Pith. sign in

REVIEW 4 major objections 5 minor 52 references

FEDEXCHANGE: Bridging the Domain Gap in Federated Object Detection for Free

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Federated object detection across heterogeneous visual domains can be improved by having the server exchange client-trained decoder weights on alternating rounds, without adding any client-side computation.

desk verdict A plausible server-side exchange trick for federated detection, but the load-bearing clustering assumption is never validated and the headline gains are cherry-picked. read the letter →

arxiv 2509.10503 v1 pith:TEU7JOLQ submitted 2025-09-01 cs.LG cs.AIcs.CV

Haolin Yuan , Jingtao Li , Weiming Zhuang , Chen Chen , Lingjuan Lyu This is my paper
classification cs.LGcs.AIcs.CV
keywords federatedlearningobjectdetectiondomainadaptationmodelexchangeserver-sideclusteringfoundationedgecomputingcross-domaingeneralization
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

FEDEXCHANGE argues that the cross-domain gap in federated object detection can be bridged without asking clients to do any extra work. Clients keep a frozen foundation-model backbone and train only a small detection decoder on their local domain; the server alternates between aggregating these decoders and exchanging them among clients based on cosine-distance clustering. Because all exchange logic runs server-side, clients get cross-domain knowledge at the same local compute cost as plain federated averaging. The paper reports that this simple mechanism improves mean average precision on every benchmark it tests, with the largest gains on the weakest or most underrepresented domains, such as rainy-weather driving data. The mechanism rests on the assumption that decoder weights encode the local visual domain, which the paper notes breaks down when local data is extremely scarce.

What carries the argument

The central mechanism is the dynamic model exchange protocol: after local decoder training, the server either performs standard global aggregation or, on exchange rounds, clusters all uploaded decoder weight vectors into two clusters by average-linkage cosine distance, performs in-cluster shuffling, and flips cluster indices to produce cross-cluster exchanges. The work this does is to let each local decoder continue training on weights that originated in a different visual domain, thereby transferring domain-invariant features between clients without data movement.

What would settle it

On a benchmark with known domain labels, such as SODA10M clear/overcast/rainy, compute the two-cluster cosine-distance partition of uploaded decoders and measure its agreement with the true weather labels. If agreement is no better than chance, or if the reported mAP gains survive when exchange uses randomly assigned cluster labels, then the mechanism is not doing domain-aware transfer.

Watch

Extended reading notes

Core claim

The paper proposes that the small learnable decoder in a frozen-backbone federated object detection system can act as a carrier of domain knowledge. In exchange rounds, the server clusters the uploaded decoders into two groups using average-linkage agglomerative clustering on cosine distance between flattened weight vectors, then shuffles decoders within each cluster and swaps decoders across clusters, ensuring no client receives the same decoder it had before. This mimics multiple centralized training processes running in parallel without sharing any private data. Across person-detection and driving benchmarks, FEDEXCHANGE reports higher mean average precision than aggregation-only or local

Load-bearing premise

The uploaded decoder weights carry enough information about a client's visual domain that cosine distance between weight vectors meaningfully separates domains; when each client has only 10% of its data, the benefit over plain aggregation disappears.

Editorial extensions

If this is right

  • Cross-domain federated detection can improve without local regularization or extra local memory; server-side weight exchange alone is sufficient to lift utility.
  • The largest gains appear on clients with the weakest baseline accuracy or fewest samples, so the method reduces the disparity between well-represented and underrepresented domains.
  • Aggregation frequency is a forgiving hyperparameter: utility stays nearly flat from aggregating every 2 rounds to every 50 rounds, so latency and bandwidth can be traded off against accuracy freely.
  • With half the local training data, FEDEXCHANGE beats standard federated averaging trained on the full dataset, suggesting the exchange acts as an implicit data-efficient regularizer.
  • Because only the lightweight decoder is uploaded and exchanged, communication and GPU costs remain close to plain federated averaging, unlike regularization-based baselines that create extra tensors locally.

Reading between the lines

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

  • The paper does not test this, but if decoder weights are a reliable domain fingerprint, the same server-side exchange idea could transfer to other frozen-backbone dense prediction tasks such as segmentation, and could let the server estimate client domains without ever seeing data.
  • A testable extension is to replace the fixed two-cluster partition with an adaptive cluster count and a rotation schedule over all clusters; the near-flat performance across aggregation frequencies suggests aggressive exchange may be safe.
  • The biggest jumps occur on the smallest or hardest domains, hinting that exchange acts as model-level augmentation for minority domains; monitoring cluster stability over rounds could serve as a deployment-time signal for when exchange stops helping.
  • The paper reports that distance-based exchange beats random and round-robin exchange in its ablation, but it never checks whether the cosine-distance clusters actually match the true domain labels; measuring cluster purity against known domains would directly test the mechanism's core premise.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes FEDEXCHANGE, a federated object detection framework in which a frozen foundation-model backbone is combined with a learnable decoder on each client. The server alternates between standard FedAvg-style aggregation and a server-side dynamic model exchange: decoders are clustered by pairwise cosine distance (using average linkage, fixed to two clusters), shuffled within clusters, and swapped across clusters. The claimed benefit is improved cross-domain object detection without additional local compute or communication overhead. Experiments are reported on four person-detection datasets and two driving datasets (BDD100K, SODA10M), compared against FedAvg, FedProx, and FedSTO, plus ablations of aggregation frequency, exchange strategy, and reduced training-set size.

Significance. The practical setting studied here—frozen foundation-model backbone with a learnable detection decoder on edge clients—is timely, and the idea of moving all additional operations to the server is attractive: if it works, FEDEXCHANGE would improve cross-domain generalization at zero extra local cost. The paper also includes a reasonable set of benchmark families and ablations, and the authors are candid about the assumption that decoder weights encode domain structure and about the method's failure at 10% data. However, the central mechanism is not directly validated, and the headline quantitative claim rests on a single run in one domain. Because random model exchange is shown in Table 6 to be worse than FedAvg, the method's advantage depends entirely on the clustering step being domain-meaningful. That load-bearing assumption is asserted rather than demonstrated, and the single-seed results give no sense of variance. The significance is therefore conditional: the framework is potentially useful, but the current evidence does not yet establish that the cosine-distance clustering is what drives the gains.

major comments (4)
  1. [Global-End Clustering, Eq. (3); Ablation Study, Table 6] The central assumption that cosine distance between decoder weight vectors reflects domain similarity is never directly tested. The paper itself concedes in the limited-sample section that 'the dynamic model exchange in FEDEXCHANGE relies on the assumption that model weights reflect the local domain distribution,' and at 10% data the advantage disappears. Yet no experiment reports whether the two clusters correspond to the true domain labels (e.g., weather or scene), no distance-matrix or embedding visualization is shown, and no ablation isolates the clustering step from the exchange schedule. Table 6 is particularly important: random exchange (30.30 mAP) is worse than FedAvg (30.57 mAP), while FEDEXCHANGE reaches 32.08. If the clustering is not domain-aligned, the method reduces to a random shuffle and should not beat FedAvg. Please provide a direct cluster-to-domain alignment analysis
  2. [Performance Comparison, Table 3; Abstract] The headline '1.6X better mean average precision in challenging domains, such as rainy conditions' is based on a single cell in Table 3: Rainy mAP of 75.08 for FEDEXCHANGE versus 47.02 for FedAvg. No multiple seeds, error bars, or significance tests are reported anywhere in the paper. The magnitude is also an outlier relative to other domains (Clear +1.55, Overcast +4.01, Rainy +28.06), and the cross-domain standard deviation jumps from 7.40 (FedAvg) to 19.35 (FEDEXCHANGE). This one result carries the abstract's strongest claim, so it must be shown to be stable. Please report mean and standard deviation over at least three independent runs and discuss whether the Rainy improvement persists.
  3. [Global-End Clustering; Hyper-parameter choices] The number of clusters is fixed to two in all experiments. This is a free parameter that is not ablated, and it is not obviously correct for the evaluated tasks: four person-detection domains, four BDD100K scene splits, and three SODA10M weather splits. The paper's justification—that in-cluster exchange handles models that would be better separated under a different clustering—does not address the opposite risk, namely that K=2 merges genuinely dissimilar domains and forces inappropriate cross-cluster swaps. Please ablate K (e.g., K=3,4 or a data-driven choice such as silhouette score) or provide a principled reason why two clusters is sufficient across all benchmarks.
  4. [Experiments; Algorithm 1] The evaluation protocol is ambiguous. In exchange rounds, each client receives a different decoder; only at rounds where r%T==0 is a global aggregate produced. It is not stated whether Tables 1–3 report the final aggregated decoder, each client's locally received (possibly exchanged) model, or some other combination. This matters for interpreting the average mAP and the standard deviation across domains, and for ensuring that all baselines are evaluated under the same protocol. Please specify exactly which model weights are evaluated to produce each reported number (e.g., 'the global model obtained in the final aggregation round for all methods') and confirm that the same protocol is used for FedAvg, FedProx, FedSTO, and FEDEXCHANGE.
minor comments (5)
  1. [Throughout] Several section cross-references are empty ('Section ', 'Section '), and the reference to Zhuang et al. is incomplete ('Zhuang et al. ???? COALA'). Please fill in all internal cross-references and the missing citation.
  2. [Method, Eq. (2)] Equation (2) contains a typo in the summation index: 'n = Pm n=1 ni' should likely be 'n = \sum_{i=1}^m n_i'.
  3. [Figure 4 caption] The caption for Figure 4c says 'SODA100M' instead of 'SODA10M'.
  4. [Algorithm 1] The notation in Algorithm 1 is inconsistent: line 3 uses 'f i_l' where the client index is likely intended, and the warm-up stage is not described in the main text. Please clarify the warm-up and the indexing.
  5. [Performance Comparison] The sentence 'The reason is that the model exchange enables the model to learn domain-invariant features that are hardly learned in its own domain' is a causal claim that is not directly evidenced. It would be more precise to describe this as the intended mechanism.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FEDEXCHANGE's exchange mechanism is an empirical algorithm with a testable domain-encoding assumption; no fitted parameter is relabeled as a prediction.

full rationale

FEDEXCHANGE is an empirical method, not a derivation. The server-side clustering and exchange (Eqs. 3-5) are algorithmic proposals whose value is assessed by benchmark mAP, not derived from fitted parameters. The key premise—that decoder weights encode domain distribution—is explicitly acknowledged as an assumption in the limited-sample section: 'the dynamic model exchange in FEDEXCHANGE relies on the assumption that model weights reflect the local domain distribution.' This is testable and is probed by the 50%/10% data experiments and by the exchange-strategy ablation (Table 6). If the assumption fails, the method degrades toward random exchange; that makes the assumption load-bearing for performance, but it does not make the reported improvements circular. No constants are fitted to test labels; the aggregation frequency T=2 is a hyperparameter chosen via ablation, not a predicted quantity. Self-citations (COALA, prior FL works) are used for background/infrastructure and do not carry the main claim. The large Rainy-domain gain is a robustness concern, not a circularity concern. Therefore no circular step is present.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No theoretical derivation; the approach is empirical. The main free choices are the aggregation frequency and the fixed two-cluster setting. The central assumption that decoder weights encode domain information is acknowledged by the authors but not independently verified. No invented entities.

free parameters (2)
  • aggregation frequency T = 2
    Chosen by hand as default; ablated over {2,5,10,50}; affects the alternation between exchange and aggregation rounds.
  • number of clusters = 2
    Fixed default; authors note 'difficulty in predefining the optimal number of clusters'.
assumptions (3)
  • domain assumption Model weights (decoder parameters) reflect the local domain distribution.
    Invoked in the limited-data section ('relies on the assumption that model weights reflect the local domain distribution') and underpins the whole clustering-exchange mechanism.
  • domain assumption Cosine distance between decoder weight vectors is a meaningful measure of domain similarity.
    Used to cluster decoders in the Global-End Clustering stage (Eq. 3); never validated against actual domain labels.
  • domain assumption Exchanging decoders helps learn domain-invariant features without catastrophic forgetting.
    The method's motivation; ablation shows exchange helps but the forgetting risk is acknowledged in the aggregation-frequency discussion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FEDEXCHANGE: Bridging the Domain Gap in Federated Object Detection for Free." pith.science (2026). https://pith.science/paper/TEU7JOLQ

@misc{pith2026250910503,
  author       = {Pith},
  title        = {Pith review of: FEDEXCHANGE: Bridging the Domain Gap in Federated Object Detection for Free},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TEU7JOLQ}},
  note         = {Machine review of arXiv:2509.10503}
}
read the original abstract

Federated Object Detection (FOD) enables clients to collaboratively train a global object detection model without accessing their local data from diverse domains. However, significant variations in environment, weather, and other domain specific factors hinder performance, making cross domain generalization a key challenge. Existing FOD methods often overlook the hardware constraints of edge devices and introduce local training regularizations that incur high computational costs, limiting real-world applicability. In this paper, we propose FEDEXCHANGE, a novel FOD framework that bridges domain gaps without introducing additional local computational overhead. FEDEXCHANGE employs a server side dynamic model exchange strategy that enables each client to gain insights from other clients' domain data without direct data sharing. Specifically, FEDEXCHANGE allows the server to alternate between model aggregation and model exchange. During aggregation rounds, the server aggregates all local models as usual. In exchange rounds, FEDEXCHANGE clusters and exchanges local models based on distance measures, allowing local models to learn from a variety of domains. As all operations are performed on the server side, clients can achieve improved cross domain utility without any additional computational overhead. Extensive evaluations demonstrate that FEDEXCHANGE enhances FOD performance, achieving 1.6X better mean average precision in challenging domains, such as rainy conditions, while requiring only 0.8X the computational resources compared to baseline methods.

Figures

Figures reproduced from arXiv: 2509.10503 by the authors.

Figure 1
Figure 1. Comparison of memory and communication costs [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Workflow of FEDEXCHANGE. The global server first defines the aggregation frequency T. During FOD training, clients upload their local model weights to the server at each communication round. Based on the current round, the server determines whether to perform model aggregation or dynamic model exchange. If dynamic exchange is selected, the server clusters all uploaded models into two groups using cosine distance and… view at source ↗
Figure 3
Figure 3. An illustration of dynamic model exchange work [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Evaluation of convergence of FEDEXCHANGE vs. baselines. FEDEXCHANGE performs better on multiple cross-domain person datasets and achieves better convergence than other methods on BDD100K and SODA100M datasets. T = {2, 5, 10, 50} and compare the resulting final utility.…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 41 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    C.; and Lin, D

    Chen, K.; Wang, J.; Pang, J.; Cao, Y.; Xiong, Y.; Li, X.; Sun, S.; Feng, W.; Liu, Z.; Xu, J.; Zhang, Z.; Cheng, D.; Zhu, C.; Cheng, T.; Zhao, Q.; Li, B.; Lu, X.; Zhu, R.; Wu, Y.; Dai, J.; Wang, J.; Shi, J.; Ouyang, W.; Loy, C. C.; and Lin, D. 2019. MMDetection : Open MMLab Detection Toolbox and Benchmark. arXiv preprint arXiv:1906.07155

    arXiv 2019

  4. [4]

    Chen, Z.; Duan, Y.; Wang, W.; He, J.; Lu, T.; Dai, J.; and Qiao, Y. 2023. Vision Transformer Adapter for Dense Predictions. In The Eleventh International Conference on Learning Representations

    work page 2023

  5. [5]

    Coates, A.; and Ng, A. Y. 2010. Multi-camera object detection for robotics. In 2010 IEEE International conference on robotics and automation, 412--419. IEEE

    work page 2010

  6. [6]

    Cordts, M.; Omran, M.; Ramos, S.; Scharw \"a chter, T.; Enzweiler, M.; Benenson, R.; Franke, U.; Roth, S.; and Schiele, B. 2015. The Cityscapes Dataset. In CVPR Workshop on The Future of Datasets in Vision

    work page 2015

  7. [7]

    Deng, J.; Li, W.; Chen, Y.; and Duan, L. 2021. Unbiased mean teacher for cross-domain object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4091--4101

    work page 2021

  8. [8]

    Deng, J.; Xu, D.; Li, W.; and Duan, L. 2023. Harmonious Teacher for Cross-Domain Object Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 23829--23838

    work page 2023

Show all 52 references
  1. [9]

    Han, J.; Liang, X.; Xu, H.; Chen, K.; Hong, L.; Mao, J.; Ye, C.; Zhang, W.; Li, Z.; Liang, X.; and Xu, C. 2021. SODA10M: A Large-Scale 2D Self/Semi-Supervised Object Detection Dataset for Autonomous Driving. arXiv:2106.11118

    2021 arXiv

  2. [10]

    He, M.; Wang, Y.; Wu, J.; Wang, Y.; Li, H.; Li, B.; Gan, W.; Wu, W.; and Qiao, Y. 2022 a . Cross domain object detection by target-perceived dual branch distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9570--9580

    2022

  3. [11]

    He, M.; Wang, Y.; Wu, J.; Wang, Y.; Li, H.; Li, B.; Gan, W.; Wu, W.; and Qiao, Y. 2022 b . Cross domain object detection by target-perceived dual branch distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9570--9580

    2022

  4. [12]

    He, Z.; and Zhang, L. 2020. Domain adaptive object detection via asymmetric tri-way faster-rcnn. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XXIV 16, 309--324. Springer

    2020

  5. [13]

    Hsu, C.-C.; Tsai, Y.-H.; Lin, Y.-Y.; and Yang, M.-H. 2020. Every pixel matters: Center-aware feature alignment for domain adaptive object detector. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part IX 16, 733--748. Springer

    2020

  6. [14]

    Hu, Y.; Yang, J.; Chen, L.; Li, K.; Sima, C.; Zhu, X.; Chai, S.; Du, S.; Lin, T.; Wang, W.; et al. 2023. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 17853--17862

    2023

  7. [15]

    Kim, T.; Jeong, M.; Kim, S.; Choi, S.; and Kim, C. 2019. Diversify and Match: A Domain Adaptive Representation Learning Paradigm for Object Detection. arXiv:1905.05396

    2019 arXiv

  8. [16]

    Kim, T.; Lin, E.; Lee, J.; Lau, C.; and Mugunthan, V. 2024. Navigating Data Heterogeneity in Federated Learning A Semi-Supervised Federated Object Detection

    2024

  9. [17]

    Lepetit, V.; and Berger, M.-O. 2000. A semi-automatic method for resolving occlusion in augmented reality. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), volume 2, 225--230. IEEE

    2000

  10. [18]

    H.; Su, L.; Xie, B.; Ding, Z.; Chen, C

    Li, S.; Liu, C. H.; Su, L.; Xie, B.; Ding, Z.; Chen, C. L. P.; and Wu, D. 2020 a . Discriminative Transfer Feature and Label Consistency for Cross-Domain Image Classification. IEEE Transactions on Neural Networks and Learning Systems

    2020

  11. [19]

    K.; Talwalkar, A.; and Smith, V

    Li, T.; Sahu, A. K.; Talwalkar, A.; and Smith, V. 2020 b . Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine

    2020

  12. [20]

    K.; Zaheer, M.; Sanjabi, M.; Talwalkar, A.; and Smith, V

    Li, T.; Sahu, A. K.; Zaheer, M.; Sanjabi, M.; Talwalkar, A.; and Smith, V. 2020 c . Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2: 429--450

    2020

  13. [21]

    Li, Y.-J.; Dai, X.; Ma, C.-Y.; Liu, Y.-C.; Chen, K.; Wu, B.; He, Z.; Kitani, K.; and Vajda, P. 2022. Cross-Domain Adaptive Teacher for Object Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 7581--7590

    2022

  14. [22]

    Lin, T.-Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Doll \'a r, P.; and Zitnick, C. L. 2014. Microsoft coco: Common objects in context. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13

    2014

  15. [23]

    Liu, D.; Long, C.; Zhang, H.; Yu, H.; Dong, X.; and Xiao, C. 2020. ARShadowGAN: Shadow Generative Adversarial Network for Augmented Reality in Single Light Scenes. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 8136--8145

    2020

  16. [24]

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

    2016

  17. [25]

    Maiettini, E.; Pasquale, G.; Rosasco, L.; and Natale, L. 2020. On-line object detection: a robotics challenge. Autonomous Robots, 44(5): 739--757

    2020

  18. [26]

    B.; Moore, E.; Ramage, D.; Hampson, S.; and y Arcas, B

    McMahan, H. B.; Moore, E.; Ramage, D.; Hampson, S.; and y Arcas, B. A. 2023. Communication-Efficient Learning of Deep Networks from Decentralized Data. arXiv:1602.05629

    2023 arXiv

  19. [27]

    C.; Pham, Q.-V.; Pathirana, P

    Nguyen, D. C.; Pham, Q.-V.; Pathirana, P. N.; Ding, M.; Seneviratne, A.; Lin, Z.; Dobre, O.; and Hwang, W.-J. 2022. Federated learning for smart healthcare: A survey. ACM Computing Surveys (Csur), 55(3): 1--37

    2022

  20. [28]

    T.; Mukherjee, S.; Aggarwal, J.; Lee, H.; Davis, L.; et al

    Oh, S.; Hoogs, A.; Perera, A.; Cuntoor, N.; Chen, C.-C.; Lee, J. T.; Mukherjee, S.; Aggarwal, J.; Lee, H.; Davis, L.; et al. 2011. A large-scale benchmark dataset for event recognition in surveillance video. In CVPR 2011, 3153--3160. IEEE

    2011

  21. [29]

    Oquab, M.; Darcet, T.; Moutakanni, T.; Vo, H.; Szafraniec, M.; Khalidov, V.; Fernandez, P.; Haziza, D.; Massa, F.; El-Nouby, A.; et al. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193

    2023 arXiv

  22. [30]

    Ren, S.; He, K.; Girshick, R.; and Sun, J. 2016. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6): 1137--1149

    2016

  23. [31]

    O.; Zhang, S.; and Schiele, B

    Rezaeianaran, F.; Shetty, R.; Aljundi, R.; Reino, D. O.; Zhang, S.; and Schiele, B. 2021. Seeking similarities over differences: Similarity-based domain alignment for adaptive object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 9204--9213

    2021

  24. [32]

    Saito, K.; Ushiku, Y.; Harada, T.; and Saenko, K. 2019. Strong-weak distribution alignment for adaptive object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 6956--6965

    2019

  25. [33]

    Sandler, M.; Howard, A.; Zhu, M.; Zhmoginov, A.; and Chen, L.-C. 2019. MobileNetV2: Inverted Residuals and Linear Bottlenecks. arXiv:1801.04381

    2019 arXiv

  26. [34]

    Shao, S.; Zhao, Z.; Li, B.; Xiao, T.; Yu, G.; Zhang, X.; and Sun, J. 2018. CrowdHuman: A Benchmark for Detecting Human in a Crowd. arXiv preprint arXiv:1805.00123

    2018 arXiv

  27. [35]

    Shokri, R.; and Shmatikov, V. 2015. Privacy-preserving deep learning. In Proceedings of the 22nd ACM SIGSAC conference on computer and communications security

    2015

  28. [36]

    Sun, B.; Huo, H.; Yang, Y.; and Bai, B. 2021. Partialfed: Cross-domain personalized federated learning via partial initialization. Advances in Neural Information Processing Systems, 34: 23309--23320

    2021

  29. [37]

    P.; Silver, C.; and Akilan, T

    Swaminathan, T. P.; Silver, C.; and Akilan, T. 2024. Benchmarking Deep Learning Models on NVIDIA Jetson Nano for Real-Time Systems: An Empirical Investigation. arXiv:2406.17749

    2024 arXiv

  30. [38]

    Wang, L.; Bian, J.; Zhang, L.; Chen, C.; and Xu, J. 2024. Taming Cross-Domain Representation Variance in Federated Prototype Learning with Heterogeneous Data Domains. arXiv:2403.09048

    2024 arXiv

  31. [39]

    Xie, R.; Yu, F.; Wang, J.; Wang, Y.; and Zhang, L. 2019. Multi-level domain adaptive learning for cross-domain detection. In Proceedings of the IEEE/CVF international conference on computer vision workshops, 0--0

    2019

  32. [40]

    S.; Su, C.; Walker, P.; Bian, J.; and Wang, F

    Xu, J.; Glicksberg, B. S.; Su, C.; Walker, P.; Bian, J.; and Wang, F. 2021. Federated learning for healthcare informatics. Journal of healthcare informatics research, 5: 1--19

    2021

  33. [41]

    Yang, Y.; Hui, B.; Yuan, H.; Gong, N.; and Cao, Y. 2023. \ PrivateFL \ : Accurate, differentially private federated learning via personalized data transformation. In 32nd USENIX Security Symposium (USENIX Security 23), 1595--1612

    2023

  34. [42]

    Yang, Y.; Lee, K.; Dariush, B.; Cao, Y.; and Lo, S.-Y. 2024. Follow the Rules: Reasoning for Video Anomaly Detection with Large Language Models. In Proceedings of the European Conference on Computer Vision (ECCV)

    2024

  35. [43]

    Yu, F.; Chen, H.; Wang, X.; Xian, W.; Chen, Y.; Liu, F.; Madhavan, V.; and Darrell, T. 2020. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    2020

  36. [44]

    Z.; and Cao, Y

    Yuan, H.; Hui, B.; Yang, Y.; Burlina, P.; Gong, N. Z.; and Cao, Y. 2022. Addressing heterogeneity in federated learning via distributional transformation. In European Conference on Computer Vision, 179--195. Springer

    2022

  37. [45]

    Yuan, H.; Paul, W.; Aucott, J.; Burlina, P.; and Cao, Y. 2024. PFedEdit: Personalized Federated Learning via Automated Model Editing

    2024

  38. [46]

    Z.; and Guo, G

    Zhang, S.; Xie, Y.; Wan, J.; Xia, H.; Li, S. Z.; and Guo, G. 2019. WiderPerson: A Diverse Dataset for Dense Pedestrian Detection in the Wild. IEEE Transactions on Multimedia (TMM)

    2019

  39. [47]

    Zhang, Z.; Cao, X.; Jia, J.; and Gong, N. Z. 2022. Fldetector: Defending federated learning against model poisoning attacks via detecting malicious clients. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2545--2555

    2022

  40. [48]

    Zheng, W.; Yan, L.; Gou, C.; and Wang, F.-Y. 2021. Federated meta-learning for fraudulent credit card detection. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, 4654--4660

    2021

  41. [49]

    Zheng, Y.; Huang, D.; Liu, S.; and Wang, Y. 2020. Cross-domain Object Detection through Coarse-to-Fine Feature Adaptation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 13763--13772

    2020

  42. [50]

    Zhu, X.; Pang, J.; Yang, C.; Shi, J.; and Lin, D. 2019. Adapting Object Detectors via Selective Cross-Domain Alignment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 687--696

    2019

  43. [51]

    Zhuang, W.; Chen, C.; and Lyu, L. 2023. When foundation model meets federated learning: Motivations, challenges, and future directions. arXiv preprint arXiv:2306.15546

    2023 arXiv

  44. [52]

    ???? COALA: A Practical and Vision-Centric Federated Learning Platform

    Zhuang, W.; Xu, J.; Chen, C.; Li, J.; and Lyu, L. ???? COALA: A Practical and Vision-Centric Federated Learning Platform. In Forty-first International Conference on Machine Learning

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.