Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Synthetic aerial images can diagnose detector weaknesses and guide efficient real-data collection.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 07:31 UTC pith:5H26PASG

load-bearing objection Solid empirical diagnostic loop for aerial detectors; the synthetic-to-real ranking transfer is only indirectly validated, but the matched-budget gains and multi-model consistency still make the paper worth engaging. the 3 major comments →

arxiv 2607.02718 v1 pith:5H26PASG submitted 2026-07-02 cs.CV cs.AIcs.LG

Diagnosing Aerial-View Object Detectors with Foundational Image Generative Models

Stanislav Panev , Minhyek Jeon , Vaishnavi Khindkar , Ahish Deshpande , Celso M de Melo , Shuowen Hu , Shayok Chakraborty , Fernando De la Torre This is my paper
classification cs.CV cs.AIcs.LG
keywords aerial object detectionremote sensing robustnesssynthetic data evaluationmodel diagnosisfoundational generative modelstargeted data supplementation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Aerial vehicle detectors fail in ways that real benchmarks obscure, because scene types and weather are entangled and unevenly sampled. This paper shows that foundation image generators can build a controlled synthetic testbed in which attributes such as scene type, season, and weather are isolated and verified. When three standard detectors trained on three real aerial datasets are evaluated on that testbed, the scene-wise performance ranking closely matches the weaknesses they later show on real imagery. Using those rankings to collect only a few thousand extra real images from the weakest categories (urban, industrial, desert) raises AP50 by as much as 13 percent, far more efficiently than adding the same number of randomly chosen images. The result is a practical diagnostic loop: generate controllable synthetic scenes, measure relative failure modes, then acquire the smallest real set that repairs them.

Core claim

Controlled synthetic probing of aerial-view vehicle detectors produces scene-wise performance trends that reliably predict real-world weaknesses; targeting real-data supplementation at the categories flagged by the synthetic testbed yields AP50 gains of up to 13 percent with an order-of-magnitude fewer extra images than non-targeted augmentation.

What carries the argument

A modular synthetic diagnostic framework that (1) samples an attribute taxonomy, (2) generates and edits photorealistic nadir images with a foundation text-to-image model, (3) verifies attributes with a multimodal model, and (4) evaluates pretrained detectors on the resulting attribute-partitioned test set.

Load-bearing premise

Performance rankings measured on the synthetic images must correctly order the same detectors' relative difficulty on real aerial scenes of matching type; residual generation artifacts must not reorder those difficulties.

What would settle it

Retrain the same three detectors on the original real sets, evaluate them on held-out real urban, industrial and desert imagery never used in the paper, and check whether the synthetic ranking still correctly identifies the worst real categories; if it does not, the transfer claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a modular synthetic diagnostic framework (AVODDiag) that uses foundation generative models (primarily Imagen 3) with text-guided generation, attribute-controlled editing, and MLLM-based attribute verification to build a controllable aerial-view testbed for vehicle detectors. Three architectures (Faster R-CNN, YOLOv8-M, ViTDet-B) trained on DOTAv2, LINZ, and UGRC are evaluated scene-wise on the synthetic set; urban and industrial scenes emerge as recurrent weak categories. Guided by these rankings, the authors add small real single-scene sets (Miami urban, LA industrial, Phoenix desert) and report AP50 gains of up to ~13% on the synthetic testbed, with a matched-budget targeted-vs-random ablation (Fig. 9) showing that non-targeted data is less effective or harmful. The framework is presented as model-agnostic and as a guide for efficient real-data collection rather than as a new detector.

Significance. If the synthetic-to-real ranking transfer holds, the work offers a practical, attribute-controllable alternative to expensive balanced aerial collection and goes beyond post-hoc error taxonomies (TIDE) and corruption benchmarks by isolating scene factors. Strengths include the multi-architecture, multi-dataset design (nine combinations), the explicit targeted-vs-random control under a fixed data budget, public code and datasets, and a semi-automatic annotation pipeline that substantially reduces labeling cost. The modular framing and honest discussion of closed-source model dependence and open-source limitations are also valuable for remote-sensing robustness research. The result is significant for Earth-observation pipelines where balanced real data are costly, provided the diagnostic ranking is shown to track real difficulty rather than only synthetic difficulty.

major comments (3)
  1. The abstract and Sec. 4.4 claim that “synthetic scene-wise performance trends closely match real-world weaknesses,” but the manuscript never reports scene-wise AP of the nine original model–dataset combinations on real urban/industrial/desert tiles. Weak categories are identified solely from synthetic evaluation (Fig. 7); the Miami/LA/Phoenix sets are used only for supplementation and post-supplementation gains (Tab. 3, Fig. 8). Without a direct real-scene ranking table (or at least pre-supplementation AP on those real sets), the gains are consistent with “extra real data from underrepresented geographies help” and do not establish that the synthetic ranking correctly ordered real difficulty. Residual generation artifacts could reorder relative hardness undetected. A minimal fix is to evaluate the original detectors on the three real single-scene sets before supplementation and report wh
  2. Sec. 4.4 / Tab. 3 and Fig. 8 report large AP50 gains (up to +13.4%) after targeted supplementation, but all gains are measured on the same synthetic diagnostic set used to select the weak categories. There is no corresponding evaluation of the supplemented models on held-out real multi-environment test splits (e.g., LINZ/UGRC test or a held-out real urban/industrial/desert partition). Without real-domain transfer numbers, it remains unclear whether the diagnostic loop improves real deployment performance or only synthetic-test performance. Reporting real test AP before/after supplementation for at least a subset of the nine combinations would make the central “guide efficient data collection” claim load-bearing rather than suggestive.
  3. Tab. 2: human approval of pseudo-boxes is only 58.5% on the synthetic set (and 48.3% on LA-Industrial). The pipeline merges Moondream 2 and Gemini proposals then relies on binary human filtering. Given that diagnosis and all reported APs rest on these labels, the manuscript should quantify residual label noise (e.g., inter-annotator agreement on a subsample, or AP sensitivity under stricter/looser approval) and discuss whether systematic false negatives in dense urban/industrial scenes could themselves drive the “weak category” ranking. This is especially important because the paper already notes that foundation VLMs are insufficient for reliable zero-shot aerial vehicle detection.
minor comments (5)
  1. Fig. 7 and Fig. 8 use orange arrows/boxes effectively, but axis scales differ across panels, making cross-model comparison harder; consider a shared color scale or a single summary ranking table.
  2. Supplementary Fig. 15 shows weather-attribute std. dev. nearly doubling after editing; the main text (Sec. 3.1) claims enrichment reduces imbalance—briefly note the weather exception and its implications for weather-conditioned diagnosis.
  3. The fixed 42.36 px square boxes (Supp. A.2) and exclusive use of AP50 are reasonable for center-based LINZ/UGRC labels, but should be stated once in the main evaluation protocol (Sec. 4.1) so readers do not assume standard multi-scale box evaluation.
  4. Attribute-validation threshold τ (Alg. 1) is free but never given a numerical value or sensitivity check; a short note or appendix ablation would help reproducibility.
  5. Typos/clarity: “DataEnrichmentThroughImageEditing” spacing in Sec. 3.1; “Init. Training Dataset” abbreviations in figures could be expanded once; arXiv ID and code URL are fine but ensure the camera-ready link is stable.

Circularity Check

0 steps flagged

Empirical diagnostic loop with no derivation that reduces to its inputs by construction

full rationale

This is an applied empirical CV paper, not a first-principles derivation. Detectors are trained exclusively on real data (DOTAv2/LINZ/UGRC), evaluated on a held-out synthetic attribute-controlled testbed, and then re-trained with additional real images selected by the synthetic ranking; the reported AP50 gains and targeted-vs-random comparison are measured after the fact on real data. No equation, fitted parameter, or uniqueness theorem forces the synthetic scene-wise ranking to equal any real ranking, nor is any quantity redefined as its own input. Self-citations (e.g., Fang et al. for the LINZ/UGRC datasets) supply data sources and are not load-bearing premises that close a logical loop. Attribute extraction and human-in-the-loop box approval are labeling steps, not predictions of detector performance. The transferability claim is therefore an empirical hypothesis tested by supplementation experiments, not a circular identity. Score 0 is the correct non-finding.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 1 invented entities

The central claim rests on a small set of modeling choices (taxonomy, commercial generators, similarity threshold for attribute validation) and on the empirical transferability of synthetic rankings to real data. No free parameters are fitted to the final AP numbers; the free parameters listed below are pipeline hyper-parameters chosen once. The invented entity is the diagnostic framework itself.

free parameters (3)
  • attribute-validation cosine threshold τ
    Used in Algorithm 1 to decide whether an MLLM-extracted attribute string is remapped to an existing taxonomy leaf or added as a new leaf; value not reported numerically but controls label consistency.
  • square bounding-box side length 42.36 px
    Fixed conversion from center-point annotations (LINZ/UGRC) to boxes so that AP50 remains meaningful; chosen from average vehicle size at 12.5 cm GSD.
  • synthetic image count and attribute sampling schedule
    5 453 images covering 304 attribute combinations; the uniform-then-complementary sampling schedule is a design choice that affects coverage of rare scenes.
axioms (3)
  • domain assumption Foundation text-to-image models (Imagen 3) can synthesize photorealistic nadir aerial scenes whose visual statistics are sufficiently close to real aerial imagery for relative detector rankings to transfer.
    Stated as the enabling premise of the whole diagnostic framework (Introduction and Sec. 3.1); never formally proved, only supported by the later empirical match.
  • ad hoc to paper The chosen three-level attribute taxonomy (scene type / season / weather + object count/color/type) captures the dominant axes of domain shift that matter for aerial vehicle detection.
    Taxonomy is hand-designed with ChatGPT assistance and refined by the authors (Sec. 3.1 and supplementary); alternative taxonomies are not ablated.
  • domain assumption Zero-shot MLLM box proposals plus binary human approval produce labels of sufficient quality for reliable AP50 evaluation.
    Sec. 3.1 and Table 2; approval rates of 48–69 % indicate residual noise that is accepted as tolerable.
invented entities (1)
  • AVODDiag synthetic diagnostic framework (text-guided generation + attribute editing + automated verification + scene-conditioned evaluation) independent evidence
    purpose: To isolate individual scene attributes and produce a controllable testbed that real aerial benchmarks cannot provide.
    The modular pipeline and the specific taxonomy-plus-editing procedure are introduced by the paper; independent evidence is the released code and the empirical transfer results.

pith-pipeline@v1.1.0-grok45 · 24736 in / 2703 out tokens · 34159 ms · 2026-07-12T07:31:25.957493+00:00 · methodology

0 comments
read the original abstract

Recent advances in large-scale image generative models enable photorealistic scene synthesis with controllable attributes. Beyond data augmentation, their potential as diagnostic tools for trained vision systems remains unexplored in the aerial and remote sensing domains. We introduce a synthetic diagnostic framework for aerial-view vehicle detection that combines text-guided generation, attribute-controlled editing, and automated attribute verification to construct a controllable synthetic testbed. This enables fine-grained evaluation of pretrained detectors under diverse scene types and environmental conditions that are difficult to isolate in real datasets. Across three detection architectures and three real aerial datasets, synthetic scene-wise performance trends closely match real-world weaknesses. Guided by these diagnostics, targeted supplementation with small real datasets from the identified weak categories yields improvements of up to 13% AP50 while requiring substantially fewer additional samples than non-targeted augmentation. Our results show that controlled synthetic probing can predict real-domain performance gaps and guide efficient data collection. The proposed diagnostic framework is modular and can incorporate alternative generative or vision-language models as capabilities evolve. Our code and datasets are available here: https://humansensinglab.github.io/AVODDiag/

Figures

Figures reproduced from arXiv: 2607.02718 by Ahish Deshpande, Celso M de Melo, Fernando De la Torre, Minhyek Jeon, Shayok Chakraborty, Shuowen Hu, Stanislav Panev, Vaishnavi Khindkar.

Figure 1
Figure 1. Figure 1: We leverage foundation generative models to synthesize diverse, attribute￾controlled aerial-view image datasets, thereby enabling systematic diagnosis of the weaknesses of aerial-view vehicle detectors. their potential as diagnostic tools for analyzing trained vision systems remains largely unexplored in remote sensing and aerial imagery. Aerial-view object de￾tection (AVOD) is a core component of Earth ob… view at source ↗
Figure 2
Figure 2. Figure 2: A showcase of synthetic aerial view imagery generated by Imagen 3. Each image represents a combination of a different environment and a season. We use these data to identify inherent weaknesses in object detectors. corresponding to the identified weak categories. Despite being at least an order of magnitude smaller than the original training sets, this targeted supplementation yields improvements of up to … view at source ↗
Figure 3
Figure 3. Figure 3: Attribute taxonomy tree and associated variables a (c) i that constitute the at￾tribute vector ai. Attribute Taxonomy The attribute taxonomy T ( [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the synthetic data generation pipeline. Left: Initial image gener￾ation using a text-to-image foundation model fG conditioned on prompts p (G) i derived from uniformly sampled attribute values in the taxonomy T . Right: Image editing via image-and-text-to-image generation to correct distributional biases and enrich under￾represented attribute values. with the text embeddings of all values C (c)… view at source ↗
Figure 5
Figure 5. Figure 5: Samples from the real datasets used in our experiments 1. A pretrained aerial-view vehicle detector is evaluated on the synthetic di￾agnostic dataset, and both overall and attribute-conditioned APs are com￾puted. 2. Attributes with the lowest AP are identified as underperforming categories. 3. Additional real training images corresponding to these categories are ob￾tained from public sources, labeled, and … view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of six text-to-image foundation generative models on the aerial-view image generation task. Each row is generated using the same prompt (see the supplementary material for more details). Open-source models: SD 1.5 [40], SDXL [33], and SD 3.5 [1]. Closed (commercial) models: DALL-E [28], Gemini 2.5 Flash Image [8], Imagen 3 [2]. (Zoom in to see the details better.) [PITH_FULL_IMAGE:f… view at source ↗
Figure 7
Figure 7. Figure 7: Scene-wise AP50 of three detectors trained on real aerial datasets (DOTAv2, LINZ, UGRC) and evaluated on the synthetic diagnostic set. Rows denote models; columns denote training datasets. Orange arrows highlight scene types below the overall AP (gray dashed line). Best viewed in color. ated on our synthetic diagnostic dataset to analyze performance across different scene types, one of the key image attrib… view at source ↗
Figure 8
Figure 8. Figure 8: AP50 gains per scene type provided by the data supplementation, described in Tab. 3. The values outlined in the orange box correspond to the scene types for which additional supplement data were used during supplementation. Best viewed in color. synthetic, attribute-controlled testbed, we identified systematic scene-dependent weaknesses of detectors trained on real datasets and showed that synthetic scene￾… view at source ↗
Figure 9
Figure 9. Figure 9: Targeted vs. random (non-targeted) supplementation AP gains More broadly, this study highlights controlled synthetic probing as a prac￾tical mechanism for improving remote sensing systems. Rather than replacing real data, synthetic diagnostics enable informed data acquisition by identifying where additional coverage is most impactful. In high-stakes Earth observation applications—where balanced data collec… view at source ↗
Figure 10
Figure 10. Figure 10: Urban and desert real supplementary datasets [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Los Angeles - Industrial dataset sampled from LARIAC6 source. The red squares depict the exact sampling locations of the 2,000 samples with size 48 m×48 m [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: An example of a LINZ dataset image with square bounding boxes for center￾based object detection. The red dots depict the annotated vehicle centers, and the blue circles depict the 12 px decision circles [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Scene types distributions of the three pre-training real aerial view datasets used in our experiments. urban 64.2% urban street 4.0% construction site 3.7% rooftop 3.7% urban rooftop 1.8% outdoor 1.7% Miami-Urban (Top 80%) industrial area 15.2% urban 12.4% industrial/ commercial area 7.4% rooftop 6.4% parking lot 5.9% industrial 5.9% urban street 3.5% industrial/ urban 3.4% industrial/ commercial rooftop … view at source ↗
Figure 14
Figure 14. Figure 14: Scene types distributions of the three full supplemental real aerial view datasets used in our experiments [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Image attribute distributions before (orange) and after (green) the image editing. The histogram bin standard deviations (before editing → after editing) are depicted under each plot’s title [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: illustrates four image-editing examples that successfully remove and add vehicles to the scene, change the weather conditions (from a sunny winter scene to a cloudy and snowy winter scene), and change the season (from fall to summer). no vehicles vehicles present sunny snowy fall summer vehicles present no vehicles [PITH_FULL_IMAGE:figures/full_fig_p027_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Image editing defects: failure to remove all vehicles (top left), failure to change the weather from sunny to cloudy (top right), failure to preserve the vehicle locations and colors while performing weather editing (bottom left), mixing two different camera views (bottom right) [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 6 linked inside Pith

  1. [1]

    Machine learning model (2024), available from https://huggingface.co/stabilityai/stable-diffusion-3.5-large 10, 11

    AI, S.: Stable diffusion 3.5 large. Machine learning model (2024), available from https://huggingface.co/stabilityai/stable-diffusion-3.5-large 10, 11

    2024

  2. [2]

    arXiv preprint arXiv:2408.07009 (2024) 1, 4, 10, 11

    Baldridge, J., Bauer, J., Bhutani, M., Brichtova, N., Bunner, A., Castrejon, L., Chan, K., Chen, Y., Dieleman, S., Du, Y., et al.: Imagen 3. arXiv preprint arXiv:2408.07009 (2024) 1, 4, 10, 11

    arXiv 2024

  3. [3]

    https://cdn

    Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee,J.,Guo,Y.,etal.:Improvingimagegenerationwithbettercaptions.Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf2(3), 8 (2023) 4

    2023

  4. [4]

    Springer (2020) 4 16 S

    Bolya, D., Foley, S., Hays, J., Hoffman, J.: Tide: A general toolbox for identifying objectdetectionerrors.In:EuropeanConferenceonComputerVision.pp.558–573. Springer (2020) 4 16 S. Panevet al

    2020

  5. [5]

    arXiv preprint arXiv:1906.07155 (2019) 10

    Chen, K., Wang, J., Pang, J., Cao, Y., Xiong, Y., Li, X., Sun, S., Feng, W., Liu, Z., Xu, J., Zhang, Z., Cheng, D., Zhu, C., Cheng, T., Zhao, Q., Li, B., Lu, X., Zhu, R., Wu, Y., Dai, J., Wang, J., Shi, J., Ouyang, W., Loy, C.C., Lin, D.: MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155 (2019) 10

    Pith/arXiv arXiv 1906

  6. [6]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Chen, Y., Li, W., Sakaridis, C., Dai, D., Van Gool, L.: Domain adaptive faster r-cnn for object detection in the wild. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3339–3348 (2018) 5

    2018

  7. [7]

    Ad- vances in Neural Information Processing Systems36, 58921–58937 (2023) 4

    Clark, K., Jaini, P.: Text-to-image diffusion models are zero shot classifiers. Ad- vances in Neural Information Processing Systems36, 58921–58937 (2023) 4

    2023

  8. [8]

    arXiv preprint arXiv:2507.06261 (2025) 1, 10, 11

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025) 1, 10, 11

    Pith/arXiv arXiv 2025

  9. [9]

    https://github.com/open-mmlab/mmyolo(2022) 2, 10

    Contributors, M.: MMYOLO: OpenMMLab YOLO series toolbox and benchmark. https://github.com/open-mmlab/mmyolo(2022) 2, 10

    2022

  10. [10]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Deng, J., Li, W., Chen, Y., Duan, L.: Unbiased mean teacher for cross-domain object detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4091–4101 (2021) 5

    2021

  11. [11]

    IEEE Transactions on Pattern Analysis & Machine Intelligence44(11), 7778–7796 (Nov 2022).https://doi.org/10.1109/TPAMI

    Ding, J., Xue, N., Xia, G.S., Bai, X., Yang, W., Yang, M.Y., Belongie, S., Luo, J., Datcu,M.,Pelillo,M.,Zhang,L.: ObjectDetectioninAerialImages:ALarge-Scale Benchmark and Challenges . IEEE Transactions on Pattern Analysis & Machine Intelligence44(11), 7778–7796 (Nov 2022).https://doi.org/10.1109/TPAMI. 2021.3117983,https://doi.ieeecomputersociety.org/10.1...

    doi:10.1109/tpami 2022

  12. [12]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Fang, X., Jeon, M., Qin, Z., Panev, S., De Melo, C., Hu, S., Chakraborty, S., De la Torre, F.: Adapting vehicle detectors for aerial imagery to unseen domains with weak supervision. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8088–8099 (2025) 2, 5, 9

    2025

  13. [13]

    arXiv preprint arXiv:2509.25026 (2025) 8

    Fiaz, M., Debary, H., Fraccaro, P., Paudel, D., Van Gool, L., Khan, F., Khan, S.: Geovlm-r1: Reinforcement fine-tuning for improved remote sensing reasoning. arXiv preprint arXiv:2509.25026 (2025) 8

    arXiv 2025

  14. [14]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Gaidon, A., Wang, Q., Cabon, Y., Vig, E.: Virtual worlds as proxy for multi-object tracking analysis. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4340–4349 (2016) 4

    2016

  15. [15]

    Proceedings of the International Conference on Learning Representations (2019) 4

    Hendrycks,D.,Dietterich,T.:Benchmarkingneuralnetworkrobustnesstocommon corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019) 4

    2019

  16. [16]

    In: European conference on computer vision

    Hoiem, D., Chodpathumwan, Y., Dai, Q.: Diagnosing error in object detectors. In: European conference on computer vision. pp. 340–353. Springer (2012) 4

    2012

  17. [17]

    Hu, A., Russell, L., Yeo, H., Murez, Z., Fedoseev, G., Kendall, A., Shotton, J., Corrado, G.: Gaia-1: A generative world model for autonomous driving (2023), https://arxiv.org/abs/2309.170804

    Pith/arXiv arXiv 2023

  18. [18]

    Jocher, G., Chaurasia, A., Qiu, J.: Yolov8: Ultralytics real-time object detection (2023),https://github.com/ultralytics/ultralytics, version 8, available at https://github.com/ultralytics/ultralytics5, 10

    2023

  19. [19]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G., et al.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023) 4

    2023

  20. [20]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Li, A.C., Prabhudesai, M., Duggal, S., Brown, E., Pathak, D.: Your diffusion model is secretly a zero-shot classifier. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2206–2217 (2023) 4 Diagnosing Aerial-View Detectors using Generative Models 17

    2023

  21. [21]

    In: British Machine Vision Conference (BMVC) (2018) 4

    Li, W., Saeedi, S., McCormac, J., Clark, R., Tzoumanikas, D., Ye, Q., Huang, Y., Tang, R., Leutenegger, S.: Interiornet: Mega-scale multi-sensor photo-realistic indoor scenes dataset. In: British Machine Vision Conference (BMVC) (2018) 4

    2018

  22. [22]

    In: European conference on computer vision

    Li, Y., Mao, H., Girshick, R., He, K.: Exploring plain vision transformer backbones for object detection. In: European conference on computer vision. pp. 280–296. Springer (2022) 2, 5, 10

    2022

  23. [23]

    In: European conference on computer vision

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740–755. Springer (2014) 13

    2014

  24. [24]

    In: The Twelfth International Conferenceon LearningRepresentations(2024),https://openreview.net/forum? id=TOWdQQgMJY6

    Liu, Q., Kortylewski, A., Bai, Y., Bai, S., Yuille, A.: Discovering Failure Modes of Text-guided Diffusion Models via Adversarial Search. In: The Twelfth International Conferenceon LearningRepresentations(2024),https://openreview.net/forum? id=TOWdQQgMJY6

    2024

  25. [25]

    Midjourney, Inc.: Midjourney.https://www.midjourney.com(2022), accessed: 2025-11-13 1

    2022

  26. [26]

    International Journal of Com- puter Vision126(9), 902–919 (2018) 4

    Müller, M., Casser, V., Lahoud, J., Smith, N., Ghanem, B.: Sim4cv: A photo- realistic simulator for computer vision applications. International Journal of Com- puter Vision126(9), 902–919 (2018) 4

    2018

  27. [27]

    NVIDIA, :, Agarwal, N., Ali, A., Bala, M., Balaji, Y., Barker, E., Cai, T., Chat- topadhyay, P., Chen, Y., Cui, Y., Ding, Y., Dworakowski, D., Fan, J., Fenzi, M., Ferroni, F., Fidler, S., Fox, D., Ge, S., Ge, Y., Gu, J., Gururani, S., He, E., Huang, J., Huffman, J., Jannaty, P., Jin, J., Kim, S.W., Klár, G., Lam, G., Lan, S., Leal- Taixe, L., Li, A., Li, ...

    Pith/arXiv arXiv 2025

  28. [28]

    OpenAI: Dall·e 3.https://platform.openai.com/docs/guides/images(2023), accessed: 2025-11-13 1, 10, 11

    2023

  29. [29]

    OpenAI: Gpt-image-1.https : / / platform . openai . com / docs / guides / images (2024), accessed: 2025-11-13 1

    2024

  30. [30]

    OpenAI: Gpt-5: Large multimodal model.https://openai.com/gpt- 5(2025), available athttps://openai.com/gpt-510

    2025

  31. [31]

    Paolo, F., Lin, T.t.T., Gupta, R., Goodman, B., Patel, N., Kuster, D., Kroodsma, D., Dunnmon, J.: xview3-sar: Detecting dark fishing activity using synthetic aper- tureradarimagery.AdvancesinNeuralInformationProcessingSystems35,37604– 37616 (2022) 5

    2022

  32. [32]

    In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T

    Papadopoulos, D.P., Clarke, A.D.F., Keller, F., Ferrari, V.: Training object class detectors from eye tracking data. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) Computer Vision – ECCV 2014. pp. 361–376. Springer International Publishing, Cham (2014) 8, 11

    2014

  33. [33]

    In: Kim, B., Yue, Y., Chaudhuri, S., Fragkiadaki, K., Khan, M., Sun, Y

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. In: Kim, B., Yue, Y., Chaudhuri, S., Fragkiadaki, K., Khan, M., Sun, Y. (eds.) International Conference on Learning Representations. vol. 2024, pp. 1862– 1874 (2024),https://proce...

    2024

  34. [34]

    ArXivabs/2209.14988(2022),https://api.semanticscholar.org/ CorpusID:2525960914

    Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. ArXivabs/2209.14988(2022),https://api.semanticscholar.org/ CorpusID:2525960914

    Pith/arXiv arXiv 2022

  35. [35]

    Drones9(8) (2025).https:// doi.org/10.3390/drones9080549,https://www.mdpi.com/2504-446X/9/8/549 8, 10

    Randieri, C., Ganesh, S.V., Raj, R.D.A., Yanamala, R.M.R., Pallakonda, A., Napoli, C.: Aerial autonomy under adversity: Advances in obstacle and aircraft detection techniques for unmanned aerial vehicles. Drones9(8) (2025).https:// doi.org/10.3390/drones9080549,https://www.mdpi.com/2504-446X/9/8/549 8, 10

    doi:10.3390/drones9080549 2025

  36. [36]

    In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)

    Reimers, N., Gurevych, I.: Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). p. 3982. Association for Com- putational Linguistics (2019) 10

    2019

  37. [37]

    Advances in neural information processing systems28(2015) 2, 10

    Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object de- tection with region proposal networks. Advances in neural information processing systems28(2015) 2, 10

    2015

  38. [38]

    why should i trust you?

    Ribeiro, M.T., Singh, S., Guestrin, C.: " why should i trust you?" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD interna- tional conference on knowledge discovery and data mining. pp. 1135–1144 (2016) 4

    2016

  39. [39]

    In: European conference on computer vision

    Richter, S.R., Vineet, V., Roth, S., Koltun, V.: Playing for data: Ground truth from computer games. In: European conference on computer vision. pp. 102–118. Springer (2016) 4

    2016

  40. [40]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022) 4, 10, 11

    2022

  41. [41]

    In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Ros, G., Sellart, L., Materzynska, J., Vazquez, D., Lopez, A.M.: The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3234–3243 (2016).https://doi.org/10.1109/CVPR.2016.3524

    doi:10.1109/cvpr.2016.3524 2016

  42. [42]

    In: Proceedings of the 36th International Conference on Neural Information Processing Systems

    Saharia, C., Chan, W., Saxena, S., Lit, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Mahdavi, S.S., Gontijo-Lopes, R., Salimans, T., Ho, J., Fleet, D.J.,Norouzi,M.:Photorealistictext-to-imagediffusionmodelswithdeeplanguage understanding. In: Proceedings of the 36th International Conference on Neural Information Processing Systems. NIPS ’22,...

    2022

  43. [43]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) 5

    Saito, K., Ushiku, Y., Harada, T., Saenko, K.: Strong-weak distribution alignment for adaptive object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019) 5

    2019

  44. [44]

    International Journal of Computer Vision128(2), 336–359 (Oct 2019)

    Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-cam: Visual explanations from deep networks via gradient-based localiza- tion. International Journal of Computer Vision128(2), 336–359 (Oct 2019). https://doi.org/10.1007/s11263-019-01228-7,http://dx.doi.org/10.1007/ s11263-019-01228-74

    doi:10.1007/s11263-019-01228-7 2019

  45. [45]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Shermeyer, J., Hossler, T., Van Etten, A., Hogan, D., Lewis, R., Kim, D.: Rareplanes: Synthetic data takes flight. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 207–217 (2021) 5

    2021

  46. [46]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Soni, S., Dudhane, A., Debary, H., Fiaz, M., Munir, M.A., Danish, M.S., Fraccaro, P., Watson, C.D., Klein, L.J., Khan, F.S., et al.: Earthdial: Turning multi-sensory Diagnosing Aerial-View Detectors using Generative Models 19 earth observations to interactive dialogues. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 14303–1...

    2025

  47. [47]

    In: Human Computation - Papers from the 2012 AAAI Workshop, Technical Re- port

    Su, H., Deng, J., Fei-Fei, L.: Crowdsourcing annotations for visual object detection. In: Human Computation - Papers from the 2012 AAAI Workshop, Technical Re- port. pp. 40–46. AAAI Workshop - Technical Report (2012), 2012 AAAI Workshop ; Conference date: 23-07-2012 Through 23-07-2012 8, 11

    2012

  48. [48]

    2025-06-21 (2024),https://huggingface.co/ vikhyatk/moondream2, hugging Face

    Vikhyat, K.: Moondream 2, rev. 2025-06-21 (2024),https://huggingface.co/ vikhyatk/moondream2, hugging Face. DOI: 10.57967/hf/6762 10

    doi:10.57967/hf/6762 2025

  49. [49]

    Wang, R., Chen, Z., Chen, C., Ma, J., Lu, H., Lin, X.: Compositional text-to-image synthesis with attention map control of diffusion models. In: Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Con- ference on Innovative Applications of Artificial Intelligence and Fourteenth Sym- posium on Educational Advances i...

    2024

  50. [50]

    Advances in Neural Information Processing Systems36, 54683–54695 (2023) 4

    Wu, W., Zhao, Y., Chen, H., Gu, Y., Zhao, R., He, Y., Zhou, H., Shou, M.Z., Shen, C.: Datasetdm: Synthesizing data with perception annotations using diffu- sion models. Advances in Neural Information Processing Systems36, 54683–54695 (2023) 4

    2023

  51. [51]

    Zhang, H., Li, F., Liu, S., Zhang, L., Su, H., Zhu, J., Ni, L.M., Shum, H.Y.: Dino: Detr with improved denoising anchor boxes for end-to-end object detection (2022), https://arxiv.org/abs/2203.036055

    Pith/arXiv arXiv 2022

  52. [52]

    prompt". Optionally include a terse

    Zhang,X.,Zhang,T.,Wang,G.,Zhu,P.,Tang,X.,Jia,X.,Jiao,L.:Remotesensing object detection meets deep learning: A metareview of challenges and advances. IEEE Geoscience and Remote Sensing Magazine11(4), 8–44 (2023).https://doi. org/10.1109/MGRS.2023.33123478, 10 Diagnosing Aerial-View Detectors using Generative Models 1 Diagnosing Aerial-View Object Detectors...

    doi:10.1109/mgrs.2023.33123478 2023