Pith. sign in

REVIEW 3 major objections 4 minor 82 references

A game-engine pipeline can generate large-scale, precisely annotated multi-view RGB-thermal aerial detection data, and pretraining on it beats standard ImageNet initialization on real-world benchmarks.

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 · deepseek-v4-flash

2026-08-01 11:12 UTC pith:WMFXRHSL

load-bearing objection G-MAD is a solid, useful framework and benchmark, but the annotation pipeline's origin-symmetry assumption and the confounded transfer experiments need fixing before the results can be taken at face value. the 3 major comments →

arxiv 2607.19942 v2 pith:WMFXRHSL submitted 2026-07-22 cs.CV cs.AI

G-MAD: A Game-Based Data Generation Framework for Multi-View RGB-T Aerial Object Detection

Yechan Kim , JongHyun Park , Dongho Yoon , Namhoon Jung , Moongu Jeon This is my paper
classification cs.CV cs.AI
keywords synthetic data generationRGB-thermalaerial object detectionmulti-vieworiented bounding boxpretraininggame enginebenchmark
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.

This paper introduces G-MAD, a framework that turns a military simulation game into a controllable data generator for aerial object detection. It produces synchronized multi-view visible and thermal images with automatic, geometry-based bounding boxes—both horizontal and oriented—using the game engine's internal 3D metadata. The authors build AMOD, a benchmark of over 73,000 images and roughly 383,000 annotated instances across 12 military vehicle categories, and show that training on multiple simultaneous viewpoints improves detection on unseen angles by about six mean-AP points. They also show that pretraining on AMOD, after a synthetic-to-real style transfer, improves performance on a real-world visible benchmark and a real-world thermal benchmark compared with standard initialization.

Core claim

The central claim is that a commercial tactical simulation game can be repurposed as a scalable synthetic-data pipeline for multi-view RGB-thermal aerial object detection, with annotations obtained directly from engine-level geometric metadata rather than manual labeling or image differencing. Using this pipeline, the authors construct AMOD and demonstrate that (1) training a detector on all six observation angles improves generalization to unseen viewpoints relative to single-view training, and (2) pretraining on AMOD, followed by an unpaired image-to-image domain transfer and a boundary-smoothing augmentation, yields better real-world detection accuracy than standard ImageNet initializatio

What carries the argument

The annotation pipeline is the load-bearing mechanism: for each object, the engine's boundingBoxReal function returns an upper corner of the model-space 3D bounding box; the framework assumes the lower corner is its negation, enumerates all eight corners by sign flips, transforms them to world coordinates, projects them to screen coordinates with safe-zone compensation, and then derives a horizontal box from coordinate extremes and an oriented box via convex hull plus rotating calipers. This converts synthetic captures into detection-ready labels without manual effort, and it is paired with a scenario/scene sampling formulation that decouples object layout (scenario) from camera viewpoint an

Load-bearing premise

The automatic annotation assumes that each object model's 3D bounding box is centered at the local coordinate origin, so the lower corner is exactly the negation of the upper corner; if a model's origin lies at ground level instead, all projected 2D boxes for that model will be systematically shifted, corrupting every label derived from the pipeline.

What would settle it

Render a small set of Arma3 objects—including at least one vehicle model known to have a ground-level pivot—with the G-MAD pipeline, and compare the generated HBB/OBB against boxes obtained from manual annotation or from a depth-buffer/ray-casting ground-truth renderer. If the projected boxes are consistently offset from the visible object silhouette (e.g., low Intersection-over-Union with the rendered object mask) for such models, the centered-bounding-box assumption is false and the AMOD annotations are biased.

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

If this is right

  • Multi-view training on AMOD improves detection across unseen observation angles, suggesting that simultaneous multi-view supervision is a valuable signal for viewpoint-robust aerial detection.
  • AMOD pretraining transfers to real-world visible and thermal aerial benchmarks, outperforming ImageNet initialization by a clear margin, which supports the use of game-engine synthetic data for reducing the synthetic-to-real gap.
  • The framework's automatic annotation and modular viewpoint sampler enable controlled studies of viewpoint variation, modality fusion, and domain transfer without costly real-world data collection.
  • The OBB annotations and consistent cross-view object IDs open the door to research on oriented object detection and cross-view consistent representation learning.
  • The engine-native annotation method avoids the error-prone image-differencing approaches of prior game-based pipelines, enabling multi-object scenes and configurable resolution with supersampling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The claim of 'high-quality labels' rests on an unverified geometric assumption: that every Arma3 model's 3D bounding box is centered at the local origin, so the lower corner equals the negation of the upper corner. For many vehicle models whose pivot sits at ground level, this would produce systematically misaligned 2D boxes, biasing every downstream experiment.
  • The reported multi-view improvement is partly confounded by data quantity, since the 'all angles' model sees six times more images; the authors include a supplementary analysis suggesting angular diversity adds value beyond volume, but the disentangling is not fully conclusive.
  • The thermal pretraining pathway is indirect: the game's white-hot thermal rendering is first translated with an unpaired image-to-image model and then augmented with stochastic boundary smoothing, so the final thermal domain realism is approximate and the transfer gains may be sensitive to the choice of translator and augmentation.
  • Because the annotation pipeline is geometry-based and fully automatic, the framework could be extended to generate temporal video sequences for tracking or action recognition with minimal modification, though the paper only demonstrates discrete multi-view frames.

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 / 4 minor

Summary. The paper presents G-MAD, an open-source framework for generating synthetic multi-view RGB-T aerial object detection data using the game Arma3. The framework supports user-defined scenario constraints, controllable camera viewpoint sampling, simultaneous visible/thermal capture, occlusion-aware filtering, and automatic HBB/OBB annotation from engine-level 3D geometry. The authors use G-MAD to construct and release AMOD, a large multi-view RGB-T benchmark with 73,920 images, 383.2k instances in 12 military object classes, with synchronized observations at six viewing angles and consistent object IDs. Experiments show that multi-angle training improves cross-angle AP on AMOD and that AMOD pretraining with UNIT style transfer and Stochastic Boundary Smoothing (SBS) improves AP50 over ImageNet initialization on DIOR-R and HIT-UAV, along with a qualitative unseen-real-world inference test.

Significance. If the annotation pipeline is correct, G-MAD and AMOD would fill a real gap: existing RGB-T aerial benchmarks are largely single-view, while AMOD provides synchronized multi-view RGB-T data with OBB labels, configurable viewpoint sampling, and reproducible open-source code. The framework's use of engine-native geometry rather than image differencing is a clear improvement over prior Arma3 pipelines, and the release of code and dataset is valuable to the community. The paper also explicitly includes scenario sampling, multi-map support, and GU-independent automation, which are practical contributions. However, the load-bearing claim of 'high-quality labels' rests on an unverified geometric assumption in the annotation pipeline; this directly affects the validity of every reported experiment, so the contribution is conditional on resolving that issue.

major comments (3)
  1. [Supp. §A.1, Fig. S-1] The annotation pipeline uses only the upper corner (x_m, y_m, z_m) from boundingBoxReal (select 1) and enumerates the eight model-space corners by taking all sign combinations. This is equivalent to assuming the lower corner is exactly the negation of the upper corner, i.e., that every model-space bounding box is centered at the local origin. The caption's statement that 'Arma3 assumes that each object is located at the origin' does not imply centering; a model with a ground-level pivot has a true z-extent [0, h] rather than [-h/2, h/2], and asymmetric x/y extents are equally possible. If this assumption fails for any of the 513 used assets, every projected HBB/OBB is systematically shifted, corrupting training and evaluation labels. The paper provides no validation against manual annotations, depth buffers, or rendered masks, and no documentation that boundingBoxReal returns half-extent
  2. [§2.3 and Table 4 / Supp. §C] The synthetic-to-real transfer claim in Table 4 does not isolate the contribution of AMOD itself. The pretraining pipeline includes UNIT style transfer and, for HIT-UAV, the proposed SBS augmentation; Table S-2 shows that SBS alone contributes +1.38 AP50 over AMOD without SBS (75.70 -> 77.08), while AMOD without SBS contributes +1.13 over ImageNet. Without ablations that remove style transfer, use a different synthetic source, or train directly on untranslated AMOD, the results do not establish that AMOD's content (as opposed to the transfer pipeline and augmentation) is what provides the transferable signal. At minimum, the paper should report pretraining on AMOD without UNIT transfer and with an alternative generic synthetic dataset, so readers can assess what is specific to G-MAD/AMOD. The current wording '(AMOD pretraining) consistently outperforms ImageNet initialization' overstates
  3. [Table 3 and Fig. S-4] The main-text claim that multi-view training improves generalization based on Table 3 is confounded by data quantity: the 'All' model sees six times more images than each single-angle model. The controlled comparison appears only in Fig. S-4, which correctly shows that data quantity is the dominant factor and that angular diversity provides a smaller, secondary benefit. This controlled experiment should be presented in the main text or at least summarized there, because it is the evidence that actually supports the multi-view supervision claim made in Section 3 and the abstract. Without it, a reader cannot tell whether the +6.39 AP in Table 3 is a viewpoint-diversity effect or simply a data-volume effect.
minor comments (4)
  1. [Abstract and Section 1] The paper consistently misspells 'HIT-UAV' as 'Hit-UAV' or 'HIT-UA V' in a few places (e.g., Table 4 header, Supp. §C.3, Fig. S-8). Please standardize the benchmark name.
  2. [Supp. §C.3] The sentence 'SBS can be seamlessly incorporated into other datasets as well, as can be seen in Fig. S-8' is repeated nearly verbatim; remove the duplication.
  3. [Fig. 3] In Fig. 3, the label '(b-1) / (a)' is unclear; the caption refers to '(b-1) Instance count per category' but the relationship to panel (a) is not explained. A cleaner sub-figure labeling would improve readability.
  4. [Supp. §C.2] The pipeline description says UNIT embeds both domains into a shared latent space Z, but the 'domain adaptation' framing is not precise: a UNIT model trained on image pairs (here synthetic-to-real) does not by itself guarantee a shared latent space that is meaningful for detection. A brief clarification of the training data used for UNIT (which real dataset(s) provided the target domain for translation) would help reproducibility.

Circularity Check

0 steps flagged

No circular derivation; annotation assumption is a data-quality risk, not a tautological fit.

full rationale

G-MAD is an empirical data-generation framework and dataset paper, not a derivation with fitted parameters: no quantity is fitted to a subset of data and then reported as a prediction of closely related data. The transfer experiments are validated against external benchmarks (DIOR-R, HIT-UAV) and on unseen real news imagery, so the central empirical claims are externally checkable rather than self-referential. The one load-bearing assumption is in Supp. §A.1: the annotation pipeline reads only the upper corner from boundingBoxReal and enumerates all sign combinations to generate the eight model-space corner points, which assumes an origin-centered, symmetric model-space bounding box. If Arma3 models are not centered at their local origin, all AMOD HBB/OBB labels would be systematically biased. However, this is an unvalidated geometric assumption and a benchmark-quality risk, not circular reasoning: the paper does not fit a parameter and then rename it as a prediction, and the assumption is explicitly stated rather than embedded in a self-citation. The self-citations ([29], [31]) are used for a generic class-imbalance remark and as one of several references for a routine detector choice; neither is load-bearing. Overall, no significant circularity; the score 1 reflects the minor but identifiable annotation-assumption risk without treating it as a tautological derivation.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities. The main extra assumptions concern Arma3's engine geometry and empirical choices in dataset construction; these are load-bearing because they define annotation quality.

free parameters (4)
  • SBS hyperparameters = p=0.5, sigma=11, tau_min=2, tau_max=5
    Chosen by hand for thermal pretraining augmentation (Supp. §C.3); directly affects reported HIT-UAV gains.
  • Camera altitude and FOV = z=120 m, VFOV=75 deg
    Empirically selected to yield shadows and GSD of 0.1 m/pixel (Supp. §B.2); changes image scale and content.
  • Object placement counts = k in [1,8] classes, e in [8,14] objects
    Empirically chosen scenario sampling ranges (Supp. §B.2).
  • Class balancing ratio = not specified
    Post-hoc balancing to 'make ratio as uniform as possible' (Supp. §B.2, Data balancing) is ad hoc and not a fixed rule.
axioms (4)
  • domain assumption Arma3 engine functions boundingBoxReal, modelToWorldVisual, and worldToScreen return geometrically accurate data
    The entire automatic annotation pipeline relies on these engine functions (Supp. §A); any error propagates to all labels.
  • domain assumption Each object's model-space bounding box is centered at its local origin
    The enumeration of all sign combinations of the upper corner (Supp. §A.1) is only valid if the lower corner is the exact negation of the upper corner; this is not verified for any of the 513 assets.
  • standard math Graham scan and rotating calipers yield the minimum-area enclosing rectangle of the projected points
    Standard computational geometry, cited as [20], [58].
  • domain assumption Line-of-sight checks correctly identify fully occluded objects
    Occlusion filtering relies on a direct line-of-sight test (Sec. 2.3); no validation against renderer depth buffer is given.

pith-pipeline@v1.3.0-alltime-deepseek · 19794 in / 10530 out tokens · 101567 ms · 2026-08-01T11:12:44.462716+00:00 · methodology

0 comments
read the original abstract

This work introduces G-MAD, an open-source framework that uses Arma3 to generate synchronized multi-view RGB-T data for aerial object detection. G-MAD addresses key limitations of real-world aerial dataset construction, including limited viewpoint control, imperfect RGB-T alignment and high annotation cost. The framework supports structured scenario specification, controllable multi-view camera placement, simultaneous visible/thermal capture, and automatic bounding box annotation using engine-level geometric metadata. These capabilities enable controlled studies of viewpoint variation, multi-modal fusion, and synthetic-to-real transfer in aerial object detection. Besides, using G-MAD, we construct and release AMOD, a new large-scale multi-view aerial RGB-T object detection benchmark. The source code and the dataset are available at https://unique-chan.github.io/G-MAD-Project.

Figures

Figures reproduced from arXiv: 2607.19942 by Dongho Yoon, Jonghyun Park, Moongu Jeon, Namhoon Jung, Yechan Kim.

Figure 1
Figure 1. Figure 1: The overall architecture of G-MAD. Arma3 has not been fully explored as a structured data-generation platform for multi-view RGB-T aerial object detection. Hence, we present G-MAD, an open-source framework that transforms Arma3 [4] into a structured data-generation pipeline for multi-view RGB-T aerial object detection. G-MAD1 supports sce￾nario specification, synchronized RGB-T multi-view capture, and auto… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of one of the default camera viewpoint [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Statistics of the AMOD dataset. (a) Object models [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Annotated examples from the AMOD dataset fea [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative analysis of an AMOD-trained detec [PITH_FULL_IMAGE:figures/full_fig_p004_5.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

82 extracted references · 2 linked inside Pith

  1. [1]

    Mehmet Akif Özkanoğlu and Sedat Ozer. 2022. InfraGAN: A GAN architecture to transfer visible images to infrared domain.Pattern Recognit. Lett.155 (2022), 69–76

    2022

  2. [2]

    https://github.com/cloftus96/Synthetic-Data-Generation

    [Website]. https://github.com/cloftus96/Synthetic-Data-Generation

  3. [3]

    https://github.com/ttsiapras/Arma3DatasetGen

    [Website]. https://github.com/ttsiapras/Arma3DatasetGen

  4. [4]

    [Website]. Arma3. https://arma3.com

  5. [5]

    Eunsu Baek, Keondo Park, Jiyoon Kim, and Hyung-Sin Kim. 2024. Unexplored faces of robustness and out-of-distribution: Covariate shifts in environment and sensor domains. InIEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). 22294–22303

    2024

  6. [6]

    Maciej Bala, Yin Cui, Yifan Ding, Yunhao Ge, Zekun Hao, Jon Hasselgren, Jacob Huffman, Jingyi Jin, JP Lewis, Zhaoshuo Li, et al. 2024. Edify 3d: Scalable high- quality 3d asset generation.arXiv preprint arXiv:2411.07135(2024)

    Pith/arXiv arXiv 2024

  7. [7]

    Mahroosh Banday and Brejesh Lall. 2025. Multi spectral visible-thermal IR image translation using improved u-net & conditional diffusion.Neurocomputing(2025), 131006

    2025

  8. [8]

    Hrishav Bakul Barua, Kalin Stefanov, KokSheik Wong, Abhinav Dhall, and Ganesh Krishnasamy. 2025. GTA-HDR: A large-scale synthetic dataset for HDR image reconstruction. InIEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). IEEE, 7876–7886

    2025

  9. [9]

    Octavian Blaga, AI Synetic, and David Scott. 2025. Breaking the Bottleneck: Synthetic Data as the New Foundation for Vision AI. (2025)

    2025

  10. [10]

    Lijing Cai, Xiangyu Dong, Kailai Zhou, and Xun Cao. 2024. Exploring video denoising in thermal infrared imaging: Physics-inspired noise generator, dataset, and model.IEEE Trans. Image Process.33 (2024), 3839–3854

    2024

  11. [11]

    Zizhao Chen, Yeqiang Qian, Xiaoxiao Yang, Chunxiang Wang, and Ming Yang

  12. [12]

    Gong Cheng, Junwei Han, Peicheng Zhou, and Lei Guo. 2014. Multi-class geospa- tial object detection and geographic image classification based on collection of part detectors.ISPRS J. Photogramm. Remote Sens.98 (2014), 119–132

    2014

  13. [13]

    Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. 2022. Anchor-free oriented proposal generator for object detection. IEEE Trans. Geosci. Remote Sens.60 (2022), 1–11

    2022

  14. [14]

    Sungjin Cheong, Wonho Jung, Yoon Seop Lim, and Yong-Hwa Park. 2024. Thermal-infrared remote-target detection system for maritime rescue using 3-D game-based data augmentation with GAN.IEEE Transactions on Geoscience and Remote Sensing62 (2024), 1–13

    2024

  15. [15]

    Rita Delussu, Lorenzo Putzu, and Giorgio Fumera. 2024. Synthetic data for video surveillance applications of computer vision: A review.International Journal of Computer Vision132, 10 (2024), 4473–4509

    2024

  16. [16]

    Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al . 2021. Object detection in aerial images: A large-scale benchmark and challenges.IEEE Trans. Pattern Anal. Mach. Intell.44, 11 (2021), 7778–7796

    2021

  17. [17]

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. 2017. CARLA: An open urban driving simulator. InConference on Robot Learning (CoRL). PMLR, 1–16

    2017

  18. [18]

    Aritra Dutta, Srijan Das, Jacob Nielsen, Rajatsubhra Chakraborty, and Mubarak Shah. 2024. Multiview aerial visual recognition (mavrec): Can multi-view improve aerial visual perception?. InIEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). 22678–22690

    2024

  19. [19]

    Free FLIR Thermal Dataset for Algorithm Training

    [Website]. Free FLIR Thermal Dataset for Algorithm Training. https://oem.flir. com/solutions/automotive/adas-dataset-form/

  20. [20]

    Ronald L. Graham. 1972. An efficient algorithm for determining the convex hull of a finite planar set.Info. Proc. Lett.1 (1972)

    1972

  21. [21]

    Grand Theft Auto V

    [Website]. Grand Theft Auto V. https://www.rockstargames.com/gta-v

  22. [22]

    Zonghao Han, Shun Zhang, Yuru Su, Xiaoning Chen, and Shaohui Mei. 2024. DR- AVIT: Toward diverse and realistic aerial visible-to-infrared image translation. IEEE Trans. Geosci. Remote Sens.62 (2024), 1–13

    2024

  23. [23]

    Meng-Ru Hsieh, Yen-Liang Lin, and Winston H Hsu. 2017. Drone-based object counting by spatially regularized regional proposal network. InIEEE Int. Conf. Comput. Vis. (ICCV). 4145–4153

    2017

  24. [24]

    Soonmin Hwang, Jaesik Park, Namil Kim, Yukyung Choi, and In So Kweon. 2015. Multispectral pedestrian detection: Benchmark dataset and baseline. InIEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). 1037–1045

    2015

  25. [25]

    Yuxiang Ji, Boyong He, Zhuoyue Tan, and Liaoni Wu. 2025. Game4loc: A uav geo-localization benchmark from game data. InAAAI Conf. Artif. Intell. (AAAI), Vol. 39. 3913–3921

    2025

  26. [26]

    Yuxiang Ji, Boyong He, Zhuoyue Tan, and Liaoni Wu. 2025. MMGeo: Multimodal Compositional Geo-Localization for UAVs. InIEEE Int. Conf. Comput. Vis. (ICCV). 25165–25175

    2025

  27. [27]

    Chenchen Jiang, Huazhong Ren, Fengguang Li, Zhonghua Hong, Hongtao Huo, Junqiang Zhang, and Jiuyuan Xin. 2025. Object detection from aerial multi- angle thermal infrared remote sensing images: Dataset and method.ISPRS J. Photogramm. Remote Sens.228 (2025), 438–452

    2025

  28. [28]

    Zhao Jin, Yinjie Lei, Naveed Akhtar, Haifeng Li, and Munawar Hayat. 2022. Deformation and correspondence aware unsupervised synthetic-to-real scene flow estimation for point clouds. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 7233–7243

    2022

  29. [29]

    Yechan Kim, Younkwan Lee, and Moongu Jeon. 2021. Imbalanced image clas- sification with complement cross entropy.Pattern Recognit. Lett.151 (2021), 33–40

    2021

  30. [30]

    Beomsu Kim, Gihyun Kwon, Kwanyoung Kim, and Jong Chul Ye. 2024. Unpaired image-to-image translation via neural schr\" odinger bridge. InInt. Conf. Learn. Represent. (ICLR)

    2024

  31. [31]

    Yechan Kim, SooYeon Kim, and Moongu Jeon. 2025. Nbbox: Noisy bounding box improves remote sensing object detection.IEEE Geoscience and Remote Sensing Letters22 (2025), 1–5

    2025

  32. [32]

    Alexander Kirillov et al. 2023. Segment anything. InIEEE Int. Conf. Comput. Vis. (ICCV)

    2023

  33. [33]

    Darius Lam, Richard Kuzma, Kevin McGee, Samuel Dooley, Michael Laielli, Matthew Klaric, Yaroslav Bulatov, and Brendan McCord. 2018. xview: Objects in context in overhead imagery.arXiv preprint arXiv:1802.07856(2018)

    Pith/arXiv arXiv 2018

  34. [34]

    Yunwei Lan, Zhigao Cui, Xin Luo, Chang Liu, Nian Wang, Menglin Zhang, Yanzhao Su, and Dong Liu. 2025. When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training. InIEEE Int. Conf. Comput. Vis. (ICCV). 8756–8765

    2025

  35. [35]

    Dong-Guw Lee, Myung-Hwan Jeon, Younggun Cho, and Ayoung Kim. 2023. Edge-guided multi-domain rgb-to-tir image translation for training vision tasks with challenging labels. InIEEE Int. Conf. Robot. Autom. (ICRA)

    2023

  36. [36]

    Yujie Lei, Jie Zhang, Wenjie Sun, Wei He, Jiasong Zhu, and Qingquan Li. 2025. MGNet: A Remote Sensing Oriented Object Detector Based on Multi-Cascaded Feature Selection and Geometric Constraints.IEEE Transactions on Geoscience and Remote Sensing(2025)

    2025

  37. [37]

    Ke Li et al. 2020. Object detection in optical remote sensing images: A survey and a new benchmark.ISPRS J. Photogramm. Remote Sens.159 (2020)

    2020

  38. [38]

    Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. 2023. Bbdm: Image-to-image trans- lation with brownian bridge diffusion models. InIEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). 1952–1961

    2023

  39. [39]

    Na Li, Haining Wang, Huijie Zhao, and Wen Ou. 2025. Cross-Modal Visible-to- Infrared Image Translation in Remote Sensing Guided by Thermal Features.IEEE Trans. Geosci. Remote Sens.(2025)

    2025

  40. [40]

    Wei Lin, Junyu Gao, Qi Wang, and Xuelong Li. 2021. Learning to detect anomaly events in crowd scenes from synthetic data.Neurocomputing436 (2021), 248–259

    2021

  41. [41]

    Kang Liu and Gellert Mattyus. 2015. Fast multiclass vehicle detection on aerial images.IEEE Geosci. Remote Sens. Lett.12, 9 (2015), 1938–1942

    2015

  42. [42]

    Zikun Liu, Hongzhen Wang, Lubin Weng, and Yiping Yang. 2016. Ship rotated bounding box space for ship extraction from high-resolution optical satellite images with complex backgrounds.IEEE Geosci. Remote Sens. Lett.13, 8 (2016), 1074–1078

    2016

  43. [43]

    Ming-Yu Liu, Thomas Breuel, and Jan Kautz. 2017. Unsupervised image-to-image translation networks. InAdv. Neural Inf. Process. Syst. (NeurIPS). 1–9

    2017

  44. [44]

    Ze Liu et al . 2021. Swin transformer: Hierarchical vision transformer using shifted windows. InIEEE International Conference on Computer Vision (ICCV). MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Yechan Kim et al

    2021

  45. [45]

    Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. 2022. Target-aware dual adversarial learning and a multi- scenario multi-modality benchmark to fuse infrared and visible for object detec- tion. InIEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR). 5802–5811

    2022

  46. [46]

    Jiahang Liu, Xiaozhen Wang, Mao Guo, Ruilei Feng, and Yue Wang. 2023. Shadow detection in remote sensing images based on spectral radiance separability en- hancement.IEEE Trans. Pattern Anal. Mach. Intell.46, 5 (2023), 3438–3449

    2023

  47. [47]

    Yang Long, Yiping Gong, Zhifeng Xiao, and Qing Liu. 2017. Accurate object localization in remote sensing images based on convolutional neural networks. IEEE Trans. Geosci. Remote Sens.55, 5 (2017), 2486–2498

    2017

  48. [48]

    Decao Ma, Juan Su, Bing Li, Yong Xian, Shaopeng Li, and Yao Ding. 2025. Self cycle strategy for unpaired visible-to-infrared image translation.Pattern Recognit. (2025), 112253

    2025

  49. [49]

    Jiancheng Pan, Yanxing Liu, Yuqian Fu, Muyuan Ma, Jiahao Li, Danda Pani Paudel, Luc Van Gool, and Xiaomeng Huang. 2025. Locate anything on earth: Advancing open-vocabulary object detection for remote sensing community. InAAAI Conf. Artif. Intell. (AAAI), Vol. 39. 6281–6289

    2025

  50. [50]

    Goran Paulin and Marina Ivasic-Kos. 2023. Review and analysis of synthetic dataset generation methods and techniques for application in computer vision. Artificial Intelligence Review56, 9 (2023), 9221–9265

    2023

  51. [51]

    Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. 2016. Playing for data: Ground truth from computer games. InEuropean Conference on Computer Vision (ECCV). Springer, 102–118

    2016

  52. [52]

    Kechen Song, Xiaotong Xue, Hongwei Wen, Yingying Ji, Yunhui Yan, and Qing- gang Meng. 2024. Misaligned visible-thermal object detection: A drone-based benchmark and baseline.IEEE Trans. Intell. Veh.(2024)

    2024

  53. [53]

    Jiamin Song, Nan Zhang, Zhenhao Wang, and Tian Tian. 2026. ESCVehicle: A Drone-based Visible-Infrared Vehicle Benchmark with Extensive Scene Coverage. IEEE Trans. Geosci. Remote Sens.(2026)

    2026

  54. [54]

    Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. 2022. Drone-based RGB- infrared cross-modality vehicle detection via uncertainty-aware learning.IEEE Trans. Circuits Syst. Video Technol.32, 10 (2022), 6700–6713

    2022

  55. [55]

    Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. 2022. FAIR1M: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS J. Photogramm. Remote Sens.184 (2022), 116–130

    2022

  56. [56]

    Karasawa Takumi, Kohei Watanabe, Qishen Ha, Antonio Tejero-De-Pablos, Yoshi- taka Ushiku, and Tatsuya Harada. 2017. Multispectral object detection for au- tonomous vehicles. InACM Int. Conf. Multimedia (MM) Worksh.35–43

    2017

  57. [57]

    Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. 2022. PI- AFusion: A progressive infrared and visible image fusion network based on illumination aware.Inf. Fusion83 (2022), 79–92

    2022

  58. [58]

    Godfried T Toussaint. 1983. Solving geometric problems with the rotating calipers. InIEEE Melecon, Vol. 83

    1983

  59. [59]

    [Website]. Unity. https://unity.com

  60. [60]

    [Website]. Unreal. https://www.unrealengine.com

  61. [61]

    Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. 2019. Learning from synthetic data for crowd counting in the wild. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 8198–8207

    2019

  62. [62]

    Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. 2021. Pixel-wise crowd under- standing via synthetic data.International Journal of Computer Vision129, 1 (2021), 225–245

    2021

  63. [63]

    Tao Wang, Chenyu Lin, Chenwei Tang, Jizhe Zhou, Deng Xiong, Jianan Li, Jian Zhao, and Jiancheng Lv. 2026. Adaptive image zoom-in with bounding box transformation for UAV object detection.ISPRS Journal of Photogrammetry and Remote Sensing233 (2026), 452–466

    2026

  64. [64]

    Nicholas Weir, David Lindenbaum, Alexei Bastidas, Adam Van Etten, Sean McPherson, Jacob Shermeyer, Varun Kumar, and Hanlin Tang. 2019. Spacenet mvoi: A multi-view overhead imagery dataset. InIEEE Int. Conf. Comput. Vis. (ICCV). 992–1001

    2019

  65. [65]

    Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai

  66. [66]

    Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. 2018. DOTA: A large-scale dataset for object detection in aerial images. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 3974–3983

    2018

  67. [67]

    Xingxing Xie, Gong Cheng, Jiabao Wang, Ke Li, Xiwen Yao, and Junwei Han

  68. [68]

    Ze Yang, Jingkang Wang, Haowei Zhang, Sivabalan Manivasagam, Yun Chen, and Raquel Urtasun. 2025. Genassets: Generating in-the-wild 3d assets in latent space. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 22392–22403

    2025

  69. [69]

    Haijun Zhang, Mingshan Sun, Qun Li, Linlin Liu, Ming Liu, and Yuzhu Ji. 2021. An empirical study of multi-scale object detection in high resolution UAV images. Neurocomputing421 (2021), 173–182

    2021

  70. [70]

    Min Zhao, Fan Bao, Chongxuan Li, and Jun Zhu. 2022. Egsde: Unpaired image- to-image translation via energy-guided stochastic differential equations. InAdv. Neural Inf. Process. Syst. (NeurIPS). 3609–3623

    2022

  71. [71]

    Lei Zhao, Mengwei Li, Bo Li, and Xingxing Wei. 2025. Diverse Visible-to-Thermal Image Translation via Controllable Temperature Encoding.IEEE Trans. Multime- dia(2025)

    2025

  72. [72]

    Shangdong Zheng, Yang Xu, Peng Zheng, Zhihui Wei, and Zebin Wu. 2026. See Hidden Insight From Transposition: Multi-Axis Feature Aggregation for Aerial Object Detection.IEEE Transactions on Geoscience and Remote Sensing(2026)

    2026

  73. [73]

    Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. 2022. Mmrotate: A rotated object detection benchmark using pytorch. InACM International Conference on Multimedia. 7331–7334

    2022

  74. [74]

    G-MAD: A Game-Based Data Generation Framework for Multi-View RGB-T Aerial Object Detection

    Haigang Zhu, Xiaogang Chen, Weiqun Dai, Kun Fu, Qixiang Ye, and Jianbin Jiao. 2015. Orientation robust object detection in aerial images using deep convolutional neural network. InIEEE Int. Conf. Image Process. (ICIP). IEEE, 3735–3739. Supplementary Material for “G-MAD: A Game-Based Data Generation Framework for Multi-View RGB-T Aerial Object Detection” M...

    2015

  75. [78]

    Aerial Military Object Detection in multi-view RGB-T scenarios

    rather than general-purpose game engines such as Unity [59] and Unreal [60]. While those engines offer high-fidelity rendering, they often require custom asset modeling or significant licensing costs of domain-specific 3D models of equipment and terrain [6, 9, 17, 68], for large-scale data construction. In contrast, Arma3 already provides rich 3D object m...

    2026

  76. [79]

    The ground surface may not be flat

  77. [80]

    “𝑧” is a height above ground level (AGL) at

  78. [81]

    Figure S-2: Illustration of how to construct our AMOD bench- mark

    VFOV of the drone cam is 75°. Figure S-2: Illustration of how to construct our AMOD bench- mark. (a) Geometric definition of the observing angle 𝜃 , where the camera shifts laterally by Δ𝑦 in𝑌 -axis, producing an off-nadir view. (b) Drone camera sweeping setup simulta- neously capturing the same area from nadir to very off-nadir viewpoints. (c) Examples f...

    1920

  79. [82]

    G-MAD: A Game-Based Data Generation Framework for Multi-View RGB-T Aerial Object Detection

    followed by rotating calipers [58] to find the minimum-area enclosing rectangle. Finally, as summarized in Tab. 2 (Main), in our benchmark, to facilitate future multi-view studies, the same object observed across different views is assigned a consistent object ID for each region of interest (A). Data balancing.Class imbalance is undesirable, as it may bia...

    2026

  80. [2017]

    InIEEE Int

    RGB-infrared cross-modality person re-identification. InIEEE Int. Conf. Comput. Vis. (ICCV). 5380–5389

Showing first 80 references.