Can Argus Judge Them All? Comparing VLMs Across Domains

Ebad Shabbir; Gautam Siddharth Kashyap; Harsh Joshi; Jiechao Gao; Niharika Jain; Rafiq Ali; Sarthak Jain; Usman Naseem

REVIEW 4 major objections 6 minor 20 references

Can Argus Judge Them All? Comparing VLMs Across Domains

T0 review · 4 major / 6 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 This paper argues that model-selection rankings should incorporate cross-dataset consistency, and it introduces a CDC metric under which CLIP ranks highest.

desk verdict The abstract's five-model claims don't appear in the body; the three-model benchmark is usable but the CDC theory is wrong. read the letter →

arxiv 2507.01042 v2 pith:UZ4OHAFE submitted 2025-06-23 cs.IR cs.AIcs.CL

Harsh Joshi , Gautam Siddharth Kashyap , Rafiq Ali , Ebad Shabbir , Niharika Jain , Sarthak Jain , Jiechao Gao , Usman Naseem This is my paper

classification cs.IRcs.AIcs.CL

keywords vision-languagemodelscross-datasetconsistencybenchmarkevaluationimage-textretrievalvisualquestionansweringcaptiongenerationmodelefficiency

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 argues that selecting a vision-language model by benchmark task scores alone ignores a capability-reliability gap: models with similar scores can behave very differently across datasets. It proposes ARGUS-EVAL, a framework that evaluates models on four axes—benchmark capability, cross-dataset consistency, robustness retention, and efficiency—and introduces a new consistency metric, CDC, to quantify generalization. On CLIP, BLIP, and LXMERT across retrieval, captioning, and reasoning benchmarks, the reliability ranking differs from the task ranking: CLIP has the highest consistency (0.92) and the lowest latency and memory, BLIP leads retrieval and captioning, and LXMERT leads structured reasoning. A sympathetic reading is that these tradeoffs should guide model selection and motivate hybrid architectures.

What carries the argument

The carrying object is the ARGUS-EVAL framework and its new metric CDC. CDC is computed as $\text{CDC}(M_j)=1-\frac{1}{|D|}\sum_{i=1}^{|D|}\left|\frac{a_{i,j}-\bar{a}_j}{\bar{a}_j}\right|$, where $a_{i,j}$ is model $j$'s accuracy on dataset $i$ and $\bar{a}_j$ is its average accuracy across datasets; the value is high when a model's accuracy is uniform across datasets. The paper supplements this definition with four lemmas—boundedness, perfect consistency, invariance to uniform scaling, and sensitivity to outliers—to argue that the score is interpretable. This metric is load-bearing because it is what converts the raw benchmark numbers into a reliability ranking that differs from the capability ranking.

What would settle it

Recompute the CDC table after normalizing each dataset's accuracy distribution, for example z-scoring the three models' accuracies within each dataset. If the ordering CLIP above BLIP above LXMERT is not preserved, then the reported consistency ranking depends on the raw-accuracy comparability assumption rather than on an intrinsic property of the models.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that vision-language models cannot be judged by a single task benchmark. The authors define Cross-Dataset Consistency (CDC) as one minus the mean absolute relative deviation of a model's per-dataset accuracy from its own average accuracy, and they report that on the five selected datasets CLIP scores 0.92, BLIP 0.76, and LXMERT 0.64. They also report that CLIP is the most efficient model (roughly 28–35 ms latency and 0.8–1.1 GB memory), BLIP dominates COCO and Flickr30k retrieval and captioning (for example, R@1 of 78.9% and 76.5%, CIDEr of 134.5 and 95.2), and LXMERT dominates CLEVR accuracy (96.3%) and VCR QA→R (71.5%). The paper's own inference is that generalization, specialization, and efficiency are distinct axes that must be traded off in deployment.

Load-bearing premise

The load-bearing premise is that the retrieval, captioning, and reasoning accuracy numbers fed into the CDC formula are comparable quantities, so averaging their relative deviations is meaningful; if they are not on a common scale, the CDC rankings are not interpretable.

Editorial extensions

If this is right

Model-selection workflows should report reliability alongside task accuracy, since task leaders BLIP and LXMERT are not the most consistent model, CLIP.
For resource-constrained or real-time settings, the efficiency measurements favor contrastive architectures: CLIP runs at roughly 30 ms and under 1.1 GB, while BLIP and LXMERT require several gigabytes and 115–180 ms.
The scale-invariance property of CDC implies that uniform accuracy compression, such as common quantization, does not change the consistency ranking, so CDC can be used to compare compressed variants.
The task-specific winners suggest that a hybrid architecture combining contrastive alignment, generative pretraining, and cross-modal reasoning would be the natural next step.
Any single-number ranking, including CDC, should be presented together with capability and efficiency scores rather than as a standalone leaderboard.

Reading between the lines

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

The abstract's headline comparison includes Qwen-2.5VL-3B-Instruct and Gemma, but the reported experiments cover only CLIP, BLIP, and LXMERT; applying the stated ARGUS-EVAL protocol to the larger models is a direct test of whether the abstract's Qwen result survives.
CDC as defined compares raw accuracies across retrieval, captioning, and reasoning datasets; if those numbers are normalized differently, for example z-scored within each dataset before averaging, the ranking of CLIP, BLIP, and LXMERT may change. The 0.92/0.76/0.64 values should therefore be read as relative to the authors' choice of raw accuracy.
The named robustness-retention axis is not operationalized in the results, so the full four-axis framework is not yet testable as a package.
A useful extension would be to report CDC with confidence intervals across random seeds, since the paper repeats inference but reports no variance for the CDC values.

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

Quick take: the abstract and the body disagree about what was evaluated; the headline Qwen/Gemma numbers appear nowhere. Don't trust the abstract.

What's actually here: a straightforward three-model benchmark (CLIP, BLIP, LXMERT) across COCO, Flickr30k, CLEVR, VCR, Visual Genome, with retrieval, captioning, VQA, and efficiency tables. That part is usable reference data. The tables are reasonably detailed, and the efficiency numbers (latency, memory, FLOPs) are the kind of thing practitioners want. The CDC metric is a standard dispersion measure (1 minus mean absolute relative deviation) and is fine as a descriptive statistic, but calling it novel is a stretch, and the paper does not compare it with existing reliability metrics.

The soft spots are serious. First, the abstract claims a five-model ARGUS-EVAL framework and gives Qwen-2.5VL-3B results; the body contains no Qwen or Gemma experiments, no table, no algorithm, no mention beyond the abstract. That is an internal consistency defect, and it makes the headline ranking unverifiable. Second, Lemma 4.1 (boundedness) is false. For three datasets with accuracies 0, 0, 1, the mean is 1/3 and the CDC equals 1 - (1+1+2)/3 = -1/3. So CDC is not bounded below by 0. The proof is hand-wavy and does not do the work. Third, Section 4.1 promises mean and standard deviation over random seeds, but no error bars appear anywhere. Fourth, the CDC mixes accuracies from retrieval, captioning, and reasoning tasks that are not on a common scale; the paper itself notes it switched retrieval direction for Visual Genome, so averaging relative deviations across those numbers is not obviously meaningful. Fifth, code and data are not released.

The limitations section is honest about scope, which counts in the paper's favor. The citation list is thin in places (e.g., 'Krishna and et al.' is incomplete), but that's minor.

Bottom line: the three-model benchmark on its own is a minor but citable reference point; the CDC metric is not new and its theory is wrong as stated. The abstract mismatch is the reason to send this back. I would not cite it, and I would not spend referee time on it until the authors fix the consistency and the lemma. A serious editor should probably desk reject on the grounds that the advertised results are missing; if the authors resubmit with only the three-model comparison and a corrected metric, it could be a fine short paper.

Referee Report

4 major / 6 minor

Summary. The paper proposes ARGUS-EVAL, a benchmarking framework for vision-language models, and reports experiments on CLIP, BLIP, and LXMERT across retrieval (COCO, Flickr30k, Visual Genome), visual reasoning (CLEVR, VCR), and captioning (COCO, Flickr30k), together with efficiency metrics and a proposed Cross-Dataset Consistency (CDC) metric. The body claims that BLIP leads retrieval and captioning, LXMERT leads structured reasoning, CLIP has the highest CDC (0.92) and the best efficiency, and the conclusion summarizes only the three-model comparison. The abstract additionally claims evaluation of Gemma-3-4B and Qwen-2.5VL-3B-Instruct and reports Qwen as the strongest overall capability model; no such results appear anywhere in the body. The paper also states several lemmas for CDC and promises statistical reporting that the tables do not deliver.

Significance. If the claims were supported, a reliability-oriented evaluation metric such as CDC could be practically useful for model selection, and the paper's attempt to combine capability, consistency, robustness, and efficiency is a reasonable goal. The paper has several strengths: it reports a concrete experimental protocol, includes efficiency metrics, and attempts to provide theoretical lemmas for the new metric. However, the central claims do not survive scrutiny: the abstract's five-model results are absent from the body, the boundedness lemma for CDC is false, and the CDC values in Table 5 cannot be reproduced from the paper's own reported accuracy tables. As a result, the paper's central reliability-ranking contribution is not currently established.

major comments (4)

[Abstract vs. §3.2–§5] The abstract claims that ARGUS-EVAL evaluates five models and that Qwen-2.5VL-3B-Instruct achieves the strongest overall capability (R@1=82.7%, BLEU-4=47.2%, CIDEr=141.6, CDC=0.91), while CLIP has the lowest latency and memory footprint. The body contains no Qwen-2.5VL or Gemma-3-4B results: Algorithm 1 fixes the model set as M={M1, M2, M3}, §3.2 introduces only CLIP, BLIP, and LXMERT, and Tables 1–5 report only those three models. This is an internal inconsistency: either the five-model evaluation was not run, in which case the abstract's ranking is unsupported, or it was run and the required evidence is omitted. Either way, the headline claim cannot be checked from the manuscript.
[§4.1, Eq. (4), Lemma 4.1] The boundedness lemma is false. For four datasets with accuracies {1, 0, 0, 0}, the mean is 0.25, and the mean absolute relative deviation is (0.75 + 0.25 + 0.25 + 0.25) / (4 × 0.25) = 1.5, so CDC = 1 − 1.5 = −0.5, which is outside the claimed [0, 1] interval. Since boundedness is the stated basis for interpreting CDC as a standardized scale, this theoretical guarantee needs to be corrected.
[§5.5, Table 5 vs. §4.1, Eq. (4)] The CDC scores are not reproducible from the paper's own data. Using Eq. (4) with the five per-dataset numbers in Tables 1 and 2 (COCO R@1, Flickr30k R@1, Visual Genome R@1, CLEVR accuracy, VCR Q→A accuracy) gives approximately CLIP 0.92, LXMERT 0.89, and BLIP 0.93, whereas Table 5 reports CLIP 0.92, BLIP 0.76, and LXMERT 0.64. Moreover, the note in §4.1 switches retrieval direction between datasets (image-to-text for COCO/Flickr30k, text-to-image for Visual Genome), so the underlying accuracies come from heterogeneous tasks without any stated normalization or calibration; averaging their relative deviations is not interpretable as a cross-dataset consistency score.
[§4.1, §4.2, Algorithm 1] Section 4.1 promises that for accuracy-based metrics the paper reports the mean and standard deviation over multiple random seeds unless otherwise specified, and Algorithm 1 says inference is repeated across randomized splits. No standard deviations, seed counts, or split details appear in any table or in Section 5; all reported numbers are point estimates. For a framework whose stated purpose is reliability, this missing statistical support is load-bearing.

minor comments (6)

[Table 3] The BLIP COCO row reports BLEU-1 as 78.9%, which is identical to BLIP's COCO R@1 in Table 1; this coincidence is unexplained and suggests a copy/paste error.
[Table 1] The table heading uses MRR for COCO and Flickr30k but MedR for Visual Genome, and MRR is omitted for Visual Genome; the choice is not justified in the text.
[§4.1] The sentence 'For CLEVR, we report overall classification accuracy .' contains a missing value or incomplete placeholder.
[Abstract] The model name 'Qwen-2.5VL-3BInstruct' is missing a space or hyphen; it should be 'Qwen-2.5VL-3B-Instruct'.
[References] The reference 'Krishna and et al. (2020)' is malformed; the citation should list the authors properly.
[Limitations] The Limitations section states that model size and hardware cost are not analyzed, but Table 4 reports memory usage; the distinction between model size and runtime memory should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: CDC is defined directly by Eq. (4) and the lemmas are algebraic consequences; the reported rankings are empirical measurements, not predictions fitted to themselves.

full rationale

The paper's only formal derivation is the Cross-Dataset Consistency metric (Eq. 4), which is defined directly as 1 minus the mean absolute relative deviation of per-dataset accuracies. Lemmas 4.1-4.4 (boundedness, perfect consistency, uniform scaling invariance, outlier sensitivity) are immediate algebraic consequences of this definition; none of them is loaded with an external assumption, a fitted parameter, or a self-citation. The benchmark numbers in Tables 1-5 are reported measurements, not quantities predicted from fitted parameters. There are no uniqueness theorems imported from the authors' prior work, no ansatz smuggled in via citation, and no step in which an output is defined in terms of the claim it is supposed to support. The abstract/body mismatch (ARGUS-EVAL and the Qwen/Gemma results appear only in the abstract, while the body evaluates only CLIP, BLIP, and LXMERT) is a serious reporting/consistency defect and a correctness risk, but it is not circular reasoning. Accordingly, the circularity score is 0.

Assumptions & free parameters 0 free parameters · 5 assumptions · 2 invented entities

The central claims rest on the comparability of heterogeneous task accuracies, on the representativeness of standard benchmarks, on the fidelity of official checkpoints, and on the boundedness of the proposed CDC metric. The boundedness axiom is false, and the comparability assumption is not justified, so the resulting rankings lack a sound foundation.

assumptions (5)

domain assumption Accuracy values from different datasets and tasks are directly comparable when averaged into CDC.
Equation (4) in Section 4.1 averages relative deviations of per-dataset accuracies, but retrieval R@1, VQA accuracy, and different retrieval directions are not on a common scale; Table 5 treats them as comparable.
domain assumption The benchmark datasets and their standard splits represent the deployment conditions the conclusions target.
Section 3.1 and the Limitations paragraph both acknowledge standard benchmarks may not reflect real-world diversity; the conclusions still rely on them.
domain assumption Official pretrained checkpoints and the reported inference setup faithfully represent each model family.
Section 4.2 specifies pretrained weights from official repositories, A100 GPU, batch size 32, and FP16; no checkpoint hashes or code are provided to verify.
ad hoc to paper The maximal CDC deviation is bounded such that the score lies in [0,1].
Lemma 4.1 asserts 0<=CDC<=1, but the proof does not bound the relative deviation when the mean accuracy approaches zero; a concrete counterexample yields CDC = -0.6.
domain assumption Mean absolute relative deviation is an adequate measure of cross-dataset consistency.
The metric is defined in Eq. (4); no comparison to alternative reliability metrics is provided.

invented entities (2)

Cross-Dataset Consistency (CDC) score
purpose: Summarize a model's stability across datasets as one minus the mean absolute relative deviation from its average accuracy.
The paper defines CDC and proves (incorrectly) that it is bounded in [0,1]; no external validation link to deployment reliability or human judgments is given.
ARGUS-EVAL framework
purpose: Capability-reliability-oriented evaluation framework with components P(M), RR(M), CDC(M), and E(M).
Only CDC and efficiency are implemented in the body; P(M) and RR(M) are named in the abstract but never defined.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Can Argus Judge Them All? Comparing VLMs Across Domains." pith.science (2026). https://pith.science/paper/UZ4OHAFE

@misc{pith2026250701042,
  author       = {Pith},
  title        = {Pith review of: Can Argus Judge Them All? Comparing VLMs Across Domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UZ4OHAFE}},
  note         = {Machine review of arXiv:2507.01042}
}

read the original abstract

Vision-Language Models (VLMs) are increasingly used in industry VLM applications such as retrieval systems, content generation platforms, and decision-support workflows, where model selection is commonly guided by benchmark rankings. These rankings are largely determined by retrieval, captioning, and reasoning downstream tasks; however, models with similar task performance often show substantially different behavior across datasets. This creates a Capability-Reliability Gap between benchmark performance and observed model stability. We present ARGUS-EVAL, a capability-reliability-oriented evaluation framework for VLMs that characterizes model behavior through Benchmark Capability P(M), Cross-Dataset Consistency CDC(M), Robustness Retention RR(M), and Efficiency E(M). We evaluate CLIP, BLIP, LXMERT, Gemma-3-4B, and Qwen-2.5VL-3B-Instruct across retrieval, captioning, and reasoning downstream tasks. The results reveal notable differences between capability-oriented and reliability-oriented rankings. Qwen-2.5VL-3BInstruct achieves the strongest overall capability (R@1 = 82.7%, BLEU-4 = 47.2%, CIDEr = 141.6, CDC = 0.91), whereas CLIP records the lowest latency (31 ms) and memory footprint (0.9 GB).

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 11 canonical work pages

[1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Jean-Baptiste Alayrac, Andre Recasens, Joseph Marino, and et al. 2022. Flamingo: a visual language model for few-shot learning. In NeurIPS

work page 2022
[4]

Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. 2023. https://openreview.net/forum?id=vvoWPYqZJA Instruct BLIP : Towards general-purpose vision-language models with instruction tuning . In Thirty-seventh Conference on Neural Information Processing Systems

2023
[5]

Fartash Faghri, David J Fleet, Jamie Kiros, and Sanja Fidler. 2018. Vse++: Improving visual-semantic embeddings with hard negatives. In British Machine Vision Conference (BMVC)

work page 2018
[6]

Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

work page 2017
[7]

Maryam Hashemi, Ghazaleh Mahmoudi, Sara Kodeiri, Hadi Sheikhi, and Sauleh Eetemadi. 2023. Lxmert model compression for visual question answering. arXiv preprint arXiv:2310.15325

work page Pith review arXiv 2023
[8]

Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam

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. In arXiv preprint arXiv:1704.04861

arXiv 2017

Show all 20 references

[9]

Sahar Kazemzadeh, Shikhar Sharma, Priya Agrawal, Holger Schwenk, and Dhruv Batra. 2023. Vl-bench: A comprehensive benchmark for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

2023
[10]

Ranjay Krishna and et al. 2020. Evaluating generalization in vision-and-language navigation models. In Conference on Computer Vision and Pattern Recognition (CVPR)

2020
[11]

Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888--12900. PMLR

2022
[12]

Wei Li, Kaoru Ota, and Ming Dong. 2020. Edge ai: On-demand accelerating deep neural network inference via edge computing. IEEE Transactions on Wireless Communications, 19(1):447--457

2020
[13]

Chia Xin Liang, Pu Tian, Caitlyn Heqi Yin, Yao Yua, Wei An-Hou, Li Ming, Tianyang Wang, Ziqian Bi, and Ming Liu. 2024. A comprehensive survey and guide to multimodal large language models in vision-language tasks. arXiv preprint arXiv:2411.06284

2024
[14]

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
[15]

Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. 2016. Xnor-net: Imagenet classification using binary convolutional neural networks. In ECCV

2016
[16]

Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. In NeurIPS

2019
[17]

Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. 2024. Eva-clip-18b: Scaling clip to 18 billion parameters. arXiv preprint arXiv:2402.04252

2024 arXiv
[18]

Hao Tan and Mohit Bansal. 2019. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490

2019 arXiv
[19]

Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. 2021. Beir: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Proceedings of the 2021 Conference on Neural Information Processing Systems (NeurIPS)

2021
[20]

Xiaohui Xu, Zhi Cai, Tian Chen, and Chuan Xu. 2018. Workload prediction in edge computing: A survey. In IEEE Communications Surveys & Tutorials

2018

Pith tools

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

[1] [1]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Jean-Baptiste Alayrac, Andre Recasens, Joseph Marino, and et al. 2022. Flamingo: a visual language model for few-shot learning. In NeurIPS

work page 2022

[4] [4]

Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. 2023. https://openreview.net/forum?id=vvoWPYqZJA Instruct BLIP : Towards general-purpose vision-language models with instruction tuning . In Thirty-seventh Conference on Neural Information Processing Systems

2023

[5] [5]

Fartash Faghri, David J Fleet, Jamie Kiros, and Sanja Fidler. 2018. Vse++: Improving visual-semantic embeddings with hard negatives. In British Machine Vision Conference (BMVC)

work page 2018

[6] [6]

Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

work page 2017

[7] [7]

Maryam Hashemi, Ghazaleh Mahmoudi, Sara Kodeiri, Hadi Sheikhi, and Sauleh Eetemadi. 2023. Lxmert model compression for visual question answering. arXiv preprint arXiv:2310.15325

work page Pith review arXiv 2023

[8] [8]

Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam

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. In arXiv preprint arXiv:1704.04861

arXiv 2017