Pith. sign in

REVIEW 5 major objections 4 minor 32 references

Visual-Linguistic Agent: Towards Collaborative Contextual Object Reasoning

T0 review · 5 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read VLA wraps any object detector with an MLLM referee and a CLIP re-labeler, reporting COCO AP50:95 gains up to 2.6 points with no detector retraining.

desk verdict A plausible test-time wrapper for object detection whose headline AP gains are not yet substantiated, because the evaluation never reports the flag-level accuracy, the CLIP-versus-detector comparison, or the score bookkeeping needed to compute AP. read the letter →

arxiv 2411.10252 v1 pith:U7SI3UBN submitted 2024-11-15 cs.CV

Jingru Yang , Huan Yu , Yang Jingxin , Chentianye Xu , Yin Biao , Yu Sun , Shengfeng He This is my paper · ORCID
classification cs.CV
keywords visual-linguisticagentmultimodallargelanguagemodelobjectdetectioncontextualreasoningerrorcorrectionCLIPclassificationCOCObenchmarkcollaboration
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 sets out to show that the localization strength of a conventional object detector and the world-knowledge of a multimodal large language model can be combined without retraining either one. In VLA, the detector proposes boxes and labels; the MLLM, reading the image and the detections as text, flags boxes that violate scene context; and a CLIP model re-labels the flagged crops. On COCO, the authors report consistent AP50:95 gains for Faster R-CNN, YOLOX, YOLO11, DETR, and DINO, from +1.3 to +2.6 points, with the largest gains on small objects. If the claim holds, contextual reasoning can be bolted onto detection as a post-processing wrapper rather than learned inside the network.

What carries the argument

The active ingredient is the text bridge between agents: detection outputs are formatted as prompts with class names and box coordinates, the MLLM evaluates them against its own caption and commonsense, and only flagged boxes are sent to CLIP. This flag-and-relabel loop is what the paper argues converts the MLLM's low-entropy global distribution $H(Y,R)$ into measured AP gains over the detector's IoU-weighted distribution $H_w(Y)$, with the gap expressed as information gain $IG = H_w(Y) - H(Y,R)$. CLIP is the corrective half of the loop; without it, the correction rate drops from 75.0% to 44.9%.

What would settle it

On COCO val2017 with the DINO+VLA (GPT-4o) configuration, record the ground-truth label inside every flagged box: if a large share of flagged boxes were already correctly labeled, or if CLIP's top-1 accuracy on flagged crops is not higher than DINO's own confidence-based classification on those same crops, then the reported AP gain would not survive the flag-and-relabel loop.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that detection errors are largely label errors that an MLLM can identify from global scene context and a CLIP classifier can correct locally. VLA operates in three stages: the Visual Agent detects, the Linguistic Agent generates a caption and evaluates the serialized detections for plausibility, and the Classification Vision Agent reclassifies each flagged region. The paper reports that this pipeline raises DINO's AP50:95 from 49.5 to 52.1 with GPT-4o, improves APs by up to 4.6 points for Claude, and raises the error-correction rate for DINO's label mistakes from 44.9% with the Linguistic Agent alone to 75.0% with the full VLA. The quantitative claim is that these gains are consistent across five detectors and four MLLMs, establishing a new benchmark claim for contextually coherent detection.

Load-bearing premise

The pipeline's entire measured gain rests on the MLLM flagging exactly the mislabeled boxes and no correct ones, and on CLIP classifying those flagged crops more accurately than the detector's own classification head; neither premise is directly measured.

Editorial extensions

If this is right

  • VLA gives a detector-agnostic upgrade path: any pretrained detector's outputs can be improved by the same text-prompt and relabel wrapper, with no gradient updates.
  • The reported gains concentrate on small objects and on confusable categories, so VLA could serve as a correctness layer for detection pipelines where label noise is costly.
  • The choice of MLLM matters: GPT-4o and LLaVA lead, while Claude and Gemini trail, suggesting the gain scales with the MLLM's visual reasoning quality.
  • With the Linguistic Agent alone, 44.9% of DINO label errors are corrected; adding CLIP raises this to 75.0%, so the classification agent is essential to the measured improvement.

Reading between the lines

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

  • A direct extension the paper leaves implicit: the same flag-and-relabel loop could be pointed at open-vocabulary or long-tail detectors, where the label noise that CLIP can fix is larger than on COCO's 80 classes.
  • The paper does not report the cost or latency of the extra MLLM call per image; a practical deployment would need flag-rate and API-cost per AP point, which is a natural follow-up measurement.
  • A stress test suggested by the architecture: feed VLA only the detector's own high-confidence mistakes and ask whether the MLLM still flags them; if flagging depends on obvious context, gains may shrink on cluttered images.
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

5 major / 4 minor

Summary. The paper proposes the Visual-Linguistic Agent (VLA), a training-free post-processing pipeline that couples an object detector (Faster R-CNN, YOLOX, YOLO11, DETR, or DINO) with an MLLM (GPT-4o, Claude 3.5, LLaVA, or Gemini 1.5) and CLIP. The detector produces boxes and labels, the MLLM receives the image plus a text listing of detections and flags contextually unreasonable ones, and CLIP reclassifies the flagged crops. The central claim is that VLA improves AP50:95 on COCO by up to 3 percentage points across multiple detectors, with Table 1 reporting gains of +1.3 to +2.7 points and Table 2 reporting DINO+VLA (GPT-4o) at 52.1 vs. 49.5 baseline. An ablation in Tables 3 and 4 reports a 75.0% correction rate on 1,327 pre-identified DINO label errors and corresponding AP gains. Section 3.1 motivates the method with an information-gain formalism (Eqs. 1-3), and Section 5 closes with a short Limitations paragraph.

Significance. If the reported gains are real and reproducible, VLA would be a useful and general contribution: it is a lightweight, model-agnostic wrapper that improves detection accuracy without retraining, and the paper tests a broad set of detector and MLLM combinations. The paper is honest in stating a limitation about NMS-duplicate boxes at the end, and the basic three-stage architecture is clearly described. However, the empirical support as presented is incomplete: no code, prompts, COCO split, detector checkpoints, or run-to-run variance are provided, and the correction-rate metric is not tied to the flag-level behavior that actually drives AP. The theoretical entropy formalism is not operationalized. Therefore the significance is currently conditional on additional diagnostic and reproducibility evidence.

major comments (5)
  1. [§4.1, Table 1] The central quantitative claim is not reproducible or auditable from the information given. The paper does not state which COCO split (val2017, test-dev, or a custom subset) was used, how many images were processed, the exact prompts sent to each MLLM (only illustrative snippets appear in §3.2), the versions/checkpoints of Faster R-CNN, YOLOX, YOLO11, DETR, DINO, and CLIP, or the API inference settings (temperature, max tokens, retries). No error bars or run-to-run variance are reported, and the baseline numbers (e.g., DINO 49.5, YOLO11 48.5) are not cited to a source or shown to come from the authors' own runs. Since the abstract's 'up to 3%' claim rests entirely on these numbers, this omission is load-bearing.
  2. [§3.2.2, §3.2.3, Table 3] The two premises that generate the AP gains are never measured. Premise 1 is that the MLLM flags exactly the mislabeled boxes; the paper reports no per-flag precision or recall, no false-alarm rate, and no count of correct detections that were flagged and later changed to wrong labels. Premise 2 is that CLIP, on the flagged crop, is a better classifier than the detector's classification head; no head-to-head comparison is reported. Table 3's Corrected Rate of 75.0% is computed only on 1,327 pre-identified DINO label errors (ED), so it excludes false positives with no ground-truth match and any new errors introduced by reclassification. Consequently, the measured CR cannot by itself validate the AP improvements in Table 1.
  3. [§4.2, Figure 3] The reclassification protocol is underspecified in a way that directly affects mAP computation. It is never stated whether a flagged detection is deleted from the output or only relabeled, nor how the confidence score is set after CLIP's prediction; both choices change the ranked list of detections and therefore the average precision. In addition, Figure 3 shows a correction to 'moon', which is not one of COCO's 80 categories, but the paper does not explain how out-of-vocabulary reclassifications are encoded in the 'COCO-style JSON' used for evaluation or how they are scored (e.g., whether they become false positives or are ignored).
  4. [§3.1, Eqs. (1)-(3)] The information-gain formalism is not operationalized. No definition is given for P(y_i) and P(y_i, r_ij) in terms of detector or MLLM outputs, no method for estimating IG from the experiments is described, and no measured IG value is reported anywhere. The narrative that VLA reduces entropy is therefore untested and currently serves only as motivation, not as a derivation or explanation of the reported gains. This does not by itself refute the empirical claim, but it means the paper's theoretical framing is unsupported as written.
  5. [§4.1-§4.3] The evaluation does not address the risk that the MLLMs and CLIP were pretrained on COCO images, which is a concrete correctness risk for a post-hoc reasoning system evaluated on COCO. A minimal safeguard would be to report results on a held-out set not present in the MLLM/CLIP pretraining corpora (or on an out-of-distribution benchmark such as Objects365 or LVIS) and/or to test whether flagging behavior depends on image-specific memorization. Without such a test, the claimed generalization across 'multiple detection models' could be confounded by data overlap.
minor comments (4)
  1. [Abstract and Contribution 3] The abstract's 'up to 3% improvement' overstates Table 1's maximum AP50:95 gain of 2.7 points (Faster R-CNN); please report the actual range of deltas and reconcile the wording.
  2. [§5 Limitations] The Limitations paragraph mentions only redundant bounding boxes after NMS; the more consequential limitations (missing flag-level accuracy, no relabeling/confidence protocol, no data-contamination check) should also be acknowledged.
  3. [References] Reference [8] is cited for GPT-4o, Claude 3.5, LLaVA, and Gemini 1.5, but the entry is a control-engineering benchmark paper on GPT-4, Claude 3 Opus, and Gemini 1.0 Ultra; please cite the primary model reports or API documentation instead.
  4. [Tables 2, 3, Figure 4] The spelling 'LLaV A' is inconsistent with 'LLaVA' used in the text; please standardize.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central AP claim is an externally measured benchmark result, and the entropy formalism is not used to generate predictions.

full rationale

The paper's central claim is that VLA improves AP50:95 by up to 3 points on COCO over standard detectors. This claim is supported by direct experimental comparisons in Tables 1, 2, and 4 against external baselines such as Faster R-CNN, YOLOX, YOLO11, DETR, and DINO. No numerical parameter is fitted to the target metric and then reported as a prediction; the correction-rate ablation in Tables 3 and 4 is an auxiliary analysis, not the basis for the central claim. The information-gain equations (1)--(3) are motivational and are not used to compute detections or AP, so the derivation chain does not reduce to its inputs. The pipeline's flagging and reclassification steps are evaluated through the final COCO-style JSON and mAP, giving independent grounding. The noted weaknesses--lack of flag-level precision/recall, no head-to-head CLIP versus detector classification comparisons, and potential pretraining contamination of GPT-4o or CLIP on COCO--are external-validity or evaluation-robustness concerns, not circularity. There is also no load-bearing self-citation: the cited references are to external systems and papers, and no uniqueness theorem from the authors is invoked to force a choice. Accordingly, the paper is self-contained against external benchmarks and receives a circularity score of 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 1 invented entities

The paper introduces no fitted numerical parameters; the claimed gains are measured, not derived. Its load-bearing assumptions are behavioral: the MLLM's flags must be accurate and CLIP's relabels must beat the detector's classifier. The only new construct is the VLA orchestration itself, which has no external evidence independent of the reported COCO numbers.

assumptions (3)
  • domain assumption The MLLM can accurately judge whether a detection label matches the image content from text coordinates and a scene caption.
    Section 3.2.2 assumes this filtering step works; no per-flag false-alarm rate is reported.
  • domain assumption CLIP is a more accurate classifier than the detector's classification head on flagged regions.
    Section 3.2.3 delegates relabeling to CLIP; no head-to-head comparison with the detector's classifier is provided.
  • domain assumption The COCO evaluation setup and baseline numbers are comparable to standard practice.
    Section 4.1 does not specify train/val split, checkpoint sources, NMS settings, or prompt details; comparability is assumed.
invented entities (1)
  • Visual-Linguistic Agent (VLA)
    purpose: Orchestrate an MLLM critic and CLIP classifier to refine detector outputs
    The framework is new in this paper; no code or independent benchmark exists yet, and its components are pre-existing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Visual-Linguistic Agent: Towards Collaborative Contextual Object Reasoning." pith.science (2026). https://pith.science/paper/U7SI3UBN

@misc{pith2026241110252,
  author       = {Pith},
  title        = {Pith review of: Visual-Linguistic Agent: Towards Collaborative Contextual Object Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U7SI3UBN}},
  note         = {Machine review of arXiv:2411.10252}
}
read the original abstract

Multimodal Large Language Models (MLLMs) excel at descriptive tasks within images but often struggle with precise object localization, a critical element for reliable visual interpretation. In contrast, traditional object detection models provide high localization accuracy but frequently generate detections lacking contextual coherence due to limited modeling of inter-object relationships. To address this fundamental limitation, we introduce the \textbf{Visual-Linguistic Agent (VLA), a collaborative framework that combines the relational reasoning strengths of MLLMs with the precise localization capabilities of traditional object detectors. In the VLA paradigm, the MLLM serves as a central Linguistic Agent, working collaboratively with specialized Vision Agents for object detection and classification. The Linguistic Agent evaluates and refines detections by reasoning over spatial and contextual relationships among objects, while the classification Vision Agent offers corrective feedback to improve classification accuracy. This collaborative approach enables VLA to significantly enhance both spatial reasoning and object localization, addressing key challenges in multimodal understanding. Extensive evaluations on the COCO dataset demonstrate substantial performance improvements across multiple detection models, highlighting VLA's potential to set a new benchmark in accurate and contextually coherent object detection.

Figures

Figures reproduced from arXiv: 2411.10252 by the authors.

Figure 1
Figure 1. An example from the proposed Visual-Linguistic-Agent (VLA) paradigm. Here, the Visual Agent (YOLO) detects objects in an image but misclassifies the moon as an orange. The Linguistic Agent (MLLM) evaluates these detection results and, using its reasoning capabilities, identifies and corrects the error. This example demonstrates the effectiveness of integrating visual detection with linguistic reasoning to enhance th… view at source ↗
Figure 2
Figure 2. Comparison of traditional object detection (a) and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The proposed Visual-Linguistic-Agent (VLA) paradigm. The Visual Agent (e.g., YOLO) detects objects and generates bounding boxes with class labels, which are passed to the Linguistic Agent (MLLM) for reasoning and contextual analysis. Based on the MLLM’s assessment, false detections are filtered, and the Classification Visual Agent corrects erroneous detections. This collaboration between agents enhances object detec… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Examples of error correction by VLA using DINO with MLLMs as the Linguistic Agent. The figure highlights common [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 17 canonical work pages

  1. [1]

    Vqa: Visual question answering

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision, 2015. 2

    work page 2015

  2. [2]

    End-to- end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to- end object detection with transformers. In European Confer- ence on Computer Vision, pages 213–229. Springer, 2020. 2, 6

    work page 2020

  3. [3]

    Spatial memory for context reasoning in object detection

    Xinlei Chen and Abhinav Gupta. Spatial memory for context reasoning in object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 4086– 4096, 2017. 2, 3

    work page 2017

  4. [4]

    Palm- e: An embodied multimodal language model

    Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm- e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2, 3

    arXiv 2023

  5. [5]

    Relation networks for object detection

    Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object detection. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3588–3597, 2018. 3

    work page 2018

  6. [6]

    Dac-detr: Divide the attention layers and conquer

    Zhengdong Hu, Yifan Sun, Jingdong Wang, and Yi Yang. Dac-detr: Divide the attention layers and conquer. Advances in Neural Information Processing Systems, 36, 2024. 2

    work page 2024

  7. [7]

    Hugging face

    Shashank Mohan Jain. Hugging face. In Introduction to transformers for NLP: With the hugging face library and models to solve problems, pages 51–67. Springer, 2022. 3

    work page 2022

  8. [8]

    Capabilities of large language models in control engi- neering: A benchmark study on gpt-4, claude 3 opus, and gemini 1.0 ultra

    Darioush Kevian, Usman Syed, Xingang Guo, Aaron Havens, Geir Dullerud, Peter Seiler, Lianhui Qin, and Bin Hu. Capabilities of large language models in control engi- neering: A benchmark study on gpt-4, claude 3 opus, and gemini 1.0 ultra. arXiv preprint arXiv:2404.03647, 2024. 6

    arXiv 2024

Show all 32 references
  1. [9]

    Yolov11: An overview of the key architectural enhancements

    Rahima Khanam and Muhammad Hussain. Yolov11: An overview of the key architectural enhancements. arXiv preprint arXiv:2410.17725, 2024. 6

    2024 arXiv

  2. [10]

    Seed-bench: Bench- marking multimodal large language models

    Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Bench- marking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 13299–13308, 2024. 2, 3

    2024

  3. [11]

    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. ICML,

  4. [12]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740–755. Springer, 2014. 2

    2014

  5. [13]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485,

    arXiv

  6. [14]

    Cigar: Cross-modality graph reasoning for domain adaptive object detection

    Yabo Liu, Jinghua Wang, Chao Huang, Yaowei Wang, and Yong Xu. Cigar: Cross-modality graph reasoning for domain adaptive object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 23776–23786, 2023. 2, 3

    2023

  7. [15]

    Rt-gcn: Gaussian-based spatiotemporal graph convolutional network for robust traffic prediction.In- formation Fusion, 102:102078, 2024

    Yutian Liu, Soora Rasouli, Melvin Wong, Tao Feng, and Tianjin Huang. Rt-gcn: Gaussian-based spatiotemporal graph convolutional network for robust traffic prediction.In- formation Fusion, 102:102078, 2024. 2, 3

    2024

  8. [16]

    Compositional chain-of-thought prompting for large multimodal models

    Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 14420–14431, 2024. 3

    2024

  9. [17]

    Improving multimodal datasets with image captioning

    Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Se- woong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. Advances in Neural Infor- mation Processing Systems, 36, 2024. 2

    2024

  10. [18]

    Faster r-cnn: Towards real-time object detection with region proposal networks

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137–1149, 2016. 6

    2016

  11. [19]

    Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face

    Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems, 36, 2024. 3

    2024

  12. [20]

    Vipergpt: Visual inference via python execution for reasoning

    D ´ıdac Sur´ıs, Sachit Menon, and Carl V ondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11888–11898, 2023. 3

    2023

  13. [21]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth´ee Lacroix, Baptiste Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2, 3

    2023 arXiv

  14. [22]

    Sw-yolox: A yolox-based real-time pedestrian detector with shift window- mixed attention mechanism

    Chi-Yi Tsai, Run-Yu Wang, and Yu-Chen Chiu. Sw-yolox: A yolox-based real-time pedestrian detector with shift window- mixed attention mechanism. Neurocomputing, 606:128357,

  15. [23]

    Robust motor- cycle helmet detection in real-world scenarios: Using co- detr and minority class enhancement

    Hao V o, Sieu Tran, Duc Minh Nguyen, Thua Nguyen, Tien Do, Duy-Dinh Le, and Thanh Duc Ngo. Robust motor- cycle helmet detection in real-world scenarios: Using co- detr and minority class enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Reco...

    2024

  16. [24]

    Chain-of-thought prompting elicits reasoning in large lan- guage models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large lan- guage models. Advances in Neural Information Processing Systems, 35:24824–24837, 2022. 3

    2022

  17. [25]

    Spatial-aware graph relation network for large-scale object detection

    Hang Xu, Chenhan Jiang, Xiaodan Liang, and Zhenguo Li. Spatial-aware graph relation network for large-scale object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9298– 9307, 2019. 2, 3

    2019

  18. [26]

    Dino: Detr with improved denoising anchor boxes for end-to-end object de- tection

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object de- tection. In The Eleventh International Conference on Learn- ing Representations. 6

  19. [27]

    Dino: Detr with improved denoising anchor boxes for end-to-end object detection

    Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2

    2022 arXiv

  20. [28]

    Ms-detr: Efficient detr training with mixed supervision

    Chuyang Zhao, Yifan Sun, Wenhao Wang, Qiang Chen, Er- rui Ding, Yi Yang, and Jingdong Wang. Ms-detr: Efficient detr training with mixed supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17027–17036, 2024. 2

    2024

  21. [29]

    Rgrn: Relation-aware graph reasoning network for object de- tection

    Jianjun Zhao, Jun Chu, Lu Leng, Chaolin Pan, and Tao Jia. Rgrn: Relation-aware graph reasoning network for object de- tection. Neural Computing and Applications, 35(22):16671– 16688, 2023. 2, 3

    2023

  22. [30]

    Semantic relation reasoning for shot- stable few-shot object detection

    Chenchen Zhu, Fangyi Chen, Uzair Ahmed, Zhiqiang Shen, and Marios Savvides. Semantic relation reasoning for shot- stable few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8782–8791, 2021. 3

    2021

  23. [31]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mo- hamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3

    2023 arXiv

  24. [32]

    An efficient two-state gru based on feature attention mechanism for sen- timent analysis

    Muhammad Zulqarnain, Rozaida Ghazali, Muhammad Aamir, and Yana Mazwin Mohmad Hassim. An efficient two-state gru based on feature attention mechanism for sen- timent analysis. Multimedia Tools and Applications, 83(1): 3085–3110, 2024. 2, 3

    2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.