Pith. sign in

REVIEW 3 major objections 5 minor 57 references

Drone Detection using Deep Neural Networks Trained on Pure Synthetic Data

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A drone detector trained only on synthetic images nearly matches real-data training, scoring 97.0% AP50 on MAV-Vid versus 97.8% for the real-data baseline.

desk verdict Worth reading for the dataset and the ablation, but the pure-synthetic transfer claim is overstated. read the letter →

arxiv 2411.09077 v1 pith:CUX2NA7P submitted 2024-11-13 cs.CV cs.AIcs.LG

Mariusz Wisniewski , Zeeshan A. Rana , Ivan Petrunin , Alan Holt , Stephen Harman This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords dronedetectionsyntheticdatastructureddomainrandomizationsim-to-realtransferFasterR-CNNobjectMAV-Vidbenchmark
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 asks whether a drone detector can be trained entirely on synthetic images and still work on real drone footage. It answers yes: a Faster R-CNN trained on a synthetic dataset rendered with structured domain randomization reaches an AP50 of 97.0% on the MAV-Vid real-world benchmark, against 97.8% for the same detector trained on real MAV-Vid images. The model is trained once and tested without fine-tuning on three real datasets, so the result is a single set of weights rather than a separate model per dataset. If the claim holds, drone-detection training data can be generated cheaply with automated pixel-accurate labels, reducing the cost and human effort of building real-world datasets.

What carries the argument

Structured domain randomization (SDR) is the central mechanism: it generates synthetic training images that keep a realistic scene context while randomizing global parameters such as camera position, focal length, background, and lighting, so the network learns drone shape rather than dataset-specific cues. The pipeline renders drone models in varied environments, produces pixel-accurate segmentation masks automatically, and uses a Faster R-CNN with a ResNet-50 backbone, initialized with MS COCO pretrained weights, as the detector. The camera-bounds randomization is the main tested lever on apparent drone size and therefore on transfer performance, while noise, JPEG compression, and distractor styles are secondary variables.

What would settle it

Hold out a portion of MAV-Vid before any configuration choice, fix the 40 m bound and the noise/JPEG settings in advance, retrain, and evaluate; if AP50 falls well below 97.0%, the reported transfer is inflated by tuning to the test set.

Watch

Extended reading notes

Core claim

The paper's central claim is that structured domain randomization, a way of generating realistic synthetic scenes while randomizing global parameters such as lighting, camera pose, and focal length, makes a synthetic drone dataset usable for real-world detection. The authors render five drone models in varied background environments, randomize camera bounds from 20 m to 320 m, and train a Faster R-CNN with a ResNet-50 backbone on the rendered images with exact segmentation masks. The same weights, evaluated without any real-data fine-tuning, score a mean AP50 of 97.0% on MAV-Vid compared with 97.8% for the reference Faster R-CNN trained on real MAV-Vid data; on Drone-vs-Bird the synthetic-trained model scores 49.8% versus 63.2%, and on Anti-UAV 67.8% versus 97.7%. The paper reports that JPEG compression, noise, birds, generic or realistic distractors, and random backgrounds did not meaningfully improve over the plain drones-only dataset, and it discloses that the network is initialized with MS COCO pretrained weights.

Load-bearing premise

The published 97.0% AP50 assumes that the 40 m camera bound and the noise/JPEG augmentations were chosen before looking at the three real test datasets, so the score is an out-of-domain estimate rather than an in-sample selection.

Editorial extensions

If this is right

  • A detector trained on synthetic images alone can come within about one percentage point of a real-data-trained detector on MAV-Vid, so for similar camera scales and backgrounds, synthetic data can substitute for costly real drone footage.
  • Because the same weights are tested on three datasets without fine-tuning, synthetic training can yield a single general-purpose detector, though its accuracy varies strongly with target-domain artifacts such as camera overlays and night footage.
  • The apparent drone-size distribution in synthetic data, controlled by camera bounds, is a first-order transfer factor: 20 m to 80 m bounds all work, while 320 m bounds break the detector.
  • Adding noise gave a small consistent gain and JPEG compression had negligible effect, so rendering losslessly is not the main sim-to-real obstacle; bridging video compression and camera artifacts would be more valuable.
  • None of the tested domain randomization styles improved over the plain drones-only dataset, challenging the usual advice that distractors help sim-to-real transfer in object detection.

Reading between the lines

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

  • A fully synthetic training run from random initialization, without MS COCO pretrained weights, would clarify how much of the reported transfer comes from the synthetic data alone rather than from generic real-image features learned during pretraining.
  • Because the synthetic dataset includes unused segmentation masks, the same generator can be extended to train segmentation or temporal models; making drone paths consistent across frames would open the way to video-based detectors.
  • The Anti-UAV performance gap suggests the highest-value additions to the synthetic domain are camera overlays and nighttime lighting, both of which are cheaper to render than to record.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 presents a Faster R-CNN detector trained on synthetic drone images generated in Blender using structured domain randomization, and evaluates it on three real-world datasets (MAV-Vid, Drone-vs-Bird, Anti-UAV). The headline result is an AP50 of 97.0% on MAV-Vid, compared with 97.8% for the real-data-trained Faster R-CNN benchmark of Isaac-Medina et al. The authors report ablations of camera bounds, data augmentations, dataset size, and domain randomization styles, and they repeat each configuration 8 times to report means with 95% confidence intervals.

Significance. If the result held as stated, it would provide a practically useful demonstration of sim-to-real transfer for drone detection, with implications for reducing data collection and labelling costs. The manuscript's strengths include release of code and datasets, use of publicly available test datasets, and a deliberate attempt to quantify run-to-run variability through repeated training. However, two load-bearing caveats—the use of MS COCO pretrained weights and the selection of the default configuration based on test-set performance—mean the current evidence does not support the claims of 'purely synthetic' training and clean out-of-distribution generalization.

major comments (3)
  1. [Section II-B] The manuscript repeatedly claims that the model is 'trained on a purely synthetic dataset' (title, abstract, Section I-D), but Section II-B states that 'the network is pretrained on the MS COCO dataset.' Since COCO is a large real-image dataset, the model has already seen a substantial amount of real-world visual data before any synthetic training. This undermines the central claim that the observed sim-to-real transfer is attributable solely to synthetic data. The authors should either train from random initialization and report that result, or explicitly reframe the claim as 'fine-tuned from COCO on synthetic data' and discuss the potential contribution of pretraining to the transfer performance.
  2. [Section III-A and III-B] The default training configuration (40 m camera bounds; JPEG compression and noise enabled) was selected based on performance on the same three test datasets used for the headline results. Figure 5 and its accompanying text state that 'Both MAV-Vid and Anti-UAV perform best on the 40 m dataset,' and Figure 6 shows that JPEG+noise was adopted as the default. Because the final AP50 of 97.0% is obtained from a configuration chosen by inspecting test-set performance, it is an in-sample maximum over the tested configurations rather than an unbiased estimate of out-of-distribution generalization. This selection bias is a genuine threat to the central claim, and a proper held-out validation split (or a correction for multiple comparisons, or presentation of all configurations without elevating one to default) is needed.
  3. [Section III-E and Table I] The abstract and Section III-E describe the Isaac-Medina et al. model as an 'equivalent model trained on real-world data,' but the comparison is not tightly controlled: the benchmark model was trained separately on each dataset, with a single run, and potentially with different training procedures and pretraining choices. The paper acknowledges some of these differences in Section III-E, but the equivalence language overstates the degree of control. The comparison should be framed more cautiously, and the limitations of a single-run, different-procedure baseline should be explicitly stated when interpreting the 97.0% versus 97.8% gap.
minor comments (5)
  1. [Section IV] In the conclusion, the sentence 'The model translates poorly to the Anti-UAV dataset, achieving AP0.5 of 67.8%, compared with 97.7% for an equivalent model trained on the DvB dataset' should say 'Anti-UAV dataset' instead of 'DvB dataset,' since the 97.7% figure corresponds to the Anti-UAV benchmark.
  2. [Section III-C] The dataset size study does not report confidence intervals for most dataset sizes (50 to 2,500 images) because only single runs were performed; this is mentioned in the text but not in the figure caption, so the reader may misread the points as having the same reliability as the 8-run averages elsewhere.
  3. [Section II-A1] The sentence 'A limitation of Blender is that it does not model the function to focus on infinity within its camera model' is awkwardly phrased; consider rewording to 'does not model focusing at infinity within its camera model.'
  4. [Section II-C] The definition of AP50 is only implicit through the IoU discussion; a brief explicit definition of average precision at IoU 0.5 would make the paper more self-contained for readers outside the detection community.
  5. [Table I] Table I reports only point estimates for the authors' results, despite the paper's emphasis on means with 95% confidence intervals; adding the confidence intervals (or pointing to the corresponding figures) would make the table consistent with the stated repeatability methodology.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline 97.0% AP50 on MAV-Vid is selected from camera-bounds and augmentation ablations performed on the same test datasets, making the reported sim-to-real transfer an in-sample optimum rather than a clean prediction.

  1. fitted input called prediction [Section III (intro, III-A, III-B); Table I]
    "Unless otherwise specified, the experiments use the training parameters described in section II-B, a camera bounding size of 40 m is used, a dataset size of 5,000 is used, JPEG compression is enabled, noise is enabled... Both MAV-Vid and Anti-UAV perform best on the 40 m dataset... Adding noise seems to have a slight effect, producing a mean AP0.5 of 97.1% on MAV-Vid... Adding both JPEG compression and noise does not produce a significant difference compared with just adding noise."

    The default configuration behind the headline result of 97.0% on MAV-Vid (40 m camera bounds, JPEG+noise, 5,000 images) was chosen after the paper ran ablations on the same three test datasets and identified which settings performed best on those datasets. The paper states that MAV-Vid and Anti-UAV 'perform best on the 40 m dataset' and adopts 40 m as default; it likewise adopts the augmentation variant with the highest observed means on these test sets. Therefore the reported sim-to-real AP50 is not an out-of-distribution transfer estimate but the best in-sample configuration selected using the test labels.

full rationale

The central numerical claim, 97.0% AP50 on MAV-Vid from a synthetic-only training pipeline, is weakened by test-set-driven hyperparameter selection rather than by a definitional or self-citation circularity. The paper itself reports the camera-bounds and augmentation ablations on MAV-Vid, DvB, and Anti-UAV and then sets the default configuration to the best-performing values on those datasets. The headline number is thus an in-sample maximum over the tried configurations. This does not fully invalidate the transfer finding—other configurations are close (96.6–97.1 on MAV-Vid)—but it does mean the claim 'achieves AP50 of 97.0% when evaluated on MAV-Vid' is not a clean out-of-domain prediction. Separately, the phrase 'purely synthetic dataset' is undercut by the use of MS COCO pretrained weights, but that is a claim-accuracy issue, not a circularity. The comparison to Isaac-Medina et al. is an external benchmark, and the self-citations to Wisniewski et al. [43], [44] are contextual rather than load-bearing. Overall, the primary circularity burden is the test-set-selected configuration behind the reported transfer metric, warranting a score of 6 rather than lower.

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

The central claim rests on the assumption that COCO pretraining is compatible with 'purely synthetic' training, on the comparability of the Isaac-Medina benchmark, and on the validity of test-set-selected hyperparameters as an unbiased estimate. These are domain assumptions rather than mathematical axioms, and no new physical entities are introduced.

free parameters (5)
  • Camera bound size = 40 m
    Selected after observing that 40m gave the best AP50 on MAV-Vid and Anti-UAV (Figure 5); used as the default in the final comparison. This is hyperparameter tuning on the test data.
  • Data augmentation schedule (noise + JPEG) = 50% images compressed in range 0-95%; Gaussian noise enabled
    The augmentations were kept because they slightly improved mean AP50 across test datasets (Figure 6); another test-data-driven choice.
  • Training dataset size = 5,000 images
    Chosen after the dataset size study showed diminishing returns beyond 5,000 (Figure 7); the paper notes that only 5,000 and 15,000 runs were repeated.
  • Focal length randomization range = 15 mm to 300 mm
    Fixed arbitrary range used for all datasets; not ablated.
  • Training hyperparameters = learning rate 0.0003, momentum 0.9, weight decay 0.0005, 10 epochs
    Taken from PyTorch defaults, not optimized; included for completeness.
assumptions (4)
  • domain assumption MS COCO pretrained weights provide a real-world image prior while the model is still described as 'trained on a purely synthetic dataset'.
    The paper states the network is pretrained on MS COCO (Section II-B) but the abstract and conclusion call the training purely synthetic. The novelty claim depends on ignoring this initialization.
  • domain assumption The Isaac-Medina et al. benchmark results are directly comparable even though they trained one model per dataset and used possibly different test subsets.
    Section III-E compares AP50 values without explicitly confirming identical evaluation splits or protocols.
  • domain assumption The ground-truth labels in MAV-Vid, DvB, and Anti-UAV are accepted as correct despite acknowledged imperfections.
    Section II-C notes labels are imperfect and that AP/AP75 are de-emphasized, but AP50 comparisons still depend on those labels.
  • domain assumption Eight training runs are enough to estimate the mean with a 95% CI assuming a Gaussian distribution.
    Section II-D bases repeatability on Picard's seed study and assumes normality despite mentioning black swan events.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Drone Detection using Deep Neural Networks Trained on Pure Synthetic Data." pith.science (2026). https://pith.science/paper/CUX2NA7P

@misc{pith2026241109077,
  author       = {Pith},
  title        = {Pith review of: Drone Detection using Deep Neural Networks Trained on Pure Synthetic Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CUX2NA7P}},
  note         = {Machine review of arXiv:2411.09077}
}
read the original abstract

Drone detection has benefited from improvements in deep neural networks, but like many other applications, suffers from the availability of accurate data for training. Synthetic data provides a potential for low-cost data generation and has been shown to improve data availability and quality. However, models trained on synthetic datasets need to prove their ability to perform on real-world data, known as the problem of sim-to-real transferability. Here, we present a drone detection Faster-RCNN model trained on a purely synthetic dataset that transfers to real-world data. We found that it achieves an AP_50 of 97.0% when evaluated on the MAV-Vid - a real dataset of flying drones - compared with 97.8% for an equivalent model trained on real-world data. Our results show that using synthetic data for drone detection has the potential to reduce data collection costs and improve labelling quality. These findings could be a starting point for more elaborate synthetic drone datasets. For example, realistic recreations of specific scenarios could de-risk the dataset generation of safety-critical applications such as the detection of drones at airports. Further, synthetic data may enable reliable drone detection systems, which could benefit other areas, such as unmanned traffic management systems. The code is available https://github.com/mazqtpopx/cranfield-synthetic-drone-detection alongside the datasets https://huggingface.co/datasets/mazqtpopx/cranfield-synthetic-drone-detection.

Figures

Figures reproduced from arXiv: 2411.09077 by the authors.

Figure 1
Figure 1. The experimental process is split into 3 parts: synthetic [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Synthetic datasets generated by using different styles. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. A zoomed-in example of an inaccurate ground truth [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Results of the bounding size study for 20 m, 40 m, [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 4
Figure 4. Figure 4: A representation selected bounds within which the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 7
Figure 7. Figure 7: Results of the dataset size study for 50, 100, 250, 500, [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 54 canonical work pages

  1. [1]

    Increase in use of drones for prison smuggling,

    B. News, “Increase in use of drones for prison smuggling,” BBC News, Apr. 2023

    work page 2023

  2. [2]

    Drugs, weapons ’smuggled to prisoners by drone’,

    ——, “Drugs, weapons ’smuggled to prisoners by drone’,” BBC News, Feb. 2022

    work page 2022

  3. [3]

    Heathrow airport: Drone sighting halts departures,

    ——, “Heathrow airport: Drone sighting halts departures,” BBC News, Jan. 2019

    work page 2019

  4. [4]

    Flights diverted at East Midlands airport after drone sightings,

    C. Skopeliti, “Flights diverted at East Midlands airport after drone sightings,” The Guardian, Jun. 2022

    work page 2022

  5. [5]

    Dublin airport: Flights suspended for 30 minutes after drone sightings,

    “Dublin airport: Flights suspended for 30 minutes after drone sightings,” https://www.bbc.com/news/articles/c72mvndez5jo, Feb. 2023

    work page 2023

  6. [6]

    UK Counter-Unmanned Aircraft Strategy,

    “UK Counter-Unmanned Aircraft Strategy,” p. 38

    work page

  7. [7]

    Defending Airports from UAS: A Survey on Cyber-Attacks and Counter-Drone Sensing Tech- nologies,

    G. Lykou, D. Moustakas, and D. Gritzalis, “Defending Airports from UAS: A Survey on Cyber-Attacks and Counter-Drone Sensing Tech- nologies,” Sensors, vol. 20, no. 12, p. 3537, Jun. 2020

    work page 2020

  8. [8]

    ImageNet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” Communications of the ACM, vol. 60, no. 6, pp. 84–90, May 2017

    work page 2017

Show all 57 references
  1. [9]

    You Only Look Once: Unified, Real-Time Object Detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You Only Look Once: Unified, Real-Time Object Detection,” May 2016

    2016

  2. [10]

    YOLO9000: Better, Faster, Stronger,

    J. Redmon and A. Farhadi, “YOLO9000: Better, Faster, Stronger,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, Jul. 2017, pp. 6517–6525

    2017

  3. [11]

    SSD: Single Shot MultiBox Detector,

    W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “SSD: Single Shot MultiBox Detector,” 2016, vol. 9905, pp. 21–37

    2016

  4. [12]

    Fast R-CNN,

    R. Girshick, “Fast R-CNN,” Sep. 2015

    2015

  5. [13]

    Faster R-CNN: To- wards Real-Time Object Detection with Region Proposal Networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: To- wards Real-Time Object Detection with Region Proposal Networks,” arXiv:1506.01497 [cs], Jan. 2016

    2016 arXiv

  6. [14]

    Unmanned Aerial Vehicle Visual Detection and Tracking using Deep Neural Networks: A Performance Benchmark,

    B. K. S. Isaac-Medina, M. Poyser, D. Organisciak, C. G. Willcocks, T. P. Breckon, and H. P. H. Shum, “Unmanned Aerial Vehicle Visual Detection and Tracking using Deep Neural Networks: A Performance Benchmark,” in 2021 IEEE/CVF International Conference on Computer Vision Worksh...

    2021

  7. [15]

    Attention Is All You Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention Is All You Need,” Dec. 2017. 12

    2017

  8. [16]

    End-to-End Object Detection with Transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-End Object Detection with Transformers,” May 2020

    2020

  9. [17]

    Towards Out-Of-Distribution Generalization: A Survey,

    J. Liu, Z. Shen, Y . He, X. Zhang, R. Xu, H. Yu, and P. Cui, “Towards Out-Of-Distribution Generalization: A Survey,” Jul. 2023

    2023

  10. [18]

    A Comprehensive Survey on Transfer Learning,

    F. Zhuang, Z. Qi, K. Duan, D. Xi, Y . Zhu, H. Zhu, H. Xiong, and Q. He, “A Comprehensive Survey on Transfer Learning,” Jun. 2020

    2020

  11. [19]

    Domain Adaptation for Visual Applications: A Comprehen- sive Survey,

    G. Csurka, “Domain Adaptation for Visual Applications: A Comprehen- sive Survey,” Mar. 2017

    2017

  12. [20]

    Domain General- ization: A Survey,

    K. Zhou, Z. Liu, Y . Qiao, T. Xiang, and C. C. Loy, “Domain General- ization: A Survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–20, 2022

    2022

  13. [21]

    Domain randomization for transferring deep neural networks from sim- ulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from sim- ulation to the real world,” in 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . Vancouver, BC: IEEE, Sep. ...

    2017

  14. [22]

    Training Deep Networks with Synthetic Data: Bridging the Reality Gap by Domain Randomization,

    J. Tremblay, A. Prakash, D. Acuna, M. Brophy, V . Jampani, C. Anil, T. To, E. Cameracci, S. Boochoon, and S. Birchfield, “Training Deep Networks with Synthetic Data: Bridging the Reality Gap by Domain Randomization,” arXiv:1804.06516 [cs], Apr. 2018

    2018 arXiv

  15. [23]

    Structured Domain Randomization: Bridging the Reality Gap by Context-Aware Synthetic Data,

    A. Prakash, S. Boochoon, M. Brophy, D. Acuna, E. Cameracci, G. State, O. Shapira, and S. Birchfield, “Structured Domain Randomization: Bridging the Reality Gap by Context-Aware Synthetic Data,” in 2019 International Conference on Robotics and Automation (ICRA) , May 2019, pp. ...

    2019

  16. [24]

    Applying Domain Randomization to Synthetic Data for Object Category Detection,

    J. Borrego, A. Dehban, R. Figueiredo, P. Moreno, A. Bernardino, and J. Santos-Victor, “Applying Domain Randomization to Synthetic Data for Object Category Detection,” Jul. 2018

    2018

  17. [25]

    Benchmarking Domain Randomisation for Visual Sim-to-Real Transfer,

    R. Alghonaim and E. Johns, “Benchmarking Domain Randomisation for Visual Sim-to-Real Transfer,” May 2021

    2021

  18. [26]

    On Pre- Trained Image Features and Synthetic Images for Deep Learning,

    S. Hinterstoisser, V . Lepetit, P. Wohlhart, and K. Konolige, “On Pre- Trained Image Features and Synthetic Images for Deep Learning,” arXiv:1710.10710 [cs], Nov. 2017

    2017 arXiv

  19. [27]

    Drone Detection Using YOLOv5,

    B. Aydin and S. Singha, “Drone Detection Using YOLOv5,” Eng, vol. 4, no. 1, pp. 416–433, Mar. 2023

    2023

  20. [28]

    Detection and Recognition of Drones Based on a Deep Convolutional Neural Network Using Visible Imagery,

    F. Samadzadegan, F. Dadrass Javan, F. Ashtari Mahini, and M. Gho- lamshahi, “Detection and Recognition of Drones Based on a Deep Convolutional Neural Network Using Visible Imagery,” Aerospace, vol. 9, no. 1, p. 31, Jan. 2022

    2022

  21. [29]

    An Object Detection Algorithm for Rotary- Wing UA V Based on AWin Transformer,

    Y . Fan, O. Li, and G. Liu, “An Object Detection Algorithm for Rotary- Wing UA V Based on AWin Transformer,” IEEE Access , vol. 10, pp. 13 139–13 150, 2022

    2022

  22. [30]

    A Modified YOLOv4 Deep Learning Network for Vision-Based UA V Recognition,

    F. Dadrass Javan, F. Samadzadegan, M. Gholamshahi, and F. Ashatari Mahini, “A Modified YOLOv4 Deep Learning Network for Vision-Based UA V Recognition,” Drones, vol. 6, no. 7, p. 160, Jul. 2022

    2022

  23. [31]

    Exploitation of data augmentation strategies for improved UA V detection,

    L. Freudenmann, L. Sommer, and A. Schumann, “Exploitation of data augmentation strategies for improved UA V detection,” in Automatic Target Recognition XXXI, vol. 11729. SPIE, Apr. 2021, pp. 119–132

    2021

  24. [32]

    Detecting aerial objects: Drones, birds, and helicopters,

    C. Mediavilla, L. Nans, D. Marez, and S. Parameswaran, “Detecting aerial objects: Drones, birds, and helicopters,” in Artificial Intelligence and Machine Learning in Defense Applications III , J. Dijk, Ed. Online Only, Spain: SPIE, Sep. 2021, p. 18

    2021

  25. [33]

    Small Flying Object Detection and Tracking in Digital Airport Tower through Spatial- Temporal ConvNets,

    P. Thai, S. Alam, N. Lilith, and B. Nguyen, “Small Flying Object Detection and Tracking in Digital Airport Tower through Spatial- Temporal ConvNets,” In Review, Preprint, Mar. 2023

    2023

  26. [34]

    Spatio-Temporal Semantic Segmentation for Drone Detection,

    C. Craye and S. Ardjoune, “Spatio-Temporal Semantic Segmentation for Drone Detection,” in 2019 16th IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS). Taipei, Taiwan: IEEE, Sep. 2019, pp. 1–5

    2019

  27. [35]

    TransVisDrone: Spatio-Temporal Transformer for Vision-based Drone-to-Drone Detec- tion in Aerial Videos,

    T. Sangam, I. R. Dave, W. Sultani, and M. Shah, “TransVisDrone: Spatio-Temporal Transformer for Vision-based Drone-to-Drone Detec- tion in Aerial Videos,” Aug. 2023

    2023

  28. [36]

    On Rendering Synthetic Images for Training an Object Detector,

    A. Rozantsev, V . Lepetit, and P. Fua, “On Rendering Synthetic Images for Training an Object Detector,” Computer Vision and Image Under- standing, vol. 137, pp. 24–37, Aug. 2015

    2015

  29. [37]

    UA V detection with a dataset augmented by domain randomization,

    D. Marez, S. Borden, and L. Nans, “UA V detection with a dataset augmented by domain randomization,” in Geospatial Informatics X, vol. 11398. SPIE, May 2020, pp. 39–50

    2020

  30. [38]

    Using Images Rendered by PBRT to Train Faster R-CNN for UA V Detection,

    J. Peng, C. Zheng, T. Cui, Y . Cheng, and L. Si, “Using Images Rendered by PBRT to Train Faster R-CNN for UA V Detection,” in

  31. [39]

    International Conference in Central Europe on Computer Graphics, Visualization and Computer Vision’2017 , 2018

    2017

  32. [40]

    Quantifying the Simulation–Reality Gap for Deep Learning-Based Drone Detection,

    T. R. Dieter, A. Weinmann, S. J ¨ager, and E. Brucherseifer, “Quantifying the Simulation–Reality Gap for Deep Learning-Based Drone Detection,” Electronics, vol. 12, no. 10, p. 2197, Jan. 2023

    2023

  33. [41]

    DronePose: The identification, segmentation, and orientation detection of drones via neural networks,

    S. Scholes, A. Ruget, G. Mora-Martin, F. Zhu, I. Gyongy, and J. Leach, “DronePose: The identification, segmentation, and orientation detection of drones via neural networks,” arXiv:2112.05488 [cs], Dec. 2021

    2021 arXiv

  34. [42]

    Scarce Data Driven Deep Learning of Drones via Generalized Data Distribution Space,

    C. Li, S. C. Sun, Z. Wei, A. Tsourdos, and W. Guo, “Scarce Data Driven Deep Learning of Drones via Generalized Data Distribution Space,” arXiv:2108.08244 [cs], Aug. 2021

    2021 arXiv

  35. [43]

    Drone Detection Using Depth Maps,

    A. Carrio, S. Vemprala, A. Ripoll, S. Saripalli, and P. Campoy, “Drone Detection Using Depth Maps,” Aug. 2018

    2018

  36. [44]

    Drone Model Classification Using Convolutional Neural Network Trained on Synthetic Data,

    M. Wisniewski, Z. A. Rana, and I. Petrunin, “Drone Model Classification Using Convolutional Neural Network Trained on Synthetic Data,” p. 20, 2022

    2022

  37. [45]

    Drone Model Identification by Convolutional Neural Network from Video Stream,

    ——, “Drone Model Identification by Convolutional Neural Network from Video Stream,” in 2021 IEEE/AIAA 40th Digital Avionics Systems Conference (DASC). San Antonio, TX, USA: IEEE, Oct. 2021, pp. 1–8

    2021

  38. [46]

    Blender - a 3D modelling and rendering package

    B. Foundation, “Blender - a 3D modelling and rendering package.”

  39. [47]

    Microsoft COCO: Common Objects in Context,

    T.-Y . Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick, J. Hays, P. Perona, D. Ramanan, C. L. Zitnick, and P. Doll ´ar, “Microsoft COCO: Common Objects in Context,” Feb. 2015

    2015

  40. [48]

    PyTorch: An Imperative Style, High- Performance Deep Learning Library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “PyTorch: An Imperative Style, High- P...

  41. [49]

    Vision/torchvision/models/detection/faster rcnn.py at main · pytorch/vision,

    “Vision/torchvision/models/detection/faster rcnn.py at main · pytorch/vision,” https://github.com/pytorch/vision/blob/main/ torchvision/models/detection/faster rcnn.py

  42. [50]

    On the Impact of Lossy Image and Video Compression on the Performance of Deep Convolutional Neural Network Architectures,

    M. Poyser, A. Atapour-Abarghouei, and T. P. Breckon, “On the Impact of Lossy Image and Video Compression on the Performance of Deep Convolutional Neural Network Architectures,” Jul. 2020

    2020

  43. [51]

    Jung, “Imgaug,” Nov

    A. Jung, “Imgaug,” Nov. 2023

    2023

  44. [52]

    Adaptive Inattentional Framework for Video Object Detection With Reward-Conditional Training,

    A. Rodriguez-Ramos, J. Rodriguez-Vazquez, C. Sampedro, and P. Cam- poy, “Adaptive Inattentional Framework for Video Object Detection With Reward-Conditional Training,” IEEE Access, vol. 8, pp. 124 451– 124 466, 2020

    2020

  45. [53]

    Drone-vs-Bird Detection Challenge at IEEE A VSS2019,

    A. Coluccia, A. Fascista, A. Schumann, L. Sommer, M. Ghenescu, A. O. Avenue, and T. Piatrik, “Drone-vs-Bird Detection Challenge at IEEE A VSS2019,” p. 7

  46. [54]

    Drone vs. Bird Detection: Deep Learning Algorithms and Results from a Grand Challenge,

    A. Coluccia, A. Fascista, A. Schumann, L. Sommer, A. Dimou, D. Zarpalas, M. M ´endez, D. de la Iglesia, I. Gonz ´alez, J.-P. Mercier, G. Gagn´e, A. Mitra, and S. Rajashekar, “Drone vs. Bird Detection: Deep Learning Algorithms and Results from a Grand Challenge,” Sensors, vol. ...

    2021

  47. [55]

    Anti-UA V: A Large Multi-Modal Benchmark for UA V Tracking,

    N. Jiang, K. Wang, X. Peng, X. Yu, Q. Wang, J. Xing, G. Li, J. Zhao, G. Guo, and Z. Han, “Anti-UA V: A Large Multi-Modal Benchmark for UA V Tracking,”arXiv:2101.08466 [cs], Feb. 2021

    2021 arXiv

  48. [56]

    Torch.manual seed(3407) is all you need: On the influence of random seeds in deep learning architectures for computer vision,

    D. Picard, “Torch.manual seed(3407) is all you need: On the influence of random seeds in deep learning architectures for computer vision,” May 2023

    2023

  49. [57]

    N. N. Taleb, The Black Swan: The Impact of the Highly Improbable . Random house, 2007, vol. 2

    2007

Pith tools

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