters Kp=3, Kw=8, lambda_video:lambda_frame:lambda_patch=5:5:1, alpha=0.05, and beta=0.001 appear to be selected on DVTMD with no validation split or selection procedure described. If these choices were made by evaluating on the test set, the reported gains may be optimistic. It is also unclear whether the same hyperparameters were used for the CapERA experiments. Please state the model-selection protocol and report results across seeds or with error bars.
minor comments (6)
  1. [Table 1] The caption reads 'Table 3.2: Dataset Statistics'; it should be 'Table 1'.
  2. [Table 3] The UniAdapter row contains a malformed entry '2.07.0' and '- / -', which should be cleaned for readability.
  3. [Eq. (11)] The text says 'we augment each patch with the sentence feature' but the equation computes word saliency; it should say 'each word'.
  4. [Eqs. (18)–(19)] In Eq. (18), the denominator L_v is not defined, and Eq. (19) sums over L words although the selected set has size K_w. Please clarify the indexing.
  5. [Table 4] The Video+Frame and Video+Frame+Patch rows have identical T2V R@1/R@5/R@10 values, which contradicts the sentence in Sec. 5.3 that 'each additional level consistently improves retrieval performance.' This should be discussed.
  6. [Eq. (14)] The video-level feature v used in patch selection is not defined in that subsection; presumably it is the mean-pooled video embedding from Sec. 4.2.1.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: TCMA's modules are learned end-to-end and CapERA provides independent support; DVTMD caption validity is a data-quality risk, not circularity.

full rationale

TCMA's method is trained end-to-end with a hierarchical contrastive-plus-Pearson loss (Eqs. 20–26); no component is solved for from the reported retrieval scores. The dynamic temperature (Eq. 8) is a learned MLP, not a fit to R@1 targets, and the word/patch selection modules are also learned, end-to-end modules rather than fitted inputs renamed as predictions. DVTMD's captions are indeed generated by Qwen2.5-VL and Qwen3 without human verification (Sec. 3.1, Steps 2–3), which is a benchmark-validity and reproducibility concern, but it is not derivation circularity: the model's outputs are not defined in terms of those captions, and the same method is also evaluated on the externally annotated CapERA benchmark. Self-citations [21] and [22] appear only as related-work background ('Recent efforts in parameter-efficient fine-tuning [21] and large-scale benchmarking [22] further demonstrated the promise of FMs in this domain') and are not load-bearing for TCMA's claims. No equation reduces to its own input and no fitted parameter is presented as an independent prediction. The SOTA claim rests on empirical comparison and is vulnerable to dataset-quality issues, but not to circular reasoning.

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

No theoretical entities (new particles, forces, dimensions) are introduced. DVTMD is a dataset artifact, not an explanatory entity. The central claim rests mainly on the reliability of LLM-generated captions and on standard contrastive-learning assumptions.

free parameters (4)
  • Kp (patches selected per frame) = 3
    Chosen empirically in Section 5.1; ablation in Table 6 shows small variations around 2-8.
  • Kw (words selected per caption) = 8
    Chosen empirically in Section 5.1; ablation Table 5 shows differences of about 0.9 R@1 versus no selection.
  • Hierarchical loss weights lambda_video:lambda_frame:lambda_patch = 5:5:1
    Set in Section 5.1 without sensitivity analysis; directly controls contribution of each alignment level.
  • Pearson regularization weights alpha, beta = alpha=0.05, beta=0.001
    Set empirically in Section 5.1; no ablation is reported for these values.
assumptions (4)
  • domain assumption CLIP's pretrained visual and text encoders transfer to aerial video frames without domain-specific architectural changes.
    Section 4.1 initializes with CLIP and fine-tunes with very low learning rates; if CLIP features are weak for aerial perspectives, the entire aggregation stack is built on poor inputs.
  • domain assumption Qwen2.5-VL and Qwen3 generate objective, factual captions that faithfully describe ERA frames and videos.
    Section 3.1 Steps 2-3 rely entirely on VLM outputs; no human validation or agreement metric is reported.
  • domain assumption All in-batch non-matching pairs are true negatives during contrastive training.
    Section 4.3 Eq. 20 treats every j not equal to i as a negative. DVTMD has five captions per video, so if two captions from the same video appear in one batch, the loss incorrectly pushes them apart.
  • domain assumption ERA's 25 event categories and the adopted train/test split are appropriate for training and evaluating drone video-text retrieval.
    DVTMD is built on ERA (Section 3.1), and the split 1,473 train / 1,391 test is used without explaining how it was created.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TCMA: Text-Conditioned Multi-granularity Alignment for Drone Cross-Modal Text-Video Retrieval." pith.science (2026). https://pith.science/paper/3TRGZ7OK

@misc{pith2026251010180,
  author       = {Pith},
  title        = {Pith review of: TCMA: Text-Conditioned Multi-granularity Alignment for Drone Cross-Modal Text-Video Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3TRGZ7OK}},
  note         = {Machine review of arXiv:2510.10180}
}
read the original abstract

Unmanned aerial vehicles (UAVs) have become powerful platforms for real-time, high-resolution data collection, producing massive volumes of aerial videos. Efficient retrieval of relevant content from these videos is crucial for applications in urban management, emergency response, security, and disaster relief. While text-video retrieval has advanced in natural video domains, the UAV domain remains underexplored due to limitations in existing datasets, such as coarse and redundant captions. Thus, in this work, we construct the Drone Video-Text Match Dataset (DVTMD), which contains 2,864 videos and 14,320 fine-grained, semantically diverse captions. The annotations capture multiple complementary aspects, including human actions, objects, background settings, environmental conditions, and visual style, thereby enhancing text-video correspondence and reducing redundancy. Building on this dataset, we propose the Text-Conditioned Multi-granularity Alignment (TCMA) framework, which integrates global video-sentence alignment, sentence-guided frame aggregation, and word-guided patch alignment. To further refine local alignment, we design a Word and Patch Selection module that filters irrelevant content, as well as a Text-Adaptive Dynamic Temperature Mechanism that adapts attention sharpness to text type. Extensive experiments on DVTMD and CapERA establish the first complete benchmark for drone text-video retrieval. Our TCMA achieves state-of-the-art performance, including 45.5% R@1 in text-to-video and 42.8% R@1 in video-to-text retrieval, demonstrating the effectiveness of our dataset and method. The code and dataset will be released.

Figures

Figures reproduced from arXiv: 2510.10180 by the authors.

Figure 1
Figure 1. Illustration of the characteristic in the drone text-video retrieval task. Video-level: global sentence [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The construction pipeline of our DVTMD benchmark: Step 1) frame sampling, Step 2) detailed [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The designed prompt template for generating detailed and objective frame-level descriptions. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The designed prompt template for generating summarized captions. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Graphical analysis of the captions. (a) Caption length distribution across videos, with most [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Similarity heatmaps of video-caption pairs based on weighted BLEU and METEOR scores. (a) [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: The overview of our proposed TCMA. It consists of three components: (1) Text and Visual [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: (a) Frame Aggregation Module, which carries out sentence-guided frame aggregation to form [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: In the bottom-left case, our model accurately identifies the number of objects (“two [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 9
Figure 9. Figure 9: Visualization of text–video retrieval examples. Our retrieval results are compared with other [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Predictive Modeling for High Impact Active Learning Classrooms

    physics.ed-ph 2026-03 unverdicted novelty 6.0 of 10

    Four classroom-time variables predict physics concept learning, and classes with 10–20% group worksheets, 20–40% group clickers, and ≥2 student questions per hour show effect sizes above 2.

Reference graph

Works this paper leans on

40 extracted references · 7 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Y. Tian, F. Lin, Y. Li, T. Zhang, Q. Zhang, X. Fu, J. Huang, X. Dai, Y. Wang, C. Tian, B. Li, Y. Lv, L. Kovács, F.-Y. Wang, Uavs meet llms: Overviews and perspectives towards agentic low-altitude mobility, Information Fusion 122 (2025) 103158

  2. [2]

    Zhang, S

    J. Zhang, S. Xu, Y. Zhao, J. Sun, S. Xu, X. Zhang, Aerial orthoimage generation for uav remote sensing: Review, Information Fusion 89 (2023) 91–120

  3. [3]

    S. P. H. Boroujeni, A. Razi, S. Khoshdel, F. Afghah, J. L. Coen, L. O’Neill, P. Fule, A. Watts, N.-M. T. Kokolakis, K. G. Vamvoudakis, A comprehensive survey of research towards ai-enabled unmanned aerial systems in pre-, active-, and post-wildfire manage- ment, Information Fusion 108 (2024) 102369

  4. [4]

    K. W. Tong, Z. Shi, G. Zhu, Y. Duan, Y. Hou, E. Q. Wu, L. Zhu, Large-scale aerial scene perception based on self-supervised multi-view stereo via cycled generative adversarial network, Information Fusion 109 (2024) 102399

  5. [5]

    J. Xu, T. Mei, T. Yao, Y. Rui, Msr-vtt: A large video description dataset for bridging video and language, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 5288–5296

  6. [6]

    D. Chen, W. B. Dolan, Collecting highly parallel data for paraphrase evaluation, in: Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, 2011, pp. 190–200

  7. [7]

    Bashmal, Y

    L. Bashmal, Y. Bazi, M. M. Al Rahhal, M. Zuair, F. Melgani, Capera: Captioning events in aerial videos, Remote Sensing 15 (8) (2023) 2139

  8. [8]

    L. Mou, Y. Hua, P. Jin, X. X. Zhu, Era: A data set and deep learning benchmark for event recognition in aerial videos [software and data sets], IEEE Geoscience and Remote Sensing Magazine 8 (4) (2020) 125–133

Show all 40 references
  1. [9]

    Mishkin, J

    A.Radford, J.W.Kim, C.Hallacy, A.Ramesh, G.Goh, S.Agarwal, G.Sastry, A.Askell, P. Mishkin, J. Clark, et al., Learning transferable visual models from natural language supervision, in: International conference on machine learning, PmLR, 2021, pp. 8748– 8763. 20

  2. [10]

    H. Luo, L. Ji, M. Zhong, Y. Chen, W. Lei, N. Duan, T. Li, Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning, Neurocomputing 508 (2022) 293–304

  3. [11]

    S. K. Gorti, N. Vouitsis, J. Ma, K. Golestan, M. Volkovs, A. Garg, G. Yu, X-pool: Cross-modal language-video attention for text-video retrieval, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5006– 5015

  4. [12]

    Y. Liu, P. Xiong, L. Xu, S. Cao, Q. Jin, Ts2-net: Token shift and selection transformer for text-video retrieval, in: European conference on computer vision, Springer, 2022, pp. 319–335

  5. [13]

    B. Fang, W. Wu, C. Liu, Y. Zhou, Y. Song, W. Wang, X. Shu, X. Ji, J. Wang, Uatvr: Uncertainty-adaptive text-video retrieval, in: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 2023, pp. 13723–13733

  6. [14]

    L. Shen, T. Hao, T. He, S. Zhao, Y. Zhang, P. Liu, Y. Bao, G. Ding, Tempme: Video temporal token merging for efficient text-video retrieval, arXiv preprint arXiv:2409.01156 (2024)

  7. [15]

    Abdullah, Y

    T. Abdullah, Y. Bazi, M. M. Al Rahhal, M. L. Mekhalfi, L. Rangarajan, M. Zuair, Textrs: Deep bidirectional triplet network for matching text to remote sensing images, Remote Sensing 12 (3) (2020) 405

  8. [16]

    Z. Yuan, W. Zhang, K. Fu, X. Li, C. Deng, H. Wang, X. Sun, Exploring a fine-grained multiscale method for cross-modal remote sensing image retrieval, arXiv preprint arXiv:2204.09868 (2022)

  9. [17]

    Zhang, J

    W. Zhang, J. Li, S. Li, J. Chen, W. Zhang, X. Gao, X. Sun, Hypersphere-based remote sensing cross-modal text–image retrieval via curriculum learning, IEEE Transactions on Geoscience and Remote Sensing 61 (2023) 1–15

  10. [18]

    J. Pan, Q. Ma, C. Bai, A prior instruction representation framework for remote sensing image-text retrieval, in: Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 611–620

  11. [19]

    Z. Ji, C. Meng, Y. Zhang, Y. Pang, X. Li, Knowledge-aided momentum contrastive learning for remote-sensing image text retrieval, IEEE Transactions on Geoscience and Remote Sensing 61 (2023) 1–13

  12. [20]

    J. Li, D. Li, C. Xiong, S. Hoi, Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, in: International conference on machine learning, PMLR, 2022, pp. 12888–12900

  13. [21]

    Y. Yuan, Y. Zhan, Z. Xiong, Parameter-efficient transfer learning for remote sensing image–text retrieval, IEEE Transactions on Geoscience and Remote Sensing 61 (2023) 1–14. 21

  14. [22]

    Y.Zhan, Z.Xiong, Y.Yuan, SkyEyeGPT:Unifyingremotesensingvision-languagetasks via instruction tuning with large language model, ISPRS Journal of Photogrammetry and Remote Sensing 221 (2025) 64–77

  15. [23]

    Huang, Y

    J. Huang, Y. Chen, S. Xiong, X. Lu, Visual contextual semantic reasoning for cross- modaldroneimage-textretrieval, IEEETransactionsonGeoscienceandRemoteSensing (2024)

  16. [24]

    Bashmal, S

    L. Bashmal, S. M. A. Mehmadi, Y. Bazi, M. M. A. Rahhal, M. Zuair, Text-to-event retrieval in aerial videos, IEEE Geoscience and Remote Sensing Letters 21 (2024) 1–5

  17. [25]

    Z. He, S. Sunkara, X. Zang, Y. Xu, L. Liu, N. Wichers, G. Schubiner, R. Lee, J. Chen, Actionbert: Leveraging user actions for semantic understanding of user interfaces, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35, 2021, pp. 5931– 5938

  18. [26]

    Miech, J.-B

    A. Miech, J.-B. Alayrac, L. Smaira, I. Laptev, J. Sivic, A. Zisserman, End-to-end learn- ing of visual representations from uncurated instructional videos, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9879– 9889

  19. [27]

    J. Lei, L. Li, L. Zhou, Z. Gan, T. L. Berg, M. Bansal, J. Liu, Less is more: Clipbert for video-and-language learning via sparse sampling, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 7331–7341

  20. [28]

    M. Bain, A. Nagrani, G. Varol, A. Zisserman, Frozen in time: A joint video and im- age encoder for end-to-end retrieval, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 1728–1738

  21. [29]

    P. Jin, H. Li, Z. Cheng, K. Li, X. Ji, C. Liu, L. Yuan, J. Chen, Diffusionret: Gen- erative text-video retrieval with diffusion model, in: Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 2470–2481

  22. [30]

    W. Wu, H. Luo, B. Fang, J. Wang, W. Ouyang, Cap4video: What can auxiliary captions do for text-video retrieval?, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 10704–10713

  23. [31]

    X. Jin, B. Zhang, W. Gong, K. Xu, X. Deng, P. Wang, Z. Zhang, X. Shen, J. Feng, Mv-adapter: Multimodal video transfer learning for video text retrieval, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 27144–27153

  24. [32]

    K. Tian, R. Zhao, Z. Xin, B. Lan, X. Li, Holistic features are almost sufficient for text- to-video retrieval, in: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 17138–17147

  25. [33]

    Y. Ge, Y. Ge, X. Liu, D. Li, Y. Shan, X. Qie, P. Luo, Bridging video-text retrieval with multiple choice questions (2022).arXiv:2201.04850. 22

  26. [34]

    Wang, Y.-L

    Z. Wang, Y.-L. Sung, F. Cheng, G. Bertasius, M. Bansal, Unified coarse-to-fine align- ment for video-text retrieval, in: Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 2816–2827

  27. [35]

    K. Tian, Y. Cheng, Y. Liu, X. Hou, Q. Chen, H. Li, 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, 2024, pp. 5207–5214

  28. [36]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y. Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y. Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, J. Lin, Qwen2.5-vl technical report (2025).arXiv:2502.13923

  29. [37]

    Y. Yao, T. Yu, A. Zhang, C. Wang, J. Cui, H. Zhu, T. Cai, H. Li, W. Zhao, Z. He, Q. Chen, H. Zhou, Z. Zou, H. Zhang, S. Hu, Z. Zheng, J. Zhou, J. Cai, X. Han, G. Zeng, D. Li, Z. Liu, M. Sun, Minicpm-v: A gpt-4v level mllm on your phone (2024). arXiv:2408.01800

  30. [38]

    Z. Chen, W. Wang, Y. Cao, Y. Liu, Z. Gao, E. Cui, J. Zhu, S. Ye, H. Tian, Z. Liu, L. Gu, X. Wang, Q. Li, Y. Ren, Z. Chen, J. Luo, J. Wang, T. Jiang, B. Wang, C. He, B. Shi, X. Zhang, H. Lv, Y. Wang, W. Shao, P. Chu, Z. Tu, T. He, Z. Wu, H. Deng, J. Ge, K. Chen, K. Zhang, L. Wa...

  31. [39]

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, C. Zheng, D. Liu, F. Zhou, F. Huang, F. Hu, H. Ge, H. Wei, H. Lin, J. Tang, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Zhou, J. Lin, K. Dang, K. Bao, K. Yang, L. Yu, L. Deng, M. L...

  32. [40]

    H. Lu, Y. Huo, G. Yang, Z. Lu, W. Zhan, M. Tomizuka, M. Ding, Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling (2023).arXiv:2302. 06605. 23

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.