Pith. sign in

REVIEW 2 major objections 5 minor 30 references

Coverage Matters: MarginMerge for Compressing Multi-Vector Visual Document Retrievers

T0 review · 2 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read MarginMerge compresses multi-vector visual document retrievers by preserving query-relevant coverage, cutting stored vectors by 90–95% while retaining 97–99% of average nDCG@5 across six datasets and two backbones.

desk verdict The compression results are probably real, but the 'zero-shot' transfer claim is contaminated by the prototype bank. read the letter →

arxiv 2608.02969 v1 pith:TIZWIJB5 submitted 2026-08-04 cs.IR

classification cs.IR
keywords multi-vectorretrievalvisualdocumentindexcompressionquery-relevantcoverageMaxSimlateinteractionColPaliQwenlearnedrepresentativesynthesis
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

Multi-vector visual document retrievers such as ColPali and ColQwen store hundreds of patch embeddings per page, which gives strong retrieval but large indexes and costly late-interaction scoring. The paper argues that aggressive compression should not pick the individually most salient patches; it should preserve query-relevant coverage, the diverse page regions that could become the best MaxSim match for some query. The authors introduce MarginMerge, which selects coverage-aware anchors, clusters the original patch embeddings, and trains a small shared network to synthesize one representative vector per cluster at index time, leaving the online MaxSim scorer unchanged. Across six datasets and two backbones, MarginMerge reports the highest matched query-agnostic average nDCG@5 at 5% and 10% vector retention, preserving 97–99% of the full-index average while storing only 5–10% of the vectors, and it transfers to unseen datasets and retention ratios without retraining. If correct, the work reframes index compression as a set-coverage problem and explains why dense rendered pages compress more easily than natural images.

What carries the argument

The load-bearing object is the coverage-aware anchor selection objective $F_d(A) = \sum_{t=1}^M \omega_t \max_{a \in A} C_{ta}$, where $C_{ti} = \exp(-(b_t(d) - z_t^\top v_i)/\tau_c)$ is the coverage that patch $i$ provides for query prototype $z_t$, and $b_t(d) = \max_i z_t^\top v_i$ is the prototype's best match in the document. Since each $C_{ta}$ is monotone submodular in the anchor set, greedy anchor selection carries the standard $(1-1/e)$ approximation guarantee. The anchors define a partition of the document patches; a small shared network $g_\theta$ produces per-patch log-weights that combine anchor similarity, prototype relevance, and a learned residual, and the cluster representative is $r_c = \mathrm{Norm}(\sum_{i \in C_c} \alpha_i v_i)$, a normalized convex combination of original patch vectors. A ranking-margin distillation loss then matches the compressed margins $m^\theta_j = \hat S(q,d^+) - \hat S(q,d^-_j)$ to the full-index margins $m^{\mathrm{full}}_j$, weighted toward pairs near the ranking boundary, so the compression preserves document ordering rather than absolute scores. This machinery is what lets the method stay inside the standard MaxSim interface while changing the representation.

What would settle it

Build a target corpus whose evaluation queries are shifted so that the winning document patches lie in regions never activated by any training-query token, then run MarginMerge at 5% vector retention; if its nDCG@5 falls to or below uniform random retention (the baseline it currently beats), the prototype-bank premise is falsified.

Watch

Extended reading notes

Core claim

At its center, the paper claims that compressing a multi-vector document representation is a set-coverage problem rather than a patch-ranking problem. Because MaxSim gives each query token the score of its best-matching document patch, the evidence that must survive compression is the union, over the query distribution, of document regions that can serve as the strongest match; keeping several near-duplicate high-scoring patches is less valuable than representing complementary regions. The paper formalizes this with a prototype bank of query directions, a per-patch coverage score, and a monotone submodular anchor objective that greedy selection optimizes to the $(1-1/e)$ approximation guarantee. MarginMerge clusters the original patches by coverage-aware anchors, then synthesizes one representative per cluster as a normalized convex combination of its patches using a 1,057-parameter shared network, trained to match full-index ranking margins. The reported result is that on six datasets over both ColQwen2.5 and ColPali, this preserves 97–99% of mean nDCG@5 while storing 5–10% of the document vectors, with transfer to held-out datasets and retention ratios without retraining.

Load-bearing premise

The method assumes that the 128 query-prototype directions built from training-query tokens are representative of the query directions the index will actually face; if test queries activate document regions far outside those directions, the coverage scores, anchors, and learned weights lose their grounding.

Editorial extensions

If this is right

  • At 5–10% vector retention, a frozen ColQwen2.5 or ColPali index retains 97–99% of average nDCG@5, so large visual-document corpora can be served at a fraction of storage and MaxSim cost.
  • At 5% retention, MarginMerge reduces ranking flips relative to geometric merging by about 41% on average across the six ColQwen2.5 datasets, making the compressed ranking more faithful to the full index.
  • A single network trained at 5% retention transfers to 10% and 20% retention and to datasets held out from training (DocVQA, InfoVQA, TAT-DQA) without retraining.
  • Patch selection methods based on predicted importance or salience can underperform uniform random retention, while diversity-aware selection preserves retrieval quality, supporting the coverage-over-salience account.
  • Compressibility orders as dense rendered documents > scene-text photographs > natural photographs, matching the effective-rank redundancy statistic across nine datasets.

Reading between the lines

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

  • Editorial inference: the coverage principle should apply to any late-interaction retriever, including ColBERT-style text retrieval, where the same MaxSim account predicts that merging clusters of substitutable token embeddings by query direction would beat salience pruning.
  • Editorial inference: re-estimating the prototype bank on a small sample of target-distribution queries, or updating it at index refresh time, could make the method robust to domain shift without retraining the weighting network.
  • Editorial inference: the indexing bottleneck is greedy anchor selection, about 1.7 seconds per TAT-DQA document in the reported implementation; submodularity guarantees suggest sampling or lazy evaluation could scale this to billions of pages.
  • Editorial inference: because representatives are already normalized and cluster-local, quantizing them to bytes should compound with MarginMerge additively; the paper defers byte-level evaluation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes MarginMerge, an index-time compression method for frozen ColPali/ColQwen multi-vector retrievers. It selects coverage-aware anchors from a prototype bank of query directions, clusters document patch vectors, and learns a lightweight shared network to synthesize one representative per cluster, keeping the standard MaxSim retrieval interface. Experiments on six datasets and two backbones report nDCG@5 at 5% and 10% vector retention, claiming 97-99% preservation of full-index quality while reducing stored vectors by 90-95%, plus transfer to unseen datasets and retention ratios. The supplementary material contains per-seed results, ablations, a leave-one-dataset-out study, proofs, and a code archive.

Significance. If the empirical claims hold, this is a practically useful and well-scoped contribution: compression happens once at index time, retrieval uses the unmodified MaxSim engine, and the reported numbers are internally consistent. The paper is strengthened by per-seed reporting, a documented leakage audit, a fixed configuration across datasets, and a clean conditional proof of the submodularity result in Supplement C.1. However, the zero-shot transfer evidence is weakened by the construction of the prototype bank, and the flagship coverage-aware anchor component is not the best variant in the full system. These issues bear directly on the paper's main contributions and need to be addressed before publication.

major comments (2)
  1. [Appendix A.1 and Experimental Setup (with Eqs. 4-5)] The paper describes DocVQA, InfoVQA, and TAT-DQA as zero-shot/unseen and states that 'The zero shot datasets are excluded from all fitting stages.' Appendix A.1, however, states that 'The global prototype bank uses training side query tokens from all six retrieval datasets.' The prototype bank Z enters the coverage score C_ti in Eq. (4) and the anchor objective F_d(A) in Eq. (5), and it also defines the prototype-relevance features consumed by the learned network. The compressed representations of documents in the three 'unseen' datasets are therefore constructed with query-direction information drawn from those same datasets (their training-side splits). This makes the evaluation transductive rather than zero-shot and could inflate the reported DocVQA, InfoVQA, and TAT-DQA numbers. The leave-one-dataset-out study in Table S2 does remove the target from prototype construction, but it covers only ArxivQA, TabFQuad, and Flickr, and on Flickr MarginMerge remains below geometric merging (0.6840 vs 0.7536). Please rerun the zero-shot evaluation with a prototype bank built only from the three training datasets, or restrict the claim to 'unseen by the weighting network' and clearly label the current protocol as transductive.
  2. [Table 4] In the full configuration, k-center anchors with the learned full objective achieve a higher average nDCG@5 than the reported coverage-aware anchors: 0.875 vs 0.865 at 5% retention and 0.886 vs 0.883 at 10% retention on ColQwen2.5. Since the reported MarginMerge configuration uses coverage-aware anchors, the paper's headline numbers are not the best configuration of its own method. More importantly, the claim that coverage-aware anchor selection is a core component of the final method is not supported by this table: the benefit of coverage-aware anchors appears only when the representative is fixed (e.g., 'Retained anchor' and 'Response centroid' rows), while learned synthesis makes the anchor rule largely irrelevant or slightly harmful. The paper should either report the k-center variant as the main configuration, with a clear explanation of why the precommitted coverage-aware anchor rule was retained, or substantially revise the claim that 'coverage-aware anchors' are responsible for the method's success.
minor comments (5)
  1. [Table S1] The flip-rate columns are not labeled with the method to which each number belongs; the pairs such as .0033/.0048 and .0621/.1095 are ambiguous as printed. Please add explicit column headers such as 'MarginMerge flip rate' and 'Geometric merging flip rate'.
  2. [Experimental Setup] The symbol '≈' is used in Table 1 without a definition in the table caption; the definition appears only in the Experimental Setup text. A brief note in the caption would improve readability.
  3. [Abstract and Contributions] The phrase 'The same model transfers to unseen datasets' should be qualified in the abstract, because the prototype bank uses training-side query tokens from all six datasets and only the weighting network is unseen by the three evaluation datasets.
  4. [Appendix A.4] The two spatial coordinates are set to zero in all reported runs, so the network effectively does not use spatial features. This should be stated in the main text as well, since the feature description in the main method implies spatial information is available.
  5. [Stability and Transfer] The claim that ranking flips are reduced 'by approximately 41% on average' does not obviously follow from the per-dataset flip rates in Table S1, which give an average reduction of about 36% if the first value in each pair is MarginMerge and the second is geometric merging. Please clarify the aggregation or the interpretation of the pair ordering.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; derivation is self-contained, though the disclosed prototype-bank construction makes the zero-shot evaluation transductive rather than strictly unseen.

full rationale

The claimed derivation is not circular. MarginMerge's coverage objective (Eqs. 4-5) is defined on a prototype bank Z, which is built from training-side query tokens (Appendix A.4); evaluation queries are excluded from prototype construction, hard negative mining, feature normalization, training, validation, and checkpoint selection (A.1). The margin-distillation loss (Eq. 12) fits compressed margins to full-index margins, but this is a training objective, not a prediction: the reported nDCG@5 is measured on held-out queries and documents. The sufficient-condition proof in Supplement C.2 is a conditional bound on sign preservation, not a fitted constraint. The leave-one-dataset-out study (Table S2) removes the target from all construction stages and still shows MarginMerge improving over the fixed response centroid on each held-out target, with the caveat that it remains below geometric merging on Flickr. The one substantive concern is that A.1 discloses "The global prototype bank uses training side query tokens from all six retrieval datasets," meaning DocVQA, InfoVQA, and TAT-DQA contribute training-side query tokens to the prototype bank even though they are called zero-shot because the weighting network is not trained on them. This makes the evaluation transductive for those datasets and could inflate the transfer numbers, but it is a data-construction and validity issue rather than a circularity: the predicted nDCG@5 is not defined in terms of itself, the weighting network is not optimized on those datasets, and the evaluation split is document-disjoint. No step in the derivation reduces to its own inputs, so no circular step is identified.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities beyond its own algorithmic constructs. The free parameters are standard hyperparameters, mostly fixed a priori but selected on validation. The axioms are either standard math or domain assumptions about the frozen retriever and query distribution.

free parameters (10)
  • M = 128
    Number of query prototypes in the bank, chosen by hand (Appendix A.4).
  • tau_c = 0.05
    Coverage temperature controlling how quickly coverage decays with response gap (Eq. 4).
  • tau_a = 0.1
    Anchor-similarity temperature in the log-weight (Eq. 7).
  • p = 0.5
    Exponent scaling prototype relevance in Eq. 7.
  • H = 5
    Clipping bound for the learned residual output (Appendix A.4).
  • Tm = 2
    Temperature for boundary weighting in margin loss (Eq. 12).
  • delta = 0.5
    Huber threshold in Eq. 12.
  • loss coefficients lambda_m, lambda_r, lambda_l, lambda_w, lambda_e, lambda_a = 1, 0.5, 0.5, 1e-3, 0.01, 0.01
    Weights of the six loss terms in Eq. S1, fixed across runs but selected during development.
  • number of hard negatives = 8
    Eight highest-scoring non-relevant training documents used in the margin loss.
  • network parameters of g_theta = 1,057 trainable parameters
    Learned weighting network, trained at 5% retention on ArxivQA, TabFQuad, and Flickr.
assumptions (6)
  • domain assumption MaxSim late interaction is the correct scoring function for retrieval
    The paper defines compression as preserving MaxSim scores (Eqs. 1-2), assuming the frozen retriever's interface is fixed.
  • domain assumption Training-query token distribution is representative of the test query distribution
    The prototype bank is built from training-side query tokens (Appendix A.1); transfer relies on this.
  • standard math Submodular greedy selection gives a valid approximation guarantee for the coverage objective
    Proved in Supplement C.1: each f_t is monotone submodular, so F_d is monotone submodular.
  • standard math Preserving the sign of positive-negative margins preserves document ordering
    Sufficient condition proved in Supplement C.2 via |Delta_j| < |m_full_j|.
  • domain assumption Frozen patch embeddings contain enough redundancy that a 5-10% vector budget can represent query-relevant evidence
    This is the premise of all compression; the paper supports it empirically but cannot prove it for arbitrary backbones or content.
  • ad hoc to paper Huber margin distillation with boundary weighting is a valid proxy for preserving nDCG@5
    The objective is heuristic; the paper notes margin matching is not uniformly superior to score reconstruction (Section on Margin matching).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Coverage Matters: MarginMerge for Compressing Multi-Vector Visual Document Retrievers." pith.science (2026). https://pith.science/paper/TIZWIJB5

@misc{pith2026260802969,
  author       = {Pith},
  title        = {Pith review of: Coverage Matters: MarginMerge for Compressing Multi-Vector Visual Document Retrievers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TIZWIJB5}},
  note         = {Machine review of arXiv:2608.02969}
}
read the original abstract

Multi-vector visual document retrievers such as ColPali and ColQwen achieve strong retrieval by storing fine-grained patch embeddings, but this produces large indexes and costly late-interaction scoring. We argue that effective compression should preserve query-relevant coverage, meaning the diverse document regions that may become the strongest MaxSim match across queries, rather than selecting patches independently by salience. This view also explains why dense rendered pages are easier to compress than natural images. We introduce MarginMerge, a compression method for frozen multi-vector retrievers. It selects coverage-aware anchors, clusters document patches, and uses a lightweight shared network to synthesize one representative per cluster. Compression is performed once during indexing, while retrieval keeps the standard MaxSim interface. Across six datasets on both ColQwen2.5 and ColPali, MarginMerge achieves the highest matched query-agnostic average at 5% and 10% vector retention. Compared with the uncompressed index using the same backbone, it preserves between 97% and 99% of average nDCG@5 while reducing stored document vectors by between 90% and 95%. At 5% retention, it also reduces ranking flips relative to geometric merging on all six ColQwen2.5 datasets by approximately 41% on average. The same model transfers to unseen datasets and retention ratios without retraining.

Figures

Figures reproduced from arXiv: 2608.02969 by the authors.

Figure 1
Figure 1. Motivation and overview of MarginMerge. MarginMerge compresses multi-vector representations into a compact set of query-relevant representatives while preserving the standard MaxSim retrieval pipeline. Abstract Multi vector visual document retrievers such as ColPali and ColQwen achieve strong retrieval by storing fine grained patch embeddings, but this produces large indexes and costly late in￾teraction scoring. We … view at source ↗
Figure 2
Figure 2. Overview of MarginMerge. The offline pipeline selects coverage-aware anchors and learns one representative per cluster. Ranking-margin distillation preserves full-index ordering, while retrieval uses standard MaxSim over only the compressed representatives. covering the same direction. In particular, once one selected anchor already provides high coverage for a prototype, se￾lecting another similar anchor gives litt… view at source ↗
Figure 3
Figure 3. Retrieval under stronger compression. Learned synthesis remains robust as retention falls, especially on Flickr. Values are reported in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Compressibility depends on content. Random re￾tention is effective for dense rendered pages but degrades on Flickr. Redundancy multiplicity separates the dataset groups at the corpus level [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Why query agnostic pruning fails. Learned selectors can underperform random retention on unseen queries. The gap grows with the candidate pool, and diversity preserves retrieval better than selecting patches only by predicted likelihood. unseen queries. Learned selecto…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 21 canonical work pages

  1. [1]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , month = nov, year =

    Unifying Multimodal Retrieval via Document Screenshot Embedding , author =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , month = nov, year =. doi:10.18653/v1/2024.emnlp-main.373 , url =

  2. [2]

    The Thirteenth International Conference on Learning Representations , year =

    Faysse, Manuel and Sibille, Hugues and Wu, Tony and Omrani, Bilel and Viaud, Gautier and Hudelot, C. The Thirteenth International Conference on Learning Representations , year =

  3. [3]

    2025 , doi =

    Bai, Shuai and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Song, Sibo and Dang, Kai and Wang, Peng and Wang, Shijie and Tang, Jun and Zhong, Humen and Zhu, Yuanzhi and Yang, Mingkun and Li, Zhaohai and Wan, Jianqiang and Wang, Pengfei and Ding, Wei and Fu, Zheren and Xu, Yiheng and Ye, Jiabo and Zhang, Xi and Xie, Tianbao and Cheng, Z...

  4. [4]

    arXiv preprint arXiv:2505.17166 , year =

    Mac. arXiv preprint arXiv:2505.17166 , year =

  5. [5]

    2020 , pages =

    Khattab, Omar and Zaharia, Matei , booktitle =. 2020 , pages =. doi:10.1145/3397271.3401075 , url =

  6. [6]

    2022 , address =

    Santhanam, Keshav and Khattab, Omar and Saad-Falcon, Jon and Potts, Christopher and Zaharia, Matei , booktitle =. 2022 , address =. doi:10.18653/v1/2022.naacl-main.272 , url =

  7. [7]

    2022 , pages =

    Santhanam, Keshav and Khattab, Omar and Potts, Christopher and Zaharia, Matei , booktitle =. 2022 , pages =. doi:10.1145/3511808.3557325 , url =

  8. [8]

    Token Merging: Your

    Bolya, Daniel and Fu, Cheng-Yang and Dai, Xiaoliang and Zhang, Peizhao and Feichtenhofer, Christoph and Hoffman, Judy , booktitle =. Token Merging: Your. 2023 , url =

Show all 30 references
  1. [9]

    Findings of the Association for Computational Linguistics: ACL 2025 , month = jul, year =

    Towards Storage-Efficient Visual Document Retrieval: An Empirical Study on Reducing Patch-Level Embeddings , author =. Findings of the Association for Computational Linguistics: ACL 2025 , month = jul, year =. doi:10.18653/v1/2025.findings-acl.1003 , url =

  2. [10]

    Hierarchical Patch Compression for

    Duong, Bach and Minh, Pham Nhat , booktitle =. Hierarchical Patch Compression for. 2025 , publisher =. doi:10.5220/0013732500004000 , url =

  3. [11]

    2026 , url =

    Xiao, Zilin and Ma, Qi and Gu, Mengting and Chen, Chun-cheng Jason and Chen, Xintao and Ordonez, Vicente and Mohan, Vijai , booktitle =. 2026 , url =

  4. [12]

    2025 , url =

    Yan, Yibo and Xu, Guangwei and Zou, Xin and Liu, Shuliang and Kwok, James and Hu, Xuming , journal =. 2025 , url =

  5. [13]

    Findings of the Association for Computational Linguistics: ACL 2026 , month = jul, year =

    Sculpting the Vector Space: Towards Efficient Multi-Vector Visual Document Retrieval via Prune-then-Merge Framework , author =. Findings of the Association for Computational Linguistics: ACL 2026 , month = jul, year =. doi:10.18653/v1/2026.findings-acl.1247 , url =

  6. [14]

    arXiv preprint arXiv:2601.20107 , year =

    Structural Anchor Pruning: Training-Free Multi-Vector Compression for Visual Document Retrieval , author =. arXiv preprint arXiv:2601.20107 , year =. doi:10.48550/arXiv.2601.20107 , url =

  7. [15]

    arXiv preprint arXiv:2602.21202 , year =

    Multi-Vector Index Compression in Any Modality , author =. arXiv preprint arXiv:2602.21202 , year =

  8. [16]

    2024 , doi =

    Dhulipala, Laxman and Hadian, Majid and Jayaram, Rajesh and Lee, Jason and Mirrokni, Vahab , booktitle =. 2024 , doi =

  9. [17]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Unveil: Unified Visual-Textual Integration and Distillation for Multi-modal Document Retrieval , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2025 , publisher =

  10. [18]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Towards Natural Language-Based Document Image Retrieval: New Dataset and Benchmark , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  11. [19]

    Tanaka, Ryota and Iki, Taichi and Hasegawa, Taku and Nishida, Kyosuke and Saito, Kuniko and Suzuki, Jun , booktitle =

  12. [20]

    Advances in Neural Information Processing Systems , volume =

    Fine-grained Late-interaction Multi-modal Retrieval for Retrieval Augmented Visual Question Answering , author =. Advances in Neural Information Processing Systems , volume =

  13. [21]

    Advances in Neural Information Processing Systems , volume =

    Rethinking the Role of Token Retrieval in Multi-Vector Retrieval , author =. Advances in Neural Information Processing Systems , volume =

  14. [22]

    and Okuno, Tomoyuki and Nakata, Yohei and Keutzer, Kurt and Zhang, Shanghang , booktitle =

    Zhang, Yuan and Fan, Chun-Kai and Ma, Junpeng and Zheng, Wenzhao and Huang, Tao and Cheng, Kuan and Gudovskiy, Denis A. and Okuno, Tomoyuki and Nakata, Yohei and Keutzer, Kurt and Zhang, Shanghang , booktitle =. 2025 , publisher =

  15. [23]

    2025 , doi =

    Zhang, Jiaxin and Yang, Wentao and Lai, Songxuan and Xie, Zecheng and Jin, Lianwen , booktitle =. 2025 , doi =

  16. [24]

    Multimodal

    Li, Lei and Wang, Yuqi and Xu, Runxin and Wang, Peiyi and Feng, Xiachong and Kong, Lingpeng and Liu, Qi , booktitle =. Multimodal. 2024 , publisher =

  17. [25]

    Mathew, Minesh and Karatzas, Dimosthenis and Jawahar, C. V. , booktitle =

  18. [26]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages =

    Mathew, Minesh and Bagal, Viraj and Tito, Rub. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages =. 2022 , doi =

  19. [27]

    Proceedings of the 30th ACM International Conference on Multimedia , pages =

    Towards Complex Document Understanding by Discrete Reasoning , author =. Proceedings of the 30th ACM International Conference on Multimedia , pages =. 2022 , doi =

  20. [28]

    Lawrence and Doll

    Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Bourdev, Lubomir and Girshick, Ross and Hays, James and Perona, Pietro and Ramanan, Deva and Zitnick, C. Lawrence and Doll. Microsoft. Computer Vision -- ECCV 2014 , pages =. 2014 , publisher =

  21. [29]

    Transactions of the Association for Computational Linguistics , volume =

    From Image Descriptions to Visual Denotations: New Similarity Metrics for Semantic Inference over Event Descriptions , author =. Transactions of the Association for Computational Linguistics , volume =. 2014 , doi =

  22. [30]

    2020 , publisher =

    Sidorov, Oleksii and Hu, Ronghang and Rohrbach, Marcus and Singh, Amanpreet , booktitle =. 2020 , publisher =

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.