Pith. sign in

REVIEW 3 major objections 5 minor 25 references

VehAnchor: Metadata-Free Metric Scale Recovery from Vehicle Cues in Aerial Imagery

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a single aerial image's metric scale can be recovered by measuring the most common pixel length of parked cars, achieving 6.87% median GSD error on DOTA v1.5.

desk verdict Solid in-distribution scale-recovery pipeline plus a useful VLM hallucination benchmark; the headline 6.87% error depends on a dataset-fitted prior, so 'metadata-free' is overstated. read the letter →

arxiv 2603.04277 v2 pith:VS6Z7TVS submitted 2026-03-04 cs.RO cs.AI

Yifei Chen , Chenqian Le , Jiayi Cheng , Xupeng Chen This is my paper · ORCID
classification cs.ROcs.AI
keywords groundsampledistanceaerialimageryscalerecoveryvehicleanchorskerneldensityestimationorientedboundingboxesvision-languagemodelsGPS-deniednavigation
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 tries to establish that an autonomous aerial robot can recover absolute metric scale — the ground sample distance, or metres per pixel — from a single monocular image, with no GPS, altitude, or camera metadata. The proposed tool detects small vehicles, finds the modal pixel length of their oriented bounding boxes via kernel density estimation, and divides a pre-calibrated reference length (5.045 m) by that mode. On the DOTA v1.5 benchmark it reports 6.87% median GSD error on 306 images. The paper also claims that five state-of-the-art vision-language models hallucinate scale, with median area errors above 50%, and that equipping such agents with a deterministic geometric tool reduces area-estimation error to 19.7% median with fewer catastrophic failures. The value of the claim is that a cheap, repeatable measurement can ground LLM/VLM planners in physics.

What carries the argument

The load-bearing object is the vehicle-anchored GSD identity GSD = L_ref / P_mode. It turns a frequency statement — 'small cars in this scene are about L_ref metres long' — into a pixel-to-metre conversion. Robustness comes from the KDE mode estimator, which replaces the mean of detected pixel lengths and gives a 17% relative improvement, and from the confidence score with a hard resolution guard, which prevents the tool from silently returning an estimate when vehicles are too few or the ground resolution too coarse. The pre-calibrated reference length L_ref = 5.045 m is the calibrated constant that carries the physical prior.

What would settle it

Take a diverse set of aerial images with known GSD and known typical vehicle fleet (e.g., a region dominated by long pickups or buses), compute GSD = 5.045 / P_mode, and compare to ground truth; if the median error exceeds the claimed 6.87% (or roughly doubles, as the paper's own ±0.5 m sensitivity test predicts), the central claim is falsified. Concretely, the paper's Table III already shows that a 0.5 m miss in L_ref raises median error from 6.88% to about 10.6%, so a fleet whose modal vehicle length is 5.5 m rather than 5.045 m should be visible as an out-of-band error.

Watch

Extended reading notes

Core claim

The paper's central claim is that the modal physical length of small vehicles is stable enough to act as a universal ruler in aerial images. The pipeline measures vehicle pixel lengths with an oriented object detector, filters outliers, finds the mode of the pixel-length distribution by kernel density estimation, and converts to GSD via GSD = L_ref / P_mode. L_ref = 5.045 m is calibrated from the DOTA v1.5 training annotations, where it is the KDE mode of physical small-vehicle lengths. The paper reports that this statistical anchoring achieves 6.87% median GSD error end-to-end on 306 validation images, close to the 6.88% error of the ground-truth-annotation upper bound, and that a resolutio

Load-bearing premise

The entire conversion rests on assuming that the most common physical length of small vehicles in the deployment scene matches the calibrated L_ref = 5.045 m measured from Chinese urban fleet annotations; if the scene has larger or smaller typical cars, the claimed error does not transfer.

Editorial extensions

If this is right

  • An aerial robot that loses GPS and camera metadata can still convert pixel measurements to metres with sub-10% median error whenever the scene contains at least a handful of small vehicles and sub-metre resolution.
  • LLM/VLM planners can be made scale-safe by calling a deterministic tool instead of estimating area from pixels; the paper reports a 2.6x lower category dependence and 4x fewer catastrophic failures.
  • The same anchor logic transfers to other objects with known modal lengths, allowing the method to be extended beyond vehicles.
  • The confidence score and resolution guard give the calling agent a principled reason to fall back to alternative localisation when the estimate cannot be trusted.
  • Calibration must be redone for each region/fleet; a ±0.5 m error in reference length roughly doubles the median GSD error.

Reading between the lines

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

  • The paper's reliance on a single calibrated mode suggests an obvious extension: maintain a small library of reference classes (vehicles, lane widths, shipping containers) and let the agent pick whichever anchor is present; this would lift the 33% coverage gap in vehicle-free scenes.
  • If true, the result reframes scale recovery as a statistical estimation problem rather than a learned one, and implies that future embodied agents should route metric-critical perception through handcrafted measurement channels, not through the language model's visual intuition.
  • A testable extension is to measure how the KDE mode behaves in scenes with bimodal vehicle fleets (e.g., a mix of compact sedans and delivery vans); the confidence score's distribution-concentration term would presumably degrade, but the paper does not investigate this failure mode.
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 / 5 minor

Summary. The paper proposes VehAnchor/VANGUARD, an inference-time GSD estimator that detects small vehicles with an oriented YOLO detector, filters outliers, estimates the modal pixel length via KDE, and converts it to GSD using a pre-calibrated reference length L_ref = 5.045 m (GSD = L_ref / P_mode). On the DOTA v1.5 validation set, the end-to-end pipeline reports a 6.87% median GSD error on 306 images (67% coverage), with an ablation showing KDE outperforms mean/median aggregation and with sensitivity analysis for L_ref, the outlier factor, vehicle count, and GSD range. The paper also introduces a self-constructed RS-GSD benchmark of 100 area-estimation entries and reports that five VLMs have 38–52% median area error, while the proposed pipeline achieves 19.7%, with 2.6× lower category dependence and 4× fewer catastrophic failures.

Significance. If the reported results hold beyond the evaluation distribution, the paper makes a useful and practical contribution: a simple, deterministic, explainable geometric skill that an LLM/VLM planner can call to obtain metric scale, with a confidence score and resolution guard. The KDE-based mode estimator and the explicit comparison against VLM spatial scale hallucination are valuable for the tool-augmented agent paradigm. The paper is also transparent in providing a sensitivity analysis and discussing limitations. However, the central quantitative claim is currently demonstrated only in-distribution on DOTA v1.5, and the VLM comparison relies on an unreleased, small benchmark with no repeated-run statistics. The degree of significance therefore depends on whether the authors can support the transferability implied by 'metadata-free' or appropriately restrict their claims.

major comments (3)
  1. [Section III-B, Eq. (2); Section IV-B, Eq. (6)] The headline 6.87% median error is an in-distribution, calibration-conditional result. L_ref is the KDE mode of physical vehicle lengths computed from DOTA v1.5 training annotations, and the evaluation is on the validation split of the same dataset (primarily Chinese urban scenes). Table III shows that a ±0.5 m change in L_ref roughly doubles the median error (10.59%/10.71% vs. 6.88%), and Section V-b concedes that L_ref is calibrated for Chinese urban fleets. Thus GSD_pred = L_ref / P_mode (Eq. 6) actually measures how well the DOTA fleet prior matches the evaluation fleet; it does not establish metadata-free generality. Please either add a cross-dataset evaluation (e.g., VisDrone, xView, or another geographic region) and report the sensitivity of the result to L_ref uncertainty, or restrict the title/abstract claims to 'calibrated prior' rather than 'metadata-free' recovery.
  2. [Section IV-D, Table II] The VLM comparison is not yet sufficiently rigorous to support the strong 'spatial scale hallucination' claim. The RS-GSD benchmark is self-constructed and unreleased, and VLM API outputs are stochastic; no repeated runs, confidence intervals, or statistical tests are reported. The '4× fewer catastrophic failures' and '2.6× lower category dependence' figures are based on single-run percentages. In addition, Claude Opus 4.6 + hint is an agent with programmatic image analysis tools (footnote †), which is not a pure visual estimator and should be analyzed separately or removed from the direct comparison. Please release the benchmark and prompt templates, report repeated-run medians, and give a significance test (e.g., paired bootstrap) for the VLM vs. pipeline differences.
  3. [Section IV-A; Section IV-B] The coverage/evaluation protocol is partially unclear: 306 E2E images vs. 269 GT-baseline images are evaluated on different subsets, and the headline 6.87% is for the 306-image subset while the 'intersection' comparison uses 267 images. The manuscript should state explicitly why the E2E set is larger than the GT set and report the median/mean errors on the same 267-image intersection for the headline number as well. This is important because the E2E result is the central claim, and readers need a common subset on which both the detector and the GT pipeline are compared.
minor comments (5)
  1. [Title/Abstract] The arXiv title uses 'VehAnchor' while the full text uses 'VANGUARD'. Please unify the name throughout.
  2. [General formatting] The submitted PDF contains numerous font-encoding artifacts (e.g., '/uni00000022' sequences) that obscure parts of the text. The final version should be regenerated so that all mathematical notation and references are readable.
  3. [Table II] The row label 'RS-GSD (ours)' is confusing because RS-GSD is also the name of the benchmark. Rename the row to 'VehAnchor + SAM (proposed)'.
  4. [Equation (1)] In Eq. (1), the notation P(gt)_i and GSD(gt)_i is not fully defined. State that P(gt)_i is the longer side of the ground-truth oriented bounding box in pixels and GSD(gt)_i is the ground-truth GSD of the image containing the i-th annotation.
  5. [Section IV-D] Describe the construction of the RS-GSD benchmark more completely: how the 100 entries were sampled from the 64 images, how ground-truth areas were computed, what prompt was used for each VLM, and whether any calibration or few-shot examples were provided. Also include a statement about data release.

Circularity Check

1 steps flagged · score 4.0 of 10

Reported 6.87% GSD error is an in-corpus check of the L_ref prior calibrated from DOTA v1.5 training, not an out-of-distribution validation of metadata-free scale recovery.

  1. fitted input called prediction [Section III-B Eq. (2); Section III-F Eq. (6); Section IV-B]
    "Rather than selecting an arbitrary value, we derive L_ref statistically from the DOTA v1.5 training set ... Lref = arg max_x \sum K_h(x-L_i) = 5.045m ... GSD_pred = L_ref / P_mode (6) ... End-to-end, replacing GT annotations with YOLO detections, the pipeline achieves 6.87% median error on 306 images."

    L_ref, the only scene-independent parameter in the GSD formula, is the KDE mode of physical vehicle lengths computed from DOTA v1.5 training annotations with known GSD (Eqs. 1-2). The headline accuracy is then measured on DOTA v1.5 validation images from the same corpus, using GSD_pred = L_ref / P_mode. The reported error therefore mostly reduces to how closely the validation fleet's modal length matches the training-fitted L_ref. The paper's own sensitivity analysis (L_ref=4.5m -> 10.59%, L_ref=5.5m -> 10.71%) and limitation 'L_ref is calibrated for Chinese urban fleets' confirm the result is a property of that fitted prior rather than a general metadata-free recovery. The prediction is not a tautology since P_mode is image-specific, but the central 6.87% claim is an in-distribution consi

full rationale

The load-bearing step is the calibration-evaluation overlap: L_ref is fit to DOTA v1.5 training data and the central accuracy claim is reported on DOTA v1.5 validation. Because GSD_pred = L_ref / P_mode and the validation images come from the same dataset/fleet distribution, the median error chiefly measures whether the validation vehicle-length mode equals the training-derived 5.045m. This is a fitted-input-called-prediction pattern at the corpus level. However, the paper explicitly acknowledges the dataset specificity in Limitations, provides a sensitivity analysis, and the method has independent content: KDE mode estimation, OBB detection, confidence gating, and a comparative VLM hallucination benchmark. There is no self-citation chain, no imported uniqueness theorem, and no ansatz smuggled via citation. The VLM comparison and the GT-baseline upper bound are separate empirical contributions. The circularity is partial and localized to the generality claim of the headline 6.87% number, not to the internal correctness of the geometric pipeline. Score 4 reflects this moderate, partly-acknowledged circularity rather than full tautology.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central free parameter is L_ref = 5.045 m, fit to DOTA v1.5 training annotations; alpha, GSD_max/P_thresh, the detector confidence cutoff, and the confidence-score weights are additional hand-set constants. No new physical entities are introduced.

free parameters (5)
  • L_ref = 5.045 m
    Calibrated as the KDE mode of physical vehicle lengths from DOTA v1.5 training annotations (Eq. 2); directly scales GSD_pred via Eq. (6).
  • alpha (outlier factor) = 1.5
    Hand-chosen in Eq. (3); sensitivity analysis shows 1.0 gives 7.27% median error, 'none' gives 6.74%, 1.5 selected as a compromise.
  • GSD_max / P_thresh = 0.3 m/px / ~17 px
    Manually chosen resolution guard threshold in Eq. (7), defining when the composite confidence score is capped.
  • Detector confidence cutoff = C_i < 0.1 discarded
    Chosen value in Section III-C to filter low-confidence detections.
  • Confidence-score weights = 0.35, 0.35, 0.20, 0.10
    Hand-selected weights for sample sufficiency, distribution concentration, detection quality, and anomaly detection; not ablated.
assumptions (4)
  • domain assumption Small vehicles have physical lengths concentrated around 4–5 m worldwide, so a single modal reference length suffices.
    Used to justify L_ref and the GSD conversion in Eq. (6); the paper itself qualifies this as dataset-specific.
  • domain assumption DOTA v1.5 is a valid proxy for high-altitude UAV nadir views with precise GSD metadata.
    Invoked in Section IV-A to justify evaluation on satellite imagery.
  • domain assumption Imagery is near-nadir/orthographic so that GSD is constant and OBB pixel length maps linearly to physical length via GSD.
    Required for Eq. (1) and Eq. (6); oblique/nadir distortion is listed as a limitation in Section V-b.
  • standard math KDE with Gaussian kernel and Scott's rule bandwidth gives a reliable mode estimate of the pixel-length distribution.
    Standard nonparametric density estimation (Rosenblatt/Parzen/Silverman), used in Eq. (4)-(5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of VehAnchor: Metadata-Free Metric Scale Recovery from Vehicle Cues in Aerial Imagery." pith.science (2026). https://pith.science/paper/VS6Z7TVS

@misc{pith2026260304277,
  author       = {Pith},
  title        = {Pith review of: VehAnchor: Metadata-Free Metric Scale Recovery from Vehicle Cues in Aerial Imagery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VS6Z7TVS}},
  note         = {Machine review of arXiv:2603.04277}
}
abstract

Autonomous aerial robots operating in GPS-denied or communication-degraded environments frequently lose access to camera metadata and telemetry, leaving onboard perception systems unable to recover the absolute metric scale of the scene. As LLM/VLM-based planners are increasingly adopted as high-level agents for embodied systems, their ability to reason about physical dimensions becomes safety-critical -- yet our experiments show that five state-of-the-art VLMs suffer from \emph{spatial scale hallucinations}, with median area estimation errors exceeding 50\%. We propose VehAnchor, a lightweight, deterministic \emph{Geometric Perception Skill} designed as a callable tool that any LLM-based agent can invoke to recover Ground Sample Distance (GSD) from ubiquitous environmental anchors: small vehicles detected via oriented bounding boxes, whose modal pixel length is robustly estimated through kernel density estimation and converted to GSD using a pre-calibrated reference length. The tool returns both a GSD estimate and a composite confidence score, enabling the calling agent to autonomously decide whether to trust the measurement or fall back to alternative strategies. On the DOTA~v1.5 benchmark, VehAnchor achieves 6.87\% median GSD error on 306~images. Integrated with SAM-based segmentation for downstream area measurement, the pipeline yields 19.7\% median error on a 100-entry benchmark -- with 2.6$\times$ lower category dependence and 4$\times$ fewer catastrophic failures than the best VLM baseline -- demonstrating that equipping agents with deterministic geometric tools is essential for safe autonomous spatial reasoning.

Figures

Figures reproduced from arXiv: 2603.04277 by the authors.

Figure 1
Figure 1. Our Geometric Perception Skill in action on a DOTA v1.5 image (P2593, Ground Track Field). (a) A UAV captures an image without metadata; [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. System architecture as an Embodied Agent Perception Loop. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. End-to-end GSD estimation results (YOLO11 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on three benchmark entries. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 4 linked inside Pith

  1. [1]

    SpatialVLM: Endowing vision-language models with spatial reasoning capabilities,

    B. Chen, Z. Xu, S. Kirmani, B. Ichter, D. Sadigh, L. Guibas, and F. Xia, “SpatialVLM: Endowing vision-language models with spatial reasoning capabilities,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 14 455–14 465

    2024

  2. [2]

    P. R. Wolf, B. A. DeWitt, and B. E. Wilkinson,Elements of Pho- togrammetry with Applications in GIS, 4th ed. New York: McGraw- Hill Education, 2014

    2014

  3. [3]

    Review article: Geometric processing of remote sensing images: models, algorithms and methods,

    T. Toutin, “Review article: Geometric processing of remote sensing images: models, algorithms and methods,”International Journal of Remote Sensing, vol. 25, no. 10, pp. 1893–1924, 2004

    1924

  4. [4]

    Satellite images analysis for shadow detection and building height estimation,

    G. Liasis and S. Stavrou, “Satellite images analysis for shadow detection and building height estimation,”ISPRS Journal of Pho- togrammetry and Remote Sensing, vol. 119, pp. 437–450, 2016

    2016

  5. [5]

    Towards robust monocular depth estimation: Mixing datasets for zero- shot cross-dataset transfer,

    R. Ranftl, K. Lasinger, D. Hafner, K. Schindler, and V . Koltun, “Towards robust monocular depth estimation: Mixing datasets for zero- shot cross-dataset transfer,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 3, pp. 1623–1637, 2022

    2022

  6. [6]

    Zoedepth: Zero-shot transfer by combining relative and metric depth,

    S. F. Bhat, R. Birkl, D. Wofk, P. Wonka, and M. M ¨uller, “Zoedepth: Zero-shot transfer by combining relative and metric depth,”arXiv preprint arXiv:2302.12288, 2023

    arXiv 2023

  7. [7]

    Regression tree CNN for estimation of ground sampling distance based on floating-point representation,

    J. H. Lee and S. Sull, “Regression tree CNN for estimation of ground sampling distance based on floating-point representation,”Remote Sensing, vol. 11, no. 19, p. 2276, 2019

    2019

  8. [8]

    Visual agentic system for spatial metric query answering in remote sensing images,

    Y . Wang and C. Wang, “Visual agentic system for spatial metric query answering in remote sensing images,” inEurographics 2025. The Eurographics Association, 2025

    2025

Show all 25 references
  1. [9]

    Remarks on some nonparametric estimates of a density function,

    M. Rosenblatt, “Remarks on some nonparametric estimates of a density function,”The Annals of Mathematical Statistics, vol. 27, no. 3, pp. 832–837, 1956

    1956

  2. [10]

    On estimation of a probability density function and mode,

    E. Parzen, “On estimation of a probability density function and mode,” The Annals of Mathematical Statistics, vol. 33, no. 3, pp. 1065–1076, 1962

    1962

  3. [11]

    B. W. Silverman,Density Estimation for Statistics and Data Analysis. London: Chapman and Hall, 1986

    1986

  4. [12]

    GeoChat: Grounded large vision-language model for remote sensing,

    K. Kuckreja, M. S. Danish, M. Naseer, A. Das, S. Khan, and F. S. Khan, “GeoChat: Grounded large vision-language model for remote sensing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 27 831–27 841

    2024

  5. [13]

    Vision-language models in remote sensing: Current progress and future trends,

    X. Li, C. Wen, Y . Hu, Z. Yuan, and X. X. Zhu, “Vision-language models in remote sensing: Current progress and future trends,”IEEE Geoscience and Remote Sensing Magazine, vol. 12, no. 2, pp. 56–86, 2024

    2024

  6. [14]

    RemoteCLIP: A vision language foundation model for remote sens- ing,

    F. Liu, D. Chen, Z. Guan, X. Zhou, J. Zhu, Q. Ye, L. Fu, and J. Zhou, “RemoteCLIP: A vision language foundation model for remote sens- ing,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–13, 2024

    2024

  7. [15]

    Visual spatial reasoning,

    F. Liu, G. Emerson, and N. Collier, “Visual spatial reasoning,” Transactions of the Association for Computational Linguistics, vol. 11, pp. 635–651, 2023

    2023

  8. [16]

    Reasoning paths with reference objects elicit quantitative spatial reasoning in large vision-language models,

    Y .-H. Liao, R. Mahmood, S. Fidler, and D. Acuna, “Reasoning paths with reference objects elicit quantitative spatial reasoning in large vision-language models,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Association for Computati...

    2024

  9. [17]

    DOTA: A large-scale dataset for object detection in aerial images,

    G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang, “DOTA: A large-scale dataset for object detection in aerial images,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 3974–3983

    2018

  10. [18]

    Object detection in aerial images: A large-scale benchmark and challenges,

    J. Ding, N. Xue, G.-S. Xia, X. Bai, W. Yang, M. Y . Yang, S. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang, “Object detection in aerial images: A large-scale benchmark and challenges,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 7...

    2022

  11. [19]

    Ultralytics YOLO11,

    G. Jocher and J. Qiu, “Ultralytics YOLO11,” https://github.com/ ultralytics/ultralytics, 2024, version 11.0.0

    2024

  12. [20]

    iSAID: A large-scale dataset for instance segmentation in aerial images,

    S. Waqas Zamir, A. Arora, A. Gupta, S. Khan, G. Sun, F. S. Shah, F. Zhu, L. Shao, G.-S. Xia, and X. Bai, “iSAID: A large-scale dataset for instance segmentation in aerial images,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CV...

    2019

  13. [21]

    Qwen2.5-VL technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tanget al., “Qwen2.5-VL technical report,”arXiv preprint arXiv:2502.13923, 2025

    2025 arXiv

  14. [22]

    Claude: Model overview,

    Anthropic, “Claude: Model overview,” https://docs.anthropic.com/en/ docs/about-claude/models, 2025, claude Opus 4.6 (model ID: claude- opus-4-6), accessed February 2026

    2025

  15. [23]

    Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution,

    P. Wang, S. Bai, S. Tan, S. Wang, Z. Fan, J. Bai, K. Chen, X. Liu, J. Wang, W. Geet al., “Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution,”arXiv preprint arXiv:2409.12191, 2024

    2024 arXiv

  16. [24]

    GPT-4osystem card,

    OpenAI, “GPT-4osystem card,”arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  17. [25]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 4015–4026

    2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.