Pith. sign in

REVIEW 3 major objections 6 minor 38 references

This paper claims that dense, 31-word modification texts—not a bigger model—are what push composed video retrieval past the previous state of the art.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A new benchmark with much longer, denser modification texts, plus a single-encoder fusion model, raises composed video retrieval Recall@1 by 3.4 points on its own test set.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Dense-WebVid-CoVR is a genuinely useful dataset, but the 3.4% gain is provisional until the train/test overlap is audited and the text-only ceiling is addressed. the 3 major comments →

arxiv 2508.14039 v1 pith:AFMLQRMP submitted 2025-08-19 cs.CV

Beyond Simple Edits: Composed Video Retrieval with Dense Modifications

Omkar Thawakar , Dmitry Demidov , Ritesh Thawkar , Rao Muhammad Anwer , Mubarak Shah , Fahad Shahbaz Khan , Salman Khan This is my paper
classification cs.CV
keywords composed video retrievaldense modification textDense-WebVid-CoVRcross-attention fusiongrounding text encodervideo-text retrieval benchmarkzero-shot transferfine-grained video understanding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper argues that composed video retrieval—finding a target video from a reference video plus a text describing what changed—has been held back by modification texts that are too short and generic. To fix it, the authors build Dense-WebVid-CoVR, a 1.6-million-sample benchmark whose modification texts average 31 words instead of the previous 4.6, and pair it with a model that fuses reference video, dense description, and modification text in one grounding encoder. On the new test set the model reaches 71.26% Recall@1, 3.4 points above the best prior method under the same training data, input modalities, and backbone. The paper's companion claim is that the dense modification texts themselves are the source of the gain, since the same model drops to 68.88% Recall@1 when tested with the old short texts.

Core claim

The central claim is that fine-grained composed video retrieval is primarily a data problem: short modification texts cannot express the subtle visual and temporal changes that distinguish a target video from a query video, so even a capable fusion model is starved of signal. The discovery is that replacing 4.6-word modifications with 31.2-word, context-grounded modifications—generated by an LLM from detailed video captions and manually verified—raises Recall@1 by 3.4 points, from 67.86 to 71.26, with the same model and training setup. The model itself is a unified fusion: a weighted combination of the middle-frame visual embedding and the dense description embedding is fed together with the

What carries the argument

The load-bearing object is the grounding text encoder under a unified fusion scheme. Instead of fusing query video, description, and modification text pairwise, the model forms embs = (1−α)g(q) + αe(d), then passes this weighted visual-plus-description embedding jointly with the modification text t through a grounding encoder f that uses self-attention, cross-attention, and a feed-forward network, producing f(q, d, t). Cross-attention grounds each modification word in the visual/description context, which is what makes long modification texts usable. The dataset machinery is equally central: detailed descriptions from Gemini-Pro with a BLIP-based hallucination check, modification texts from

Load-bearing premise

The load-bearing premise is that the manually verified test set measures compositional video understanding rather than distribution matching: the paper does not establish that none of the 131K training videos overlap the 3.2K test videos drawn from the same WebVid pool, nor that the GPT-4o modification texts do not let a text-only model solve the benchmark (its own text-only baseline reaches 50.12% Recall@1).

What would settle it

Compute the overlap between the 131K training video identities and the 3.2K test triplets against WebVid's video IDs; additionally, run a text-only retriever that encodes the modification text and ranks target videos by caption similarity without any query video. If a non-trivial fraction of test videos appear in training, or if the text-only retriever matches or exceeds 71.26% Recall@1, the benchmark does not require the visual modality as claimed.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Dense modification texts become the standard for composed video retrieval evaluation, since short-text benchmarks understate what models can do.
  • The unified fusion scheme replaces pairwise fusion; it is simpler, roughly three times faster, and more accurate on fine-grained queries.
  • A model trained on dense modifications transfers zero-shot to egocentric video retrieval and to composed image retrieval, so the benefit is not limited to one dataset.
  • Retrieval systems can handle queries that specify counts, colors, backgrounds, and temporal order, not just simple object swaps.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • My inference: because the paper's own text-only baseline reaches 50.12% Recall@1 on the new benchmark, a large part of the gain may come from easier lexical matching between the dense modification text and the target caption; ablations that remove caption overlap would separate true visual grounding from language matching.
  • My inference: without a published train/test disjointness check, the 3.4-point gain could partly reflect near-duplicate videos across the WebVid-2M and WebVid-8M pools; a leakage-controlled split would make the benchmark's numbers more interpretable.
  • My inference: the same dataset-generation pipeline—detailed captions, LLM difference text, and manual verification—could be applied to other retrieval domains such as audio, egocentric action, or long-video moment retrieval, where short difference text is also a bottleneck.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces Dense-WebVid-CoVR, a large-scale composed-video-retrieval benchmark built from WebVid-CoVR triplets by generating dense video descriptions with Gemini-Pro and dense modification texts with GPT-4o, with human verification of the test/validation sets and 100k training triplets. The paper also proposes a retrieval model that fuses the query video, a dense description, and the modification text in a single BLIP grounding encoder using weighted fusion plus cross-attention, trained with a contrastive loss. On the new test set the model reports 71.26% Recall@1, a 3.4-point improvement over the prior method of Thawakar et al. under the same training data, input modalities, and backbone, and it reports favorable transfer results on EgoCVR, CIRR, and FashionIQ.

Significance. If the new benchmark is valid, the dense-annotation direction is useful for composed video retrieval, and the proposed unified fusion is a simple, effective baseline. The paper’s strongest evidence is the transfer to established CIRR/FashionIQ benchmarks and the zero-shot EgoCVR results, which are not affected by the main benchmark-construction concern. However, the headline CoVR claim rests on a test set that is not independent of the text-generation pipeline, and the paper does not audit train/test overlap. The paper’s own text-only baseline (50.12 R@1, Table 2 row 11) indicates a substantial text-only ceiling that is never discussed, so the central 3.4-point claim is not yet interpretable as evidence of compositional visual understanding.

major comments (3)
  1. [Sec. 3.1–3.2, Table 2, Table 6] The evaluation benchmark is not independent of the modification-text generation process. In Sec. 3.1 and Figure 8, GPT-4o generates the modification text from the target video’s description (caption2) together with the source caption, and the same generator is used for both training and test modifications. A model trained on this distribution can learn to invert the text-to-target-caption mapping. The paper’s own text-only baseline already reaches 50.12 R@1 (Table 2, row 11), which is a large ceiling on the 71.26 R@1 multimodal score and is not addressed. The statement in Sec. 3.2 that query-structuring strategies ‘prevent text-only retrieval’ is therefore not supported quantitatively. This also confounds the ablation in Table 6 (68.88 vs 71.26): the dense text may be more predictive of the target caption, not more descriptive of the visual change. Please report text-only R@1 for both De
  2. [Sec. 3.2, Sec. 5, Table 2] Train/test video overlap is never audited. Sec. 3.2 states that training triplets are drawn from WebVid-2M and that test/validation sets are drawn from WebVid-8M, and Sec. 5 says the evaluation sets come ‘from the WebVid10M corpus.’ Since WebVid-2M and WebVid-8M are both subsets of the same source pool, the paper must demonstrate that no test query/target video or near-duplicate frame sequence appears among the 131K training videos. Without such a disjointness check, the 71.26 R@1 and the 3.4-point margin over Thawakar et al. (Table 2, rows 14–15) could be inflated by memorization or near-duplicate visual matching. Please report a video-ID and frame-level overlap analysis and rerun the main comparison on the disjoint subset, with test IDs released.
  3. [Sec. 5.1, Table 4, Table 6] The claimed benefit of dense modification texts is not separated from text-only target-caption leakage. Table 4 and Table 6 compare models trained or evaluated with Dense texts versus WebVid-CoVR texts, but neither controls for how much of the target video is recoverable from the text alone. Given that the text-only model in Table 2 already reaches 50.12 R@1, the improvements in Tables 4 and 6 could be explained by the dense texts being more informative about the target caption rather than by improved multimodal reasoning. Please add a text-only retrieval condition for each text type and show that the dense-text gain over WebVid-CoVR texts remains when text-only performance is comparable, or otherwise restrict the benchmark to modifications that cannot be solved from text alone.
minor comments (6)
  1. [Sec. 5.1] In the results paragraph, ‘our approach outperforms [30] with Recall@1 and Recall@5 score of 71.3 and 81.1’ should read 89.1 for R@5, matching Table 2.
  2. [Sec. 3.2 vs Sec. 5] The test set size is given as 3,000 in Sec. 3.2 and 3.2K in Sec. 5; please reconcile.
  3. [Eq. (1), Implementation Details] α is called a learnable parameter in Sec. 4.1 but is later described as a validation-set-derived value (0.36). Please clarify whether α is learned or tuned.
  4. [Sec. 4.1] The model uses only the middle frame of each video, while the introduction motivates temporal understanding. Please clarify what temporal information the model can actually exploit and how the temporal claims in the benchmark are measured.
  5. [Sec. 5.1] The claim that the model is ‘3× faster than [30]’ is not supported by any runtime table or measurement; please provide the comparison or remove the claim.
  6. [References] References [20] and [21] appear to be the same paper (‘Data roaming and early fusion for composed image retrieval’); please deduplicate.

Circularity Check

1 steps flagged

Dense modification texts are generated from the target video's caption, so the headline gain on Dense-WebVid-CoVR is partly encoded in the query by construction; external benchmarks keep the method itself from being fully circular.

specific steps
  1. self definitional [Section 3.1 (Modification-Text Generation) and Supplementary Fig. 8 prompt]
    "For this, we employ GPT-4o [14] to generate modification texts. We provide the model with existing triplets from WebVid, including both the original video captions and their corresponding modification texts, to guide the generation process. ... The prompt instructs the model to analyze two given captions (caption1 and caption2), identify the key differences, and formulate a concise modification text to transform caption1 into caption2."

    The query modification text t in each test triplet is produced from the target video's caption (caption2) by GPT-4o. The benchmark thus defines t as a function of the retrieval target v*: the denser t is, the more it paraphrases the target's caption. A model trained on the same generation pipeline can learn to invert that function, so high R@1 reflects text-to-video matching rather than compositional understanding. The paper's own text-only ablation reaches 50.12 R@1 (Table 2, row 11), contradicting the claimed query structuring that 'prevent[s] text-only retrieval' (Section 3.2). The 3.4% gain credited to dense modification texts (Table 6: 71.26 vs 68.88) is therefore partly built into the benchmark: dense texts are more informative of the target by construction.

full rationale

The paper's method itself is not circular: it uses external backbones (BLIP-2, CLIP) and is evaluated on independent benchmarks (CIRR, FashionIQ, EgoCVR), where it performs well. The comparison to [30] is a same-author baseline but is not used as a proof step. The main circularity lies in the Dense-WebVid-CoVR benchmark construction: modification texts are generated by GPT-4o from the source and target video captions with the explicit goal of guiding retrieval to the target, making the target a known function of the query text. The paper's own text-only result (50.12 R@1) demonstrates a large text-only ceiling, and the dense-text improvement is partly a benchmark artifact rather than evidence of compositional understanding. This is a partial, benchmark-level circularity, not a fully tautological method, so score 6 rather than higher.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claims rest on domain assumptions about the generation pipeline and about the benchmark's isolation from the training distribution. The only genuinely fitted scalar is the fusion weight α; the 0.4 similarity thresholds are ad hoc quality gates. The most consequential axiom is that GPT-4o modification texts do not encode the target caption too directly, because the entire benchmark is generated by that function at both train and test time. No new physical or mathematical entities are postulated.

free parameters (3)
  • Fusion weight α = 0.36 (tuned on validation set)
    Eq. 1: embs = (1-α)g(q) + αe(d). Learned and selected on the validation split of the same benchmark; directly controls how much of the query representation comes from text versus video.
  • Cosine similarity thresholds for caption and modification-text filtering = 0.4
    Set empirically in Section 3 and Suppl. C/E to reject hallucinated Gemini-Pro captions and low-quality modification texts; affects the composition of the training and test data.
  • Contrastive hyperparameters = τ=0.07, λ=1, β=0.5
    Temperature, loss weight, and hard-negative weighting taken from [24] (Eq. 2); standard in the literature but hand-chosen values that affect the reported recall numbers.
axioms (5)
  • domain assumption Pretrained ViT-L and BLIP-2 embeddings are fine-grained enough to support dense modification retrieval from a single middle frame (15 sampled frames).
    Section 4.1 selects the middle frame following [30, 31]; the whole approach depends on this coarse temporal sampling not destroying the fine-grained signal the benchmark is designed to test.
  • domain assumption Gemini-Pro captions that pass a BLIP cosine check at 0.4 are accurate dense descriptions.
    Section 3 and Suppl. C rely on this threshold to guarantee caption quality; no human audit of the caption set is reported, only of modification texts.
  • domain assumption GPT-4o modification texts encode true visual differences and do not leak the target caption verbatim.
    Section 3.1 generates texts from both captions; Section 3.2 adds query structuring to reduce text-only solvability, but the paper's own text-only R@1 of 50.12 (Table 2) shows substantial leakage remains.
  • domain assumption Training videos (WebVid-2M) and test videos (WebVid-8M) are disjoint within the shared WebVid-10M pool.
    Sections 3.2 and 5 define the splits without any duplicate-video audit, so the headline retrieval numbers could be inflated by memorization if overlap exists.
  • standard math Standard transformer and contrastive-learning machinery behaves as in [24, 30, 31].
    Eq. 2 and the grounding encoder follow established BLIP-2/Q-Former-style practices; treated as background knowledge.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Simple Edits: Composed Video Retrieval with Dense Modifications." pith.science (2026). https://pith.science/paper/AFMLQRMP

@misc{pith2026250814039,
  author       = {Pith},
  title        = {Pith review of: Beyond Simple Edits: Composed Video Retrieval with Dense Modifications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AFMLQRMP}},
  note         = {Machine review of arXiv:2508.14039}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Composed video retrieval is a challenging task that strives to retrieve a target video based on a query video and a textual description detailing specific modifications. Standard retrieval frameworks typically struggle to handle the complexity of fine-grained compositional queries and variations in temporal understanding limiting their retrieval ability in the fine-grained setting. To address this issue, we introduce a novel dataset that captures both fine-grained and composed actions across diverse video segments, enabling more detailed compositional changes in retrieved video content. The proposed dataset, named Dense-WebVid-CoVR, consists of 1.6 million samples with dense modification text that is around seven times more than its existing counterpart. We further develop a new model that integrates visual and textual information through Cross-Attention (CA) fusion using grounded text encoder, enabling precise alignment between dense query modifications and target videos. The proposed model achieves state-of-the-art results surpassing existing methods on all metrics. Notably, it achieves 71.3\% Recall@1 in visual+text setting and outperforms the state-of-the-art by 3.4\%, highlighting its efficacy in terms of leveraging detailed video descriptions and dense modification texts. Our proposed dataset, code, and model are available at :https://github.com/OmkarThawakar/BSE-CoVR

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 24 canonical work pages · 4 internal anchors

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 11

    Pith/arXiv arXiv 2023

  2. [2]

    isearle: Improving textual inversion for zero-shot composed image retrieval

    Lorenzo Agnolucci, Alberto Baldrati, Marco Bertini, and Alberto Del Bimbo. isearle: Improving textual inversion for zero-shot composed image retrieval. arXiv preprint arXiv:2405.02951, 2024. 4

    Pith/arXiv arXiv 2024

  3. [3]

    A clip-hitchhiker’s guide to long video retrieval

    Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisser- man. A clip-hitchhiker’s guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022. 3

    Pith/arXiv arXiv 2022

  4. [4]

    Effective conditioned and composed image retrieval combining clip-based features

    Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Al- berto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21466–21474, 2022. 8

    work page 2022

  5. [5]

    In- structpix2pix: Learning to follow image editing instructions

    Tim Brooks, Aleksander Holynski, and Alexei A Efros. In- structpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392–18402, 2023. 2

    work page 2023

  6. [6]

    Learning joint visual seman- tic matching embeddings for language-guided retrieval

    Yanbei Chen and Loris Bazzani. Learning joint visual seman- tic matching embeddings for language-guided retrieval. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXII 16, pages 136–152. Springer, 2020. 8

    work page 2020

  7. [7]

    Image search with text feedback by visiolinguistic attention learning

    Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3001–3011, 2020. 8

    work page 2020

  8. [8]

    Artemis: Attention-based retrieval with text-explicit matching and implicit similarity

    Ginger Delmas, Rafael Sampaio de Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. arXiv preprint arXiv:2203.08101, 2022. 8

    Pith/arXiv arXiv 2022

  9. [9]

    Modality-agnostic attention fusion for visual search with text feedback

    Eric Dodds, Jack Culpepper, Simao Herdade, Yang Zhang, and Kofi Boakye. Modality-agnostic attention fusion for visual search with text feedback. arXiv preprint arXiv:2007.00145, 2020. 8

    Pith/arXiv arXiv 2007

  10. [10]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 6

    Pith/arXiv arXiv 2010

  11. [11]

    Fash- ionvlp: Vision language transformer for fashion retrieval with feedback

    Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fash- ionvlp: Vision language transformer for fashion retrieval with feedback. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 14105–14115, 2022. 8

    work page 2022

  12. [12]

    Compodiff: Versatile com- posed image retrieval with latent diffusion

    Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile com- posed image retrieval with latent diffusion. arXiv preprint arXiv:2303.11916, 2023. 2, 8

    Pith/arXiv arXiv 2023

  13. [13]

    Egocvr: An egocentric bench- mark for fine-grained composed video retrieval

    Thomas Hummel, Shyamgopal Karthik, Mariana-Iuliana Georgescu, and Zeynep Akata. Egocvr: An egocentric bench- mark for fine-grained composed video retrieval. European Conference on Computer Vision (ECCV), 2024. 1, 2, 6, 7, 11

    work page 2024

  14. [14]

    Gpt-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2, 3

    Pith/arXiv arXiv 2024

  15. [15]

    SAC: Semantic Attention Composition for Text-Conditioned Image Retrieval

    Surgan Jandial, Ayush Chopra, Pinkesh Badjatiya, Pranit Chawla, Mausoom Sarkar, and Balaji Krishnamurthy. Trace: Transform aggregate and compose visiolinguistic represen- tations for image search with text feedback. arXiv preprint arXiv:2009.01485, 7:7, 2020. 8

    work page internal anchor Pith review Pith/arXiv arXiv 2009

  16. [16]

    Sac: Semantic attention composition for text-conditioned image retrieval

    Surgan Jandial, Pinkesh Badjatiya, Pranit Chawla, Ayush Chopra, Mausoom Sarkar, and Balaji Krishnamurthy. Sac: Semantic attention composition for text-conditioned image retrieval. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4021–4030, 2022. 8

    work page 2022

  17. [17]

    Vision-by-language for training-free com- positional image retrieval

    Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free com- positional image retrieval. arXiv preprint arXiv:2310.09291,

    Pith/arXiv arXiv

  18. [18]

    Dual compositional learning in interactive image re- trieval

    Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image re- trieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1771–1779, 2021. 8

    work page 2021

  19. [19]

    Cosmo: Content-style modulation for image retrieval with text feed- back

    Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feed- back. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 802–812, 2021. 8

    work page 2021

  20. [21]

    Data Roaming and Quality Assessment for Composed Image Retrieval

    Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and early fusion for composed image retrieval. arXiv preprint arXiv:2303.09429, 2(3):7, 2023. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  21. [22]

    Blip- 2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip- 2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 3, 6

    Pith/arXiv arXiv 2023

  22. [23]

    Image retrieval on real-life images with pre- trained vision-and-language models

    Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre- trained vision-and-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 2125–2134, 2021. 2, 6, 8 9

    work page 2021

  23. [24]

    Filtering, distillation, and hard negatives for vision-language pre-training

    Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6967–6977, 2023. 5, 6

    work page 2023

  24. [25]

    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. In International conference on machine learning, pages 8748–8763. PMLR, 2021. 2

    2021

  25. [26]

    Fine-tuned clip models are efficient video learners

    Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6545–6554, 2023. 2

    work page 2023

  26. [27]

    Pic2word: Mapping pictures to words for zero-shot composed image re- trieval

    Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image re- trieval. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 19305–19314,

    work page

  27. [28]

    RTIC: Residual Learning for Text and Image Composition using Graph Convolutional Network

    Minchul Shin, Yoonjae Cho, Byungsoo Ko, and Geonmo Gu. Rtic: Residual learning for text and image compo- sition using graph convolutional network. arXiv preprint arXiv:2104.03015, 2021. 8

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  28. [29]

    Gemini: a family of highly capable multimodal models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2, 3, 11

    Pith/arXiv arXiv 2023

  29. [30]

    Composed video retrieval via enriched context and discriminative embeddings

    Omkar Thawakar, Muzammal Naseer, Rao Muhammad An- wer, Salman Khan, Michael Felsberg, Mubarak Shah, and Fahad Shahbaz Khan. Composed video retrieval via enriched context and discriminative embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26896–26906, 2024. 2, 4, 5, 6, 7, 8, 11, 13, 14, 18

    work page 2024

  30. [31]

    Covr: Learning composed video retrieval from web video captions

    Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. Covr: Learning composed video retrieval from web video captions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5270–5279, 2024. 1, 2, 3, 4, 6, 7, 8, 11, 12, 13, 15

    work page 2024

  31. [32]

    Composing text and image for image retrieval-an empirical odyssey

    Nam V o, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval-an empirical odyssey. In Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, pages 6439–6448, 2019. 8

    work page 2019

  32. [33]

    Composing text and image for image retrieval-an empirical odyssey

    Nam V o, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval-an empirical odyssey. In Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, pages 6439–6448, 2019. 2

    work page 2019

  33. [34]

    Fashion iq: A new dataset towards retrieving images by natural language feedback

    Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11307–11317,

    work page

  34. [35]

    Videoclip: Contrastive pre-training for zero-shot video-text understanding

    Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 2

    Pith/arXiv arXiv 2021

  35. [36]

    Clip-vip: Adapting pre- trained image-text model to video-language representation alignment

    Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pre- trained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022

    Pith/arXiv arXiv 2022

  36. [37]

    Taco: Token- aware cascade contrastive learning for video-text alignment

    Jianwei Yang, Yonatan Bisk, and Jianfeng Gao. Taco: Token- aware cascade contrastive learning for video-text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11562–11572, 2021. 2

    work page 2021

  37. [38]

    CurlingNet: Compositional Learning between Images and Text for Fashion IQ Data

    Youngjae Yu, Seunghwan Lee, Yuncheol Choi, and Gunhee Kim. Curlingnet: Compositional learning between images and text for fashion iq data. arXiv preprint arXiv:2003.12299,

    work page internal anchor Pith review Pith/arXiv arXiv 2003

  38. [2020]

    #$%#&%'())&*(%'#+&,(-#.-/01(2(%

    8 10 Beyond Simple Edits: Composed Video Retrieval with Dense Modifications Supplementary Material A. Supplementary Materials In this appendix, we provide a detailed discussion of the related work on both Composed Image Retrieval (CoIR) and Composed Video Retrieval (CoVR). CoVR, as an extension of CoIR, brings the challenge of video temporal dynamics and ...

    work page 2019

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.