Pith. sign in

REVIEW 3 major objections 4 minor 39 references

RT-VLM: Re-Thinking Vision Language Model with 4-Clues for Real-World Object Recognition Robustness

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read Structured 4-clue supervision plus a self-critique loop makes a vision-language model robust to blur, viewpoint, occlusion, and class confusion.

desk verdict Plausible recipe with a synthetic 4-clue dataset and a self-refine loop, but the evaluation doesn't establish the robustness claim because of an unaddressed JSON-format confound and weak baselines. read the letter →

arxiv 2509.05333 v1 pith:C4UDYPJU submitted 2025-09-01 cs.CV cs.AI

classification cs.CVcs.AI
keywords objectrecognitiondomainshiftvisionlanguagemodelsself-correctionsyntheticdata4-cluessupervisionparameter-efficientfine-tuningrobustnessbenchmarks
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

This paper argues that object recognition robustness under real-world shifts comes from richer, structured supervision plus a deliberate self-correction pass, not from more data alone. It builds a synthetic training set in which every image carries four clues — bounding boxes, class names, object-level captions, and a scene-level caption — and fine-tunes a large vision-language model on this JSON-style output. At inference, the model first emits its own four clues, then re-reads them under a prompt that asks it to correct errors caused by covariate shift, viewpoint, occlusion, or class confusion. Across benchmarks isolating each shift, the full system beats the untuned base model and single-pass variants. The authors also report a cost: the fine-tuned model's free-form scene captioning degrades on standard captioning data.

What carries the argument

The 4-Clues representation: a per-image JSON schema containing bounding boxes, class names, object-level captions, and a scene-level context caption. It serves as both the training supervision and the intermediate evidence that the Re-Thinking prompt re-examines. The two-stage inference re-injects the model's own first-pass JSON with instructions to correct objects missed or mislabelled under the four failure modes; fine-tuning on the synthetic corpus is what makes the model able to emit and parse this structure.

What would settle it

A concrete check: run the base model and RT-VLM on the same images while counting how often each returns a parseable 4-Clues JSON. If the base model's valid-parse rate is far below RT-VLM's, most of the reported robustness gap could be a formatting artifact rather than perception; equalizing parse rates would shrink or eliminate the gap. Also, removing the Re-Thinking prompt and just asking 'correct your answer' should show whether the structured four-clue framing, rather than generic self-correction, drives the gain.

Watch

Extended reading notes

Core claim

The central claim is that pairing dense multi-clue supervision with a two-stage self-correction mechanism produces consistent robustness gains across four distinct domain shifts. The 4-Clues representation — boxes, classes, object captions, context caption — gives the model a shared evidence format; the Re-Thinking stage feeds the first pass back through the same model with a prompt that names the four failure modes and asks for corrections. On the reported benchmarks, the full RT-VLM reaches, for example, mAP@0.5 of 0.75 on COCO-C versus 0.48 for the base model, 0.80 on Pascal3D+ versus 0.57, and 74.11% top-1 on ImageNet-A versus 64.81%; the Re-Thinking stage adds a smaller but consistent g

Load-bearing premise

The evaluation assumes the base vision-language model's low detection scores reflect visual weakness rather than its inability to output the required JSON schema; the paper does not report valid-parse rates, so the baseline numbers could be artificially low.

Editorial extensions

If this is right

  • If the claim holds, robustness to corruption, viewpoint, occlusion, and class confusion can be improved by changing the output representation and adding a self-critique pass, without new real annotated data.
  • The Re-Thinking gain is contingent on the model being fine-tuned on the 4-Clues format; applying the same prompt to an untuned generalist yields little benefit.
  • Object-centric specialization comes at a measurable cost to open-ended scene captioning, so deployment needs to match task type.
  • The full pipeline runs on a single 24GB consumer GPU, suggesting the recipe is accessible without large-scale compute.
  • Because the training data is synthetic, the same pipeline can be regenerated for new object vocabularies or new shift types.

Reading between the lines

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

  • If the base model's lower scores are partly JSON-format failures rather than visual failures, the absolute gap between the base model and RT-VLM would shrink; reporting valid-parse rates would separate format adherence from perceptual robustness.
  • The Re-Thinking design suggests a general recipe: any structured prediction task (e.g., relations, attributes, counts) could get a self-correcting second pass if the first pass emits a parseable evidence schema.
  • A direct test: add the Re-Thinking prompt to a base model that is given a few-shot example of the 4-Clues JSON; if the gain equals that of fine-tuning, the benefit may be prompt-format rather than learned structure.
  • The synthetic-data pipeline could be stressed by measuring how detector label noise (mAP 0.83 on the synthetic domain) propagates into final recognition; lower detector quality likely caps the ceiling.
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

3 major / 4 minor

Summary. The paper proposes RT-VLM, a framework for improving object recognition robustness under four domain shifts (covariate shift, viewpoint variation, occlusion, and class confusion). The method has three components: (i) a synthetic dataset of 20,000 FLUX.1-dev images annotated with '4-Clues' (bounding boxes, class names, object-level captions, and a context-level caption, generated by a YOLO12x detector and Janus Pro 7B); (ii) parameter-efficient LoRA fine-tuning of Llama 3.2 11B Vision Instruct on prompt/response pairs that require the model to emit these four clues as JSON; and (iii) a two-stage 'Re-Thinking' inference in which the model first emits its own 4-clue prediction and then, prompted to consider the four failure modes, revises that prediction. The paper reports that the full RT-VLM + Re-Thinking system outperforms a base VLM, base VLM + Re-Thinking, and RT-VLM without Re-Thinking on custom synthetic data, COCO val2017, COCO-C, Pascal-C, Pascal3D+, ObjectNet3D, OccludedPASCAL3D+, and ImageNet-A.

Significance. If the empirical gains are real, the paper offers a practical and parameter-efficient way to make VLMs more robust, and it makes a worthwhile conceptual contribution by linking structured multimodal supervision to a self-correction loop. Deserving of credit: the framework is clearly described at a high level, runs on a single consumer GPU, and is evaluated on a broad collection of standard robustness benchmarks, not merely one in-house set. The strengths are the explicit decomposition of failure modes, the use of external benchmarks, and the consistent direction of the reported gains. However, the central comparative claim is currently not fully established because the evaluation does not account for the possibility that the base model's outputs are less parseable under the required JSON schema, and because several methodological details (loss definitions, no error bars, in-distribution custom test set) weaken confidence in the numbers.

major comments (3)
  1. [Experimental Evaluation, Tables 5-6, 9-12; Key Finding 1] All object-recognition metrics are computed from model outputs assumed to be JSON objects in the 4-Clue schema, but the paper reports no valid-parse rate, parser failure mode, or fallback extraction for any variant. For Base VLM (A) and Base VLM + Re-Thinking (B), which were never fine-tuned to emit this schema, the paper itself states in Key Finding 1 that the base model 'lacks explicit exposure to the 4-Clues representation, therefore its capacity to generate or parse that structure is limited.' Consequently, the large gaps between A/B and C/D may partly reflect malformed JSON being scored as missed detections, not a visual robustness gain. The claim of 'consistently surpasses strong baselines' is not established until parse rates are reported or a format-controlled baseline is included.
  2. [Step 2: Training RT-VLM with Multi Clue Supervision, Eqs. (5)-(8)] The composite VLM loss is not reproducible. Eq. (7) defines L_schema as -Σ_t log P(t_gold) + λ N_invalid/|y| and then equates it to -Σ_t w_s log P(t), but N_invalid, w_s, and the relationship between the two expressions are undefined. The phrase 'implicitly captured by the primary loss' is also contradictory with its presence in Eq. (5). Eq. (8) introduces w_y and γ without relating them to the label-smoothing CE in Eq. (6) or the 0.2 weight in Eq. (5). Because the method's central novelty is this supervision scheme, the objective should be stated precisely.
  3. [Experimental Setup, Tables 5, 7, 8] The custom-generated test set (5,000 synthetic images) is produced by the same FLUX.1-dev prompt protocol used to build the training corpus. Tables 5, 7, and 8 therefore measure in-distribution performance and cannot independently support transfer or robustness claims. The external benchmarks in Tables 6, 9-12 are the right evidence; the paper should either demote the custom-set numbers or characterize the domain gap between the synthetic training distribution and the custom test set. As written, this is an avoidable circularity concern.
minor comments (4)
  1. [All experimental tables] No error bars, multiple runs, or statistical significance tests are reported. Given that some gains are modest (e.g., A vs B in Table 5, and B vs A in several captioning rows), the reader cannot judge which differences are meaningful. Please report at least 3 seeds with mean and standard deviation, or justify why the reported runs are deterministic.
  2. [Tables 7 and 8] The captioning evaluation protocol is underspecified. For the custom test set, it is not stated which reference captions are used for CIDEr/BLEU-4, how object-level captions are extracted from the JSON output, or how they are matched to the annotated regions. This makes the captioning numbers difficult to interpret and reproduce.
  3. [Eq. (1)] The SIoU formula appears malformed: L_SIOU = CIOU + Λ + ∆ + Ω/2 mixes notation and does not match the standard SIoU definition. Please correct the equation or replace it with a citation, since the exact form is not used elsewhere in the paper.
  4. [Throughout] The manuscript has numerous typographical and spacing errors (e.g., 'V iewpoint' in the Introduction, inconsistent hyphenation of 'object-level' and 'view-point'). A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the training, inference, and evaluation chain is self-contained and externally benchmarked.

full rationale

The paper's central claim is empirical: fine-tuning a VLM on a synthetic 4-Clues dataset and applying a two-stage Re-Thinking inference yields superior robustness across external benchmarks. The derivation chain (dataset generation -> supervised tuning -> self-correction inference) does not reduce to its inputs by construction. No parameter is fitted to the evaluation benchmarks; the comparison variants control for the prompt (A vs. B differ only by Re-Thinking, C vs. D similarly), and the reported gains on COCO-C, Pascal-C, Pascal3D+, ObjectNet3D, OccludedPASCAL3D+, and ImageNet-A are external, not generated by the paper's own pipeline. There are no load-bearing self-citations: the cited works (FLUX, Janus-Pro, YOLOv12, Llama, LoRA, SELF-REFINE, etc.) are external tools or prior art, and there is no invocation of a self-authored uniqueness theorem or ansatz smuggled in via citation. The Re-Thinking prompt names the four targeted failure modes, but that is a design choice, not a tautology, because the baseline B receives the same prompt and the contribution of fine-tuning is isolated. The most serious concern in the paper is the unreported valid-JSON parse rate for the untuned baselines A and B in Tables 5-12; if those baselines emit malformed outputs, the detection metrics are artificially depressed. However, that is an evaluation-validity and reporting issue, not mathematical circularity: it does not make any predicted quantity equivalent to a fitted input by definition. The custom synthetic test set is in-distribution for the generator, which may inflate some results, but the external robustness benchmarks carry the central claim. Therefore, under the strict rule that circularity must be exhibited by quotation and reduction, no circular step is present.

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

The method rests on several unverified domain assumptions: the synthetic data transfers to real robustness, the automatic labels are of sufficient quality, and the evaluation protocol is fair. These are load-bearing but not defended experimentally.

free parameters (3)
  • YOLO loss hyperparameters = alpha=0.25, gamma=1.5, lambda_box=7.5, lambda_dfl=1.5, lambda_cls=0.5
    Chosen by hand to train the YOLO12x detector that labels the synthetic dataset; these directly affect label quality.
  • VLM composite loss weight = 0.2
    Weight for the schema and class-balanced focal loss terms in Eq. 5; chosen without ablation.
  • Prompt set sizes = 100 base prompts, 20 modifiers, 200 images per triplet
    Hand-designed to produce 20,000 synthetic images; the specific prompts are not listed.
assumptions (3)
  • domain assumption Synthetic images generated by FLUX.1-dev are sufficiently realistic to transfer to real-world robustness benchmarks.
    The whole method relies on this transfer; the paper only tests on synthetic test set from the same generator plus real benchmarks, but no analysis of the domain gap.
  • domain assumption The YOLO12x detector trained on 1,000 human labels provides ground-truth-quality boxes and class labels for the remaining 19,000 images.
    No measurement of label noise impact; mAP@0.5 of 0.83 indicates substantial noise.
  • domain assumption The base VLM's lower scores reflect visual robustness rather than an inability to follow the required JSON output format.
    The paper does not report parse rates or handle malformed outputs; if the base model fails to emit JSON, its detection metrics are artificially low.
invented entities (1)
  • 4-Clues annotation schema
    purpose: Provides structured supervision (boxes, class names, object captions, context caption) to teach the VLM robust recognition.
    This is a new data representation introduced by the paper; it is not an independently verifiable entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RT-VLM: Re-Thinking Vision Language Model with 4-Clues for Real-World Object Recognition Robustness." pith.science (2026). https://pith.science/paper/C4UDYPJU

@misc{pith2026250905333,
  author       = {Pith},
  title        = {Pith review of: RT-VLM: Re-Thinking Vision Language Model with 4-Clues for Real-World Object Recognition Robustness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4UDYPJU}},
  note         = {Machine review of arXiv:2509.05333}
}
read the original abstract

Real world deployments often expose modern object recognition models to domain shifts that precipitate a severe drop in accuracy. Such shifts encompass (i) variations in low level image statistics, (ii) changes in object pose and viewpoint, (iii) partial occlusion, and (iv) visual confusion across adjacent classes. To mitigate this degradation, we introduce the Re-Thinking Vision Language Model (RT-VLM) framework. The foundation of this framework is a unique synthetic dataset generation pipeline that produces images annotated with "4-Clues": precise bounding boxes, class names, detailed object-level captions, and a comprehensive context-level caption for the entire scene. We then perform parameter efficient supervised tuning of Llama 3.2 11B Vision Instruct on this resource. At inference time, a two stage Re-Thinking scheme is executed: the model first emits its own four clues, then re examines these responses as evidence and iteratively corrects them. Across robustness benchmarks that isolate individual domain shifts, RT-VLM consistently surpasses strong baselines. These findings indicate that the integration of structured multimodal evidence with an explicit self critique loop constitutes a promising route toward reliable and transferable visual understanding.

Figures

Figures reproduced from arXiv: 2509.05333 by the authors.

Figure 1
Figure 1. Three step pipeline comprising dataset generation, training, and two stage inference. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 27 canonical work pages

  1. [1]

    Bauer, A.; Trapp, S.; Stenger, M.; Leppich, R.; Kounev, S.; Leznik, M.; Chard, K.; and Foster, I. 2024. Comprehensive Exploration of Synthetic Data Generation: A Survey . arXiv preprint arXiv:2401.02524

  2. [2]

    Black Forest Labs ; Batifol, S.; Blattmann, A.; Boesel, F.; Consul, S.; Diagne, C.; Dockhorn, T.; English, J.; English, Z.; Esser, P.; Kulal, S.; Lacey, K.; Levi, Y.; Li, C.; Lorenz, D.; M \"u ller, J.; Podell, D.; Rombach, R.; Saini, H.; Sauer, A.; and Smith, L. 2025. FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Spa...

  3. [3]

    Carion, N.; Massa, F.; Synnaeve, G.; Usunier, N.; Kirillov, A.; and Zagoruyko, S. 2020. End-to-End Object Detection with Transformers . In European Conference on Computer Vision (ECCV)

  4. [4]

    Chen, X.; Fang, H.; Lin, T.-Y.; Vedantam, R.; Gupta, S.; Doll \'a r, P.; and Zitnick, C. L. 2015. Microsoft COCO Captions: Data Collection and Evaluation Server . arXiv preprint arXiv:1504.00325

  5. [5]

    DeepSeek AI ; Chen, X.; Wu, Z.; Liu, X.; Pan, Z.; Liu, W.; Xie, Z.; Yu, X.; and Ruan, C. 2025. Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling . arXiv preprint arXiv:2501.17811

  6. [6]

    Dettmers, T.; Pagnoni, A.; Holtzman, A.; and Zettlemoyer, L. 2023. QLORA: Efficient Finetuning of Quantized LLMs . arXiv preprint arXiv:2305.14314

  7. [7]

    Dong, Y.; Ruan, S.-Y.; Su, H.; Kang, C.-M.; Wei, X.; and Zhu, J. 2022. ViewFool: Evaluating the Robustness of Visual Recognition to Adversarial Viewpoints . In Advances in Neural Information Processing Systems 35 (NeurIPS)

  8. [8]

    V.; Williams, C

    Everingham, M.; Gool, L. V.; Williams, C. K. I.; Winn, J.; and Zisserman, A. 2010. The Pascal Visual Object Classes (VOC) Challenge . International Journal of Computer Vision, 88(2): 303--338

Show all 39 references
  1. [9]

    Ganin, Y.; Ustinova, E.; Ajakan, H.; Germain, P.; Larochelle, H.; Laviolette, F.; Marchand, M.; and Lempitsky, V. 2016. Domain-Adversarial Training of Neural Networks . Journal of Machine Learning Research, 17(59): 1--35

  2. [10]

    Gevorgyan, Z. 2022. SIoU Loss: More Powerful Learning for Bounding Box Regression . arXiv preprint arXiv:2205.12740

  3. [11]

    H.; de Wit, A

    Grattafiori, A.; Dubey, A.; Jauhri, A.; Pandey, A.; Williams, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; El-Nouby, A.; Liu, A. H.; de Wit, A. R. P. M. M.; Schneer, A.; Mourachko, A.; Almahairi, A.; Yang, A.; Andreassen, A.; Touati, A.; Goyal, A.; Hartshorn, A.; Ya...

  4. [12]

    Hendrycks, D.; and Dietterich, T. 2019. Benchmarking Neural Network Robustness to Common Corruptions and Perturbations . In International Conference on Learning Representations (ICLR)

  5. [13]

    Hendrycks, D.; Zhao, K.; Basart, S.; Steinhardt, J.; and Song, D. 2021. Natural Adversarial Examples . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  6. [14]

    D.; Gesmundo, A.; Attariyan, M.; and Gelly, S

    Houlsby, N.; Giurgiu, A.; Jastrzebski, S.; Morrone, B.; Laroussilhe, Q. D.; Gesmundo, A.; Attariyan, M.; and Gelly, S. 2019. Parameter-Efficient Transfer Learning for NLP . In Proceedings of the 36th International Conference on Machine Learning (ICML)

  7. [15]

    Hsiao, E.; and Hebert, M. 2012. Occlusion Reasoning for Object Detection under Arbitrary Viewpoint . In 2012 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

  8. [16]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2022. LoRA: Low-Rank Adaptation of Large Language Models . In International Conference on Learning Representations (ICLR)

  9. [17]

    H.; Zhang, P.; Zhang, H.; Yang, J.; Li, C.; Zhong, Y.; Wang, L.; Yuan, L.; Zhang, L.; Hwang, J.-H.; and Gao, J

    Li, L. H.; Zhang, P.; Zhang, H.; Yang, J.; Li, C.; Zhong, Y.; Wang, L.; Yuan, L.; Zhang, L.; Hwang, J.-H.; and Gao, J. 2022. Grounded Language-Image Pre-training . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  10. [18]

    Li, X.; Wang, W.; Wu, L.; Chen, S.; Hu, X.; Li, J.; Tang, J.; and Yang, J. 2020. Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection . In Advances in Neural Information Processing Systems 33 (NeurIPS)

  11. [19]

    Lin, T.-Y.; Goyal, P.; Girshick, R.; He, K.; and Doll \'a r, P. 2017. Focal Loss for Dense Object Detection . In Proceedings of the IEEE International Conference on Computer Vision (ICCV)

  12. [20]

    Lin, T.-Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Doll \'a r, P.; and Zitnick, C. L. 2014. Microsoft COCO: Common Objects in Context . In European Conference on Computer Vision (ECCV)

  13. [21]

    Liu, H.; Li, C.; Wu, Q.; and Lee, Y. J. 2023 a . Visual Instruction Tuning . In Advances in Neural Information Processing Systems 36 (NeurIPS)

  14. [22]

    Liu, S.; Zeng, Z.; Ren, T.; Li, F.; Zhang, H.; Yang, J.; Jiang, Q.; Li, C.; Yang, J.; Su, H.; Zhu, J.; and Zhang, L. 2023 b . Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection . arXiv preprint arXiv:2303.05499

  15. [23]

    P.; Gupta, S.; Yazdanbakhsh, A.; and Clark, P

    Madaan, A.; Tandon, N.; Gupta, P.; Hallinan, S.; Gao, L.; Wiegreffe, S.; Alon, U.; Dziri, N.; Prabhumoye, S.; Yang, Y.; Welleck, S.; Majumder, B. P.; Gupta, S.; Yazdanbakhsh, A.; and Clark, P. 2023. Self-Refine: Iterative Refinement with Self-Feedback . In Advances in Neural I...

  16. [24]

    Papineni, K.; Roukos, S.; Ward, T.; and Zhu, W.-J. 2002. BLEU: a Method for Automatic Evaluation of Machine Translation . In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL)

  17. [25]

    A.; Wang, L.; Cervantes, C

    Plummer, B. A.; Wang, L.; Cervantes, C. M.; Caicedo, J. C.; Hockenmaier, J.; and Lazebnik, S. 2015. Flickr30k Entities: Collecting Region-to-Phrase Correspondences for Richer Image-to-Sentence Models . In Proceedings of the IEEE International Conference on Computer Vision (ICCV)

  18. [26]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I. 2021. Learning Transferable Visual Models From Natural Language Supervision . In Proceedings of the 38th International Confere...

  19. [27]

    Redmon, J.; Divvala, S.; Girshick, R.; and Farhadi, A. 2016. You Only Look Once: Unified, Real-Time Object Detection . In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

  20. [28]

    Ren, S.; He, K.; Girshick, R.; and Sun, J. 2015. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks . In Advances in Neural Information Processing Systems 28 (NIPS)

  21. [29]

    Shimodaira, H. 2000. Improving predictive inference under covariate shift by weighting the log-likelihood function . Journal of Statistical Planning and Inference, 90(2): 227--244

  22. [30]

    Tian, Y.; Ye, Q.; and Doermann, D. 2025. YOLOv12: Attention-Centric Real-Time Object Detectors . arXiv preprint arXiv:2502.12524

  23. [31]

    van Rijsbergen, C. J. 1979. Information Retrieval . Butterworth-Heinemann

  24. [32]

    L.; and Parikh, D

    Vedantam, R.; Zitnick, C. L.; and Parikh, D. 2015. CIDEr: Consensus-Based Image Description Evaluation . In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

  25. [33]

    Wang, A.; Sun, Y.; Kortylewski, A.; and Yuille, A. 2020. Robust Object Detection under Occlusion with Context-Aware Compositional Nets . In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  26. [34]

    Wang, S.; Veldhuis, R.; Brune, C.; and Strisciuglio, N. 2023. A Survey on the Robustness of Computer Vision Models against Common Corruptions . arXiv preprint arXiv:2305.06024

  27. [35]

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q.; and Zhou, D. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models . In Advances in Neural Information Processing Systems 35 (NeurIPS)

  28. [36]

    B.; Su, H.; Mottaghi, R.; Guibas, L.; and Savarese, S

    Xiang, Y.; Kim, W.; Chen, W.; Ji, J.; Choy, C. B.; Su, H.; Mottaghi, R.; Guibas, L.; and Savarese, S. 2016. ObjectNet3D: A Large Scale Database for 3D Object Recognition . In European Conference on Computer Vision (ECCV)

  29. [37]

    Xiang, Y.; Mottaghi, R.; and Savarese, S. 2014. Beyond PASCAL: A Benchmark for 3D Object Detection in the Wild . In IEEE Winter Conference on Applications of Computer Vision (WACV)

  30. [38]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  31. [39]

    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 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.