Pith. sign in

REVIEW 2 major objections 5 minor 37 references

Federated training of lightweight YOLO detectors nearly matches centralized accuracy on drone imagery while keeping every image local.

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 ↗

FedAvg on non-IID KIIT-MiTA drone imagery recovers most centralized YOLO nano mAP while keeping images local, with YOLO26 nano gaining ~53% and ~68% relative mAP over single-drone baselines.

T0 review reviewed 2026-07-12 challenge →

load-bearing objection Solid controlled FedAvg+YOLO study on KIIT-MiTA: FL nearly matches centralized and beats single-drone; scope is four simulated nodes on one dataset, not a fleet proof. the 2 major comments →

arxiv 2607.02636 v1 pith:FTP4RAQJ submitted 2026-07-02 cs.LG cs.AIcs.CVcs.DC

Federated Learning for Object Detection: Enabling Collaborative Drone Learning Without Centralizing Data

Daniel M. Jimenez-Gutierrez , Enrique Zuazua , Georgios Kellaris , Joaquin del Rio , Oleksii Sliusarenko , Xabi Uribe-Etxebarria This is my paper
classification cs.LG cs.AIcs.CVcs.DC
keywords federated learningobject detectiondrone imageryYOLOprivacy-preserving trainingnon-IID dataedge AIKIIT-MiTA
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.

The reading

This paper asks whether drones can pool knowledge for object detection without ever shipping their raw aerial images to a central server. On the KIIT-MiTA military-target drone dataset, the authors compare three regimes under one protocol: each drone training alone, full data centralization, and standard federated averaging across four non-IID nodes. Federated training recovers most of the centralized gain and far outperforms isolated single-drone models. The lightest architecture they test, YOLO26 nano, reaches 0.71 mAP@0.50 and 0.47 mAP@0.50:0.95 after federation—relative lifts of roughly 53% and 68% over single-drone training—while remaining only a few points behind the centralized oracle and still running at about 99 frames per second with 6.1 GFLOPs. A further stress test under a 600 Kbps bandwidth cap shows accuracy holds while training time rises modestly and still finishes faster than centralized CPU training. The practical message is that privacy-preserving collaboration can deliver near-centralized detection quality for distributed drone fleets without moving the imagery.

Core claim

Under a shared experimental protocol on KIIT-MiTA, standard Federated Averaging of lightweight YOLO nano detectors remains close to centralized training and dramatically improves over single-drone training, with YOLO26 nano achieving relative gains of 52.89% in mAP@0.50 and 67.80% in mAP@0.50:0.95 while preserving real-time inference cost and data locality.

What carries the argument

Horizontal federated averaging (FedAvg) of YOLO nano detectors: each of four nodes runs 10 local epochs on its private non-IID partition, then shares only weights for server averaging over 10 rounds, so raw images never leave the node.

Load-bearing premise

That results from four Dirichlet-partitioned nodes on one military drone dataset are enough to support claims of scalable high performance for real multi-drone fleets with stronger heterogeneity and other aerial datasets.

What would settle it

Repeat the identical Single-drone / Centralized / Federated protocol on a second public UAV detection benchmark (or with substantially more nodes and stronger non-IID skew) and check whether federated mAP still stays within a few points of centralized while beating single-node baselines by large relative margins.

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

If this is right

  • Distributed drone fleets can improve a shared detector by exchanging only model weights, avoiding raw-image transfer costs and many privacy or regulatory barriers.
  • Lightweight YOLO nano models remain real-time after federation (≈99 FPS, 6.1 GFLOPs), so the privacy gain does not force heavier onboard hardware.
  • Even under a 600 Kbps bidirectional cap, federated training preserves accuracy and still finishes faster than centralized CPU training that would move the images.
  • Per-class gains appear across all seven military categories, so the benefit is not confined to a few dominant object types.

Where Pith is reading between the lines

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

  • The same FedAvg-plus-nano-YOLO recipe could be stress-tested on civilian UAV tasks (search-and-rescue, infrastructure inspection) where class imbalance and altitude variation differ from military targets.
  • Because the paper already notes that model updates can leak information, pairing the pipeline with secure aggregation or differential privacy is a natural next engineering step for high-security deployments.
  • If larger fleets introduce stragglers or intermittent links, the observed training-time sensitivity to bandwidth suggests partial-participation or compression variants would be the first place to look.
Share X Bluesky LinkedIn Reddit HN

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

2 major / 5 minor

Summary. The paper presents a controlled empirical study of Federated Learning (FedAvg via the Sherpa.ai platform) for drone-based object detection on the KIIT-MiTA military aerial dataset. It compares Single-drone, Centralized, and Federated training under a shared protocol across four YOLO nano detectors (YOLOv5/v8/11/26), using a four-node non-IID Dirichlet partition (α=1) of the training split and the official held-out test set. Primary metrics are mAP@0.50 and mAP@0.50:0.95 (10 seeds, mean±std), with FPS/GFLOPs and a 600 Kbps bandwidth-constrained CPU run. The central empirical claim is that Federated training remains close to Centralized while substantially outperforming Single-drone; for YOLO26 nano, Federated mAP@0.50 / mAP@0.50:0.95 are 0.7137 / 0.4695 versus Single-drone 0.4668 / 0.2798 (relative gains ~52.89% / 67.80%) and Centralized 0.7322 / 0.4959. Contributions are framed as a pipeline and comparative evaluation rather than a new detector or FL optimizer.

Significance. If the reported numbers hold under the stated protocol, the work provides a clear, reproducible demonstration that standard FedAvg can recover most of the multi-source benefit for lightweight aerial detectors without pooling raw images—an operationally relevant privacy–utility trade-off for drone fleets. Strengths include multi-seed reporting (Tables 1–2), matched training budget (100 epochs), per-class breakdowns, efficiency metrics suitable for edge deployment, and a limited-bandwidth stress test (§6.3.4). The contribution is incremental rather than algorithmic: it does not introduce a new detector or non-IID FL method, and evaluation is confined to one dataset and K=4 simulated nodes. Within that scope it is a useful application study for the FL-for-UAV community, especially given honest limitations in §7.

major comments (2)
  1. Abstract and §8 state that FL enables “scalable, high-performing… object detection across distributed drone fleets,” but the evidence is a single dataset (KIIT-MiTA), K=4 nodes, fixed Dirichlet α=1, and FedAvg only (§5.2, §6.2, Table 1). §7 correctly notes that larger fleets, stronger heterogeneity, and other UAV datasets are untested. The experimental core (Federated ≈ Centralized ≫ Single-drone under this protocol) is supported; the abstract/conclusion language should be aligned with that scope so the central claim is not overstated relative to the evidence.
  2. §5.2: non-IID partitioning assigns each multi-object image by its dominant class (most frequent boxes), then applies Dirichlet(α=1). The manuscript acknowledges this is approximate for detection, but does not report sensitivity to α, alternative multi-label partitions, or how often secondary classes co-occur across nodes. Because the paper’s privacy–utility narrative rests on “realistic” non-IID drone heterogeneity, a short sensitivity check (e.g., α∈{0.1,1,10} or class-instance-level stats) would make the load-bearing non-IID assumption more defensible without changing the study design.
minor comments (5)
  1. Figure 1 is hard to read (overlapping percentages, unclear map legend) and does not clearly connect geographic regions to the four experimental nodes; either clarify or move to supplementary material.
  2. §6.2: Federated runs use AWS T4 GPUs while Centralized/Single-drone use an RTX A4000; FPS is fairly measured on shared validation hardware, but training-time comparisons in §6.3.4 (CPU-only) should state more explicitly that absolute hours are not cross-hardware comparable to the GPU training setup.
  3. Related work (§2.2.2) cites Lu & Sun [29,30] on federated UAV detection and non-IID mitigations; a brief note on why plain FedAvg was chosen over their contrastive/focal variants would help readers place the baseline choice.
  4. Typographical/formatting: abstract “52.89%and67 .80%”; “SHERP A.AI federa ted” in Figure 1 caption area; occasional spacing issues in equations (e.g., mAP formulas in §6.1).
  5. §3.2.1 mentions VFL only to set it aside; a single sentence on why HFL matches multi-drone same-sensor imagery is enough—consider shortening to reduce tutorial bulk relative to the experimental contribution.

Circularity Check

0 steps flagged

No significant circularity: controlled empirical FL comparison on held-out mAP, not a self-referential derivation.

full rationale

This paper is an experimental systems study, not a first-principles derivation. Its central claim is that FedAvg FL on a 4-node non-IID partition of KIIT-MiTA yields mAP close to centralized training and substantially above single-drone baselines for YOLO nano models (Table 1: YOLO26 nano Federated 0.7137/0.4695 vs Single 0.4668/0.2798 vs Centralized 0.7322/0.4959). Metrics are standard COCO-style mAP computed on the official held-out test set after multi-seed runs with a matched epoch budget; relative gains are arithmetic from those measured numbers. The Dirichlet partition (α=1, dominant-class), FedAvg rounds, augmentations, and model choices are experimental controls, not definitions that force Federated mAP to equal Centralized by construction. Self-citations (FedArtML for partitioning, Sherpa.ai platform) describe tooling and are not load-bearing uniqueness theorems or smuggled ansatzes that underwrite the accuracy claims. No fitted parameter is renamed as a prediction of a related quantity, and no equation equates the target result to its inputs. Limitations (§7) openly note the narrow scope (one dataset, four nodes). The work is self-contained against its own baselines; score 0 is the correct honest finding.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

Central empirical claim rests on standard FL/CV machinery plus several experimental design choices (partition, training budget, platform). No new physical entities. Free parameters are the usual ML knobs that define the non-IID regime and optimization path; axioms are domain assumptions that the simulated four-drone setup and FedAvg protocol speak to real collaborative drone learning.

free parameters (5)
  • Dirichlet concentration α for non-IID split = 1
    α=1 sets moderate class imbalance across four nodes; different α would change heterogeneity and thus the Single vs Federated gap that drives the headline gains.
  • Local epochs per round × communication rounds = 10 local epochs, 10 rounds
    10×10 defines total federated training budget matched to 100 centralized epochs; this schedule is a free experimental choice that affects convergence and reported mAP.
  • Ultralytics augmentation hyperparameters = mosaic=1.0, mixup=0.05, scale=0.50, HSV (0.015,0.70,0.40), flip_h=0.5
    mosaic, mixup, HSV, scale, etc., selected after preliminary experiments for best validation; they influence absolute mAP for all regimes.
  • Number of federated nodes K = 4
    K=4 is fixed by design; scalability claims depend on this small fleet size.
  • Bandwidth cap for stress test = 600 Kbps
    600 Kbps bidirectional is a hand-chosen conservative stress level for the timing experiment.
axioms (4)
  • domain assumption FedAvg aggregation of YOLO weights yields a meaningful global detector under the paper's non-IID image partitions.
    Invoked throughout §3.2 and §5.1; standard FL assumption, not proved here for multi-object aerial detection.
  • ad hoc to paper Dominant-class Dirichlet assignment of multi-object images adequately models realistic drone data heterogeneity.
    §5.2 explicitly notes multi-label complexity and uses dominant class as an approximation; load-bearing for the non-IID claim.
  • domain assumption KIIT-MiTA official splits and military target classes are a valid proxy for safety-critical drone perception evaluation.
    Dataset choice §4.1; generalization limits acknowledged in §7.
  • domain assumption Matching total epoch budget (100) across Single/Centralized/Federated makes the three regimes comparable.
    §6.2 reproducibility details; common experimental convention but not uniquely justified.

reviewed 2026-07-12 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Federated Learning for Object Detection: Enabling Collaborative Drone Learning Without Centralizing Data." pith.science (2026). https://pith.science/paper/FTP4RAQJ

@misc{pith2026260702636,
  author       = {Pith},
  title        = {Pith review of: Federated Learning for Object Detection: Enabling Collaborative Drone Learning Without Centralizing Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FTP4RAQJ}},
  note         = {Machine review of arXiv:2607.02636}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Object detection is a fundamental capability for AI-driven perception in safety-critical drone and edge-vision systems, including disaster response, operational security environments, infrastructure monitoring and defense applications. Robust model performance in such environments depends on large, continuously updated datasets. However, training high-performing detectors typically requires centralizing aerial imagery, which raises privacy, regulatory, storage, and bandwidth challenges. This is especially problematic in distributed drone deployments, where visual data is generated onboard and is often impractical or undesirable to transfer to a centralized infrastructure. In this work, we apply Federated Learning (FL) for object detection, enabling drones to improve a shared model while keeping image data local and private. We implement a federated object detection pipeline using the Sherpa.ai FL platform on the KIIT-MiTA dataset, and compare it with Single-drone and Centralized baselines using mean Average Precision (mAP) at IoU thresholds of 0.50 and 0.50-0.95. In our experiments, the proposed FL approach remains close to Centralized training while dramatically improving over Single-drone training, with the best lightweight model (YOLO26 nano), suitable for deployment even on very limited edge infrastructure, achieving relative gains of 52.89% and 67.80% in mAP@0.50 and mAP@0.50:0.95, respectively. These results show that FL enables scalable, high-performing, and privacy-preserving object detection across distributed drone fleets without data centralization.

Figures

Figures reproduced from arXiv: 2607.02636 by Daniel M. Jimenez-Gutierrez, Enrique Zuazua, Georgios Kellaris, Joaquin Del Rio, Oleksii Sliusarenko, Xabi Uribe-Etxebarria.

Figure 1
Figure 1. Figure 1: Global map illustrating drone mAP@0.50 for the best federated YOLO model. 1 arXiv:2607.02636v1 [cs.LG] 2 Jul 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the object detection process in aerial drone imagery. 2.2 Related Work This subsection reviews the lines of work most relevant to this paper, namely, centralized and federated approaches to object detection in aerial and drone-based settings. We emphasize prior studies on real-time detectors, aerial-object-detection benchmarks, and privacy-preserving distributed training. This context helps… view at source ↗
Figure 3
Figure 3. Figure 3: Example images from the KIIT-MiTA dataset [16] showing some military target classes in an aerial drone scene. In our study, we utilize all annotated categories during training, resulting in a total of seven classes: artillery, missile, radar, multiple rocket launcher (M. R. Launcher), soldier, tank, and vehicle. 4.2 Preprocessing of the Dataset The preprocessing pipeline was adapted to the training workflo… view at source ↗
Figure 4
Figure 4. Figure 4: Proposed architecture for centralized training. 5 Proposed Privacy-Preserving Solution through FL In this section, we present the FL architecture and the construction of the federated nodes used in our experimental setup. 5.1 FL Architecture We implement FL as depicted in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Proposed architecture for federated training. When applying this framework to object detection, local training involves optimizing a compound loss function encompassing classification, objectness, and bounding box regression. Given the model’s computational demands, the efficiency of local computation and communication overhead becomes critical. 5.2 Creation of Nodes To simulate a realistic federated setti… view at source ↗
Figure 6
Figure 6. Figure 6: Dominant-class distribution across drones for the non-IID partition used. Each stacked bar shows the proportion of images assigned to each class according to the most frequent object class in the image. 6 Experiments In this section, we describe the experimental setup and the training configurations considered in this work, and present the main results for drone-based object detection. We evaluate model pe… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative examples of model predictions on test images. Detected objects are shown with their predicted bounding boxes, class labels, and confidence scores. Precision = T P T P + F P (6.1) In the context of object detection, a predicted bounding box is considered a true positive if its IoU with a ground truth box is greater than a given threshold (i.e., τ = 0.50), and the predicted class matches the grou… view at source ↗
Figure 8
Figure 8. Figure 8: Mean metric values comparing Single-drone, Centralized, and Federated training settings for the best model (YOLO26 nano). Among the evaluated models, YOLO26 nano achieves the best overall accuracy. In the Federated setting, it obtains the highest mAP@0.50, with a value of 0.7137, and the highest mAP@0.50:0.95, with a value of 0.4695. Compared with the Single-drone baseline, this represents an absolute gain… view at source ↗
Figure 9
Figure 9. Figure 9: Per-class mean mAP@0.50 values comparing Single-drone, Centralized, and Federated training settings for the best model (YOLO26 nano). 15 [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Training time in hours for the YOLO11 Nano model across centralized unconstrained, federated unconstrained, and federated bandwidth-constrained (600 Kbps) scenarios. The results show that the performance remained comparable to the full-bandwidth baseline, indicating that limited connectivity had a negligible impact on model performance. The constrained run completed federated training in 3 hours 20 minute… 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

37 extracted references · 1 canonical work pages

  1. [1]

    ‘‘Faster r-cnn: Towards real-time object detection with region proposal networks’’

    Shaoqing Ren et al. ‘‘Faster r-cnn: Towards real-time object detection with region proposal networks’’. In:Advances in neural information processing systems28 (2015)

    2015

  2. [2]

    ‘‘You only look once: Unified, real-time object detection’’

    Joseph Redmon et al. ‘‘You only look once: Unified, real-time object detection’’. In:Proceedings of the IEEE conference on computer vision and pattern recognition. 2016, pp. 779–788

    2016

  3. [3]

    ‘‘Yolov3: An incremental improvement’’

    Joseph Redmon and Ali Farhadi. ‘‘Yolov3: An incremental improvement’’. In:arXiv preprint arXiv:1804.02767(2018)

    Pith/arXiv arXiv 2018

  4. [4]

    Version 8.0.0

    Glenn Jocher, Ayush Chaurasia, and Jing Qiu.Ultralytics YOLOv8. Version 8.0.0. 2023.url:https: //github.com/ultralytics/ultralytics

    2023

  5. [5]

    ‘‘Focal loss for dense object detection’’

    Tsung-Yi Lin et al. ‘‘Focal loss for dense object detection’’. In:Proceedings of the IEEE international conference on computer vision. 2017, pp. 2980–2988

    2017

  6. [6]

    ‘‘Vision meets drones: A challenge’’

    Pengfei Zhu et al. ‘‘Vision meets drones: A challenge’’. In:arXiv preprint arXiv:1804.07437(2018)

    Pith/arXiv arXiv 2018

  7. [7]

    ‘‘Object detection with deep learning: A review’’

    Zhong-Qiu Zhao et al. ‘‘Object detection with deep learning: A review’’. In:IEEE transactions on neural networks and learning systems30.11 (2019), pp. 3212–3232

    2019

  8. [8]

    ‘‘Generalized uav object detection via frequency domain disentanglement’’

    Kunyu Wang et al. ‘‘Generalized uav object detection via frequency domain disentanglement’’. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2023, pp. 1064– 1073

    2023

  9. [9]

    Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation). Apr. 27, 2016. url:https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=...

    2016

  10. [10]

    Ley Orgánica 3/2018, de 5 de diciembre, de Protección de Datos Personales y garantía de los derechos digitales. Dec. 6, 2018.url:https://www.boe.es/eli/es/lo/2018/12/05/3. [11]Health Insurance Portability and Accountability Act of 1996. Aug. 21, 1996. [12]California Consumer Privacy Act. 2018

    2018

  11. [11]

    ‘‘Communication-efficient learning of deep networks from decentralized data’’

    Brendan McMahan et al. ‘‘Communication-efficient learning of deep networks from decentralized data’’. In:Artificial intelligence and statistics. PMLR. 2017, pp. 1273–1282

    2017

  12. [12]

    ‘‘Federated machine learning: Concept and applications’’

    Qiang Yang et al. ‘‘Federated machine learning: Concept and applications’’. In:ACM Transactions on Intelligent Systems and Technology (TIST)10.2 (2019), pp. 1–19

    2019

  13. [13]

    ‘‘Fedvision: An online visual object detection platform powered by federated learning’’

    Yang Liu et al. ‘‘Fedvision: An online visual object detection platform powered by federated learning’’. In:Proceedings of the AAAI conference on artificial intelligence. Vol. 34. 08. 2020, pp. 13172–13179

    2020

  14. [14]

    https://www.kaggle.com/datasets/sudipchakrabarty/kiit- mita

    Sudip Chakrabarty.KIIT-MiTA. https://www.kaggle.com/datasets/sudipchakrabarty/kiit- mita. Kaggle dataset, accessed 2026-04-22. 2025

    2026

  15. [15]

    ‘‘Rich feature hierarchies for accurate object detection and semantic segmentation’’

    Ross Girshick et al. ‘‘Rich feature hierarchies for accurate object detection and semantic segmentation’’. In:Proceedings of the IEEE conference on computer vision and pattern recognition. 2014, pp. 580–587

    2014

  16. [16]

    ‘‘Fast r-cnn’’

    Ross Girshick. ‘‘Fast r-cnn’’. In:Proceedings of the IEEE international conference on computer vision. 2015, pp. 1440–1448

    2015

  17. [17]

    ‘‘Ssd: Single shot multibox detector’’

    Wei Liu et al. ‘‘Ssd: Single shot multibox detector’’. In:Computer Vision--ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11--14, 2016, Proceedings, Part I 14. Springer. 2016, pp. 21–37

    2016

  18. [18]

    Version 7.0

    Glenn Jocher.Ultralytics YOLOv5. Version 7.0. 2020.doi: 10.5281/zenodo.3908559.url: https: //github.com/ultralytics/yolov5

    work page doi:10.5281/zenodo.3908559.url: 2020

  19. [19]

    Version 11.0.0

    Glenn Jocher and Jing Qiu.Ultralytics YOLO11. Version 11.0.0. 2024.url:https://github.com/ ultralytics/ultralytics

    2024

  20. [20]

    Version 26.0.0

    Glenn Jocher and Jing Qiu.Ultralytics YOLO26. Version 26.0.0. 2026.url:https://github.com/ ultralytics/ultralytics

    2026

  21. [21]

    ‘‘Fcos: Fully convolutional one-stage object detection’’

    Zhi Tian et al. ‘‘Fcos: Fully convolutional one-stage object detection’’. In:Proceedings of the IEEE/CVF international conference on computer vision. 2019, pp. 9627–9636

    2019

  22. [22]

    ‘‘End-to-end object detection with transformers’’

    Nicolas Carion et al. ‘‘End-to-end object detection with transformers’’. In:European conference on computer vision. Springer. 2020, pp. 213–229

    2020

  23. [23]

    ‘‘Recent advances for aerial object detection: A survey’’

    Jiaxu Leng et al. ‘‘Recent advances for aerial object detection: A survey’’. In:ACM Computing Surveys 56.12 (2024), pp. 1–36

    2024

  24. [24]

    ‘‘Microsoft coco: Common objects in context’’

    Tsung-Yi Lin et al. ‘‘Microsoft coco: Common objects in context’’. In:Computer vision--ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13. Springer. 2014, pp. 740–755

    2014

  25. [25]

    ‘‘The pascal visual object classes (voc) challenge’’

    Mark Everingham et al. ‘‘The pascal visual object classes (voc) challenge’’. In:International journal of computer vision88 (2010), pp. 303–338

    2010

  26. [26]

    ‘‘Drones in defense: Real-time vision-based military target surveillance and tracking’’

    Sudip Chakrabarty et al. ‘‘Drones in defense: Real-time vision-based military target surveillance and tracking’’. In:2025 3rd International Conference on Intelligent Systems, Advanced Computing and Communication (ISACC). IEEE. 2025, pp. 508–513

    2025

  27. [27]

    ‘‘Development of real-time unmanned aerial vehicle urban object detection system with federated learning’’

    You-Ru Lu and Dengfeng Sun. ‘‘Development of real-time unmanned aerial vehicle urban object detection system with federated learning’’. In:Journal of Aerospace Information Systems21.7 (2024), pp. 547–553

    2024

  28. [28]

    ‘‘Improving Federated Learning UAV Urban Object Detection System via Data Heterogeneity Mitigation’’

    You-Ru Lu and Dengfeng Sun. ‘‘Improving Federated Learning UAV Urban Object Detection System via Data Heterogeneity Mitigation’’. In:Journal of Aerospace Information Systems22.11 (2025), pp. 930–937

    2025

  29. [29]

    ‘‘Federated learning with non-iid data: A survey’’

    Zili Lu et al. ‘‘Federated learning with non-iid data: A survey’’. In:IEEE Internet of Things Journal (2024)

    2024

  30. [30]

    ‘‘Leaf: A benchmark for federated settings’’

    Sebastian Caldas et al. ‘‘Leaf: A benchmark for federated settings’’. In:arXiv preprint arXiv:1812.01097 (2018)

    Pith/arXiv arXiv 2018

  31. [31]

    ‘‘Federated learning: Challenges, methods, and future directions’’

    Tian Li et al. ‘‘Federated learning: Challenges, methods, and future directions’’. In:IEEE signal processing magazine37.3 (2020), pp. 50–60

    2020

  32. [32]

    ‘‘Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data’’

    Micah J Sheller et al. ‘‘Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data’’. In:Scientific reports10.1 (2020), p. 12598

    2020

  33. [33]

    ‘‘Federated learning on non-IID data: A survey’’

    Hangyu Zhu et al. ‘‘Federated learning on non-IID data: A survey’’. In:Neurocomputing465 (2021), pp. 371–390. 19 FL for Object Detection: Enabling Collaborative Drone Learning Without Centralizing Data

    2021

  34. [34]

    2023.url:https://github.com/ ultralytics/ultralytics

    Glenn Jocher, Jing Qiu, and Ayush Chaurasia.Ultralytics YOLO. 2023.url:https://github.com/ ultralytics/ultralytics

    2023

  35. [35]

    ‘‘FedArtML: A Tool to Facilitate the Generation of Non-IID Datasets in a Controlled Way to Support Federated Learning Research’’

    Daniel M Jimenez-Gutierrez et al. ‘‘FedArtML: A Tool to Facilitate the Generation of Non-IID Datasets in a Controlled Way to Support Federated Learning Research’’. In:IEEE Access(2024)

    2024

  36. [36]

    ‘‘Client selection and bandwidth allocation in wireless federated learning networks: A long-term perspective’’

    Jie Xu and Heqiang Wang. ‘‘Client selection and bandwidth allocation in wireless federated learning networks: A long-term perspective’’. In:IEEE Transactions on Wireless Communications20.2 (2020), pp. 1188–1200

    2020

  37. [37]

    ‘‘Wireless federated learning over resource-constrained networks: Digital versus analog transmissions’’

    Jiacheng Yao et al. ‘‘Wireless federated learning over resource-constrained networks: Digital versus analog transmissions’’. In:IEEE Transactions on Wireless Communications23.10 (2024), pp. 14020– 14036. 20

    2024

This paper was first reviewed by grok-4.5 on July 12, 2026.