Pith. sign in

REVIEW 5 major objections 6 minor 50 references

FeDepth: Federated Learning for Depth Estimation under Robot Heterogeneity

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read FeDepth claims that a one-time, descriptor-based soft clustering of robots before federated training yields consistently more robust depth models than standard FL or hard-clustered FL, even when client domains overlap.

desk verdict FeDepth's soft clustering is a real win on the HPE benchmark, but the abstract overclaims: BMR shows parity with PACFL, not consistent improvement. read the letter →

arxiv 2608.01129 v1 pith:2E5VKA3Z submitted 2026-08-02 cs.RO cs.CV

Ganghyeon Lee , Inha Lee , Junhee Lee , Jeongeon Lee , Sung Whan Yoon , Kyungdon Joo This is my paper
classification cs.ROcs.CV
keywords federatedlearningmonoculardepthestimationrobotheterogeneitysoftclusteringclusterednon-IIDdatadomainshiftperception
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

Federated learning lets robots train depth models without uploading raw images, but robot fleets are heterogeneous: platforms, environments, and depth ranges differ, and the resulting client distributions often overlap rather than falling into clean groups. FeDepth is a clustered federated learning method built on the claim that this overlap is exactly why standard FL and hard-clustering CFL fail. It assigns each client to multiple clusters before training, using feature-distribution descriptors from a frozen encoder, then aggregates models per cluster and gives each client the average of its clusters' models. Across the paper's two new non-IID scenarios, one mixing UGV/UAV/legged platforms and environments from M3ED, the other mixing indoor NYUv2 and outdoor KITTI, FeDepth reports consistent robustness gains over FedAvg, FedProx, FedDyn, SCAFFOLD, FedClust, and PACFL on three depth-estimation architectures. If the claim holds, heterogeneous robot fleets can collaborate privately while each platform still benefits from a model matched to its own domain.

What carries the argument

The client descriptor and the soft cluster map built from it. A client descriptor is $\mathbf{v}_c = (\mu_c, \sigma_c^2)$: the mean and variance of GeM-pooled features over sampled frames from the frozen encoder, approximating each client's feature distribution as a Gaussian with diagonal covariance. The map $\mathbf{M}^* \in \{0,1\}^{K\times N}$ is produced by thresholding the Jeffreys-divergence similarity matrix and merging subset clusters. The map is the load-bearing object: it defines which clients are aggregated into which cluster models, which cluster models each client averages at the start of a round, and which cluster models a test sequence selects at inference.

What would settle it

On the HPE split, retrain FeDepth and at round 50 recompute client descriptors from the current trained encoder instead of the frozen one, then compare the new cluster map to the one-time map, for example by adjusted Rand index. If the maps disagree substantially, or if using the refreshed map lowers Abs Rel by more than about 0.02 on NeWCRFs, the one-shot clustering assumption is the part that fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that, for federated monocular depth estimation under robot heterogeneity, the right unit of model sharing is not the global model and not a hard partition of clients, but a soft cluster map. In FeDepth, each client computes a descriptor, the mean and variance of generalized-mean-pooled features taken from a frozen pretrained encoder, and the server compares clients with the symmetric Jeffreys divergence between Gaussian approximations of these descriptors. Thresholding this similarity map and merging subset clusters yields overlapping clusters, so a client that shares a platform with one group and an environment with another can contribute to both. Training then

Load-bearing premise

The load-bearing premise is that the cluster map computed once from frozen-encoder descriptors before training stays accurate enough through all later rounds, even though the depth model weights change substantially.

Editorial extensions

If this is right

  • The two scenarios, HPE from M3ED and BMR from NYUv2 plus KITTI, become reusable testbeds for federated robot perception under non-IID data.
  • Soft clustering retains most of the benefit of hard clustering even when domains are cleanly separated, as the BMR results show, so one method covers both overlapping and separable heterogeneity.
  • Unseen-client generalization works without test-time personalization: a new robot sequence selects clusters by descriptor similarity and averages the chosen models.
  • The mechanism is model-agnostic in the tested sense, giving consistent gains for bin-based, CRF-based, and DCT-based depth heads.
  • Robots can keep raw images local; only descriptors and model weights move to the server.

Reading between the lines

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

  • Editorial inference: because the cluster map is frozen after one descriptor extraction, its stability against the evolving model is untested; a natural stress test is recomputing descriptors from mid-training checkpoints and checking whether the map drifts.
  • Editorial inference: the descriptor mechanism is task-agnostic, so FeDepth could plausibly transfer to other dense prediction tasks such as semantic segmentation or optical flow in federated robot settings, not just depth.
  • Editorial inference: the overlap ratio, set here to 20 percent, and the threshold tau jointly control the method; in larger fleets or with stronger domain shifts the optimal overlap may differ, and the paper does not give a rule for setting it.
  • Editorial inference: a practical deployment question left open is how to handle clients whose distribution changes over time, such as seasonal or lighting drift, since the once-computed clusters would not adapt unless the pipeline is re-run.
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

5 major / 6 minor

Summary. The paper proposes FeDepth, a clustered federated learning framework for monocular depth estimation under robot heterogeneity. The authors introduce two new non-IID benchmarks: HPE (platform-environment combinations from M3ED) and BMR (bi-modal indoor/outdoor depth ranges from NYUv2 and KITTI). FeDepth performs a one-time soft clustering of clients using descriptors extracted with a frozen encoder (GeM pooling of features; mean and variance approximated as Gaussian; Jeffreys divergence similarity), assigns each client to possibly multiple clusters, trains cluster-specific models via FedAvg-style aggregation, and at inference selects a weighted combination of cluster models based on descriptor similarity to the test sequence. Experiments cover three depth architectures (AdaBins, NeWCRFs, DCDepth); ablations compare hard vs. soft clustering, overlap ratio sensitivity, and domain-label-based grouping. The central claim is that FeDepth consistently improves robustness over standard and clustered FL baselines across architectures.

Significance. If the claims hold, the paper makes a useful contribution by (i) identifying an important and under-studied setting—overlapping client distributions in federated robot perception—and (ii) showing that a simple descriptor-based soft clustering can beat hard-clustering CFL baselines in that setting. The HPE benchmark and the use of three depth architectures are strengths. The paper also ships detailed supplementary material, including full pseudocode, cluster visualizations, per-sequence splits, and standard deviations, which is commendable. However, the significance is tempered by two issues: the abstract's 'consistently improves' claim is not supported by the BMR results, and the clustering method is a pre-training heuristic whose stability is not validated. The HPE results are strong and support the framework's value in the overlapping-domain scenario; the BMR results show parity or slight inferiority to hard clustering, which is a reasonable outcome but must be stated accurately.

major comments (5)
  1. [Abstract / Sec. 4.2 / Table 2] The abstract and Sec. 4.2 claim that FeDepth 'consistently improves robustness over standard FL and clustered FL baselines across multiple depth estimation architectures.' This is contradicted by the BMR results in Table 2 and Table A2. For AdaBins, PACFL† achieves Abs Rel 0.124 vs. FeDepth 0.126, Sq Rel 0.263 vs. 0.283, and RMSE 2.022 vs. 2.150. For NeWCRFs, FeDepth is better only on RMSE (1.499 vs. 1.501) and ties on Sq Rel, while losing on Abs Rel and RMSElog. For DCDepth, FeDepth wins on Sq Rel (0.127 vs. 0.130) and ties on Abs Rel and the δ thresholds. The paper even concedes 'a minor performance gap remains compared to PACFL†' for AdaBins. The blanket 'consistently improves' claim is internally inconsistent with these numbers. The claim must be qualified to the HPE scenario, or BMR must be reported as parity/approximate parity with hard clustering.
  2. [Sec. 3.3 / Algorithm 2] The soft clustering is performed once before training using descriptors from a frozen encoder, and the cluster map M* is never updated during federated training. The model weights change substantially over rounds, so the client feature distributions with respect to the frozen encoder may shift, making the clusters stale. The paper provides no measurement of cluster stability across rounds, no ablation in which clustering is recomputed at intermediate rounds, and no analysis of whether the overlap structure remains valid. This is load-bearing because the claimed advantage of FeDepth over hard clustering rests on the cluster map being a good representation of client relationships throughout training. If the clusters become stale, the method could underperform a method that updates clustering (e.g., IFCA-style), or the reported gains could be an artifact of favorable pre-training geometry.
  3. [Sec. 4.1 / Table 3b] The overlap ratio (and the threshold τ that induces it) is selected on the test set. The text states 'in HPE, the clustering threshold τ is selected to yield an overlap ratio of 20%,' and Table 3b reports sensitivity over 0–100% overlap, with 20% best. Because the overlap ratio is a hyperparameter tuned on the same test split used for evaluation, the reported HPE improvements are partly the result of oracle selection. The paper should clarify whether the 20% choice generalizes (e.g., selected by validation-like protocol or consistent across architectures) and should report variance across random seeds or multiple test splits. Without this, the magnitude of the claimed gains is uncertain.
  4. [Sec. 4.2 / BMR evaluation] The BMR comparison is incomplete in an important way: the paper compares only FeDepth and PACFL† in the main table, omitting FedClust from Table 2 (FedClust results appear only in Table A2). For AdaBins, FedClust outperforms FeDepth on Abs Rel (0.133 vs. 0.126? Actually FedClust 0.133 vs. FeDepth 0.126, so FeDepth wins; for NeWCRFs, FedClust 0.105 vs. FeDepth 0.089, so FeDepth wins; for DCDepth, FedClust 0.087 vs. FeDepth 0.082, so FeDepth wins). This omission is not harmful to FeDepth's standing, but the main text should state FedClust results to support the claim that FeDepth is 'comparable to hard clustering.' More importantly, the claim 'FeDepth is not limited to overlapping domains' is supported, but the claim 'consistently improves' is not. Rewrite the BMR paragraph to state that FeDepth achieves parity with PACFL† on two architectures and is slightly worse on AdaBins, while still
  5. [Sec. 4.1 / Baseline fairness] PACFL† is a modified version of PACFL that uses FeDepth's descriptor extractor, but the original PACFL is not evaluated. The paper should report the original PACFL (with raw images or the authors' features) to show that the descriptor extractor is not the source of improvement. In Table 1, PACFL† is worse than FedAvg for AdaBins (0.602 vs. 0.434) but better for NeWCRFs and DCDepth. This suggests the descriptor extractor matters substantially, and the claim that FeDepth's advantage comes from soft clustering (rather than from the descriptor quality) is not fully established without the original PACFL baseline. Please add original PACFL or at least discuss why direct comparison is infeasible.
minor comments (6)
  1. [Abstract / Conclusion] The abstract and conclusion should be rephrased to 'FeDepth consistently improves over standard FL and hard-clustering baselines in the HPE scenario, and performs comparably in the BMR scenario where domains are clearly separated.'
  2. [Eq. (2) / Notation] In Eq. (2), the GeM pooling uses the same symbol p for the pooling exponent and for the data-weight p_k^c in Eq. (4); this is notationally confusing. Please use a different symbol (e.g., α) for the GeM exponent.
  3. [Sec. 3.3 / Inference weighting] In Eq. (5) and Algorithm 4, the weights p_k are defined as inverse similarity but the normalization (p_k = S_k^{-1} / sum_j S_j^{-1}) is clear only in the algorithm. Please state the normalization in the main text.
  4. [Fig. 4 / Caption] The caption is long and includes the text 'Soft Cluster Merging' and 'Cluster Matching & Model Averaging' but does not explain the colors or the meaning of the vertical/horizontal axes of the cluster map. Please make the caption self-contained.
  5. [Sec. 4.1 / Computation] The statement '2–4 days per experiment using four RTX 4090 GPUs' is given without information on the number of communication rounds, local epochs, or batch size. This makes the cost claim unverifiable; please add the hyperparameter table (or reference the supplement) in the main text.
  6. [References] The related work correctly cites the cited works, but the letter '†' in PACFL† is introduced in Table 1 and used in the text; please define it in the main text (it is defined in the experiments section but the definition appears after the table caption). Also, the reference to 'prior work [5]' in Sec. 4.2 (ZoeDepth) should specify what prior work showed, as it is not immediately clear.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: FeDepth's cluster map derives from frozen-encoder descriptors and is benchmarked against external FL baselines; only mild test-set tuning of the overlap ratio τ, and the abstract's 'consistently improves' claim is over-stated relative to the paper's own Table 2 (a consistency, not circularity, issue).

  1. fitted input called prediction [Sec. 4.1 (hyperparameter selection) and Tab. 3b (Ablation study), used for Tab. 1 (HPE results)]
    "In HPE, the clustering threshold τ is selected to yield an overlap ratio of 20%. ... As a result, we adopt an overlap ratio of 20%, which yields the best performance in our experiments."

    The hyperparameter controlling the cluster map (τ/overlap ratio) is selected by sweeping the headline metric (Abs Rel) on the HPE test set in Tab. 3b, and the same test set is then used to report FeDepth's claimed improvements in Tab. 1. The reported advantage is therefore partly a consequence of picking the parameter value that maximizes it on the evaluation set, rather than a parameter-free derivation. This is a mild form of fitted-input selection: the soft-clustering advantage is NOT forced by construction, since overlap 10% (0.253) and 30% (0.257) also beat the hard-clustering result (0.264) in Tab. 3b, so the central claim has independent content. Additionally, in BMR, τ is deliberately set so that cluster assignments approximate hard clustering, pre-arranging the observed parity with

full rationale

The derivation chain is otherwise self-contained and externally falsifiable. Descriptors V are extracted once from a frozen pre-trained encoder (Alg. 1) before any federated training; the similarity map A is computed with Jeffreys divergence (Eq. 3); the binary cluster map M is thresholded, subset-merged, and singleton-reassigned (Alg. 2); cluster models are aggregated by FedAvg weighting and client models by cluster-average (Eq. 4); and inference selects cluster models by descriptor similarity with the same frozen encoder and same τ (Eq. 5). No step is defined in terms of the depth-estimation loss or the reported metrics, and the cluster map is never updated based on training outcomes. The comparisons (Tab. 1, Tab. 2, Tab. A1, Tab. A2) are against external baselines (FedAvg, FedProx, FedDyn, SCAFFOLD, FedClust, PACFL†) on unseen test clients, so the core claim is independently testable. The only self-citation by the present authors is reference [27] (Slabins), which is a peripheral related-work mention and not load-bearing. There is no imported uniqueness theorem and no ansatz smuggled in via citation. The genuine weaknesses are correctness risks, not circularity: (1) the overlap ratio τ is tuned on the HPE test set whose results it later supports (flagged above, scored mildly); (2) the abstract's claim of 'consistently improves robustness over ... clustered FL baselines across multiple depth estimation architectures' is contradicted by Tab. 2, where PACFL† is better than Ours on AdaBins (Abs Rel 0.124 vs 0.126, Sq Rel 0.263 vs 0.283, RMSE 2.022 vs 2.150) and on NeWCRFs Abs Rel (0.087 vs 0.089), a gap the paper partially concedes ('a minor performance gap remains compared to PACFL†'); (3) the clustering is performed once on frozen-encoder features and never updated, an assumption whose stability is unmeasured. These are internal-consistency and validation issues to be weighed in the correctness pass, not reductions of the claimed result to its inputs.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The client descriptor is a representation, not an entity. The main things the paper pulls in are the clustering structure and the test-set-tuned threshold.

free parameters (4)
  • clustering threshold tau = HPE: set to yield 20% overlap; BMR: set to approximate hard clustering
    The threshold determines cluster structure and is selected based on test results, so it is a fitted parameter.
  • overlap ratio = 20%
    Chosen for best performance in Tab. 3b, which is test-set tuning.
  • GeM pooling exponent p = unspecified
    Appears in Eq. (2) but no value is given; it affects descriptor quality.
  • number of sampled keyframes n = unspecified
    Used in Alg. 1 but no value is given; affects descriptor stability.
assumptions (3)
  • domain assumption Frozen pre-trained encoder provides discriminative features for client domain similarity.
    Section 3.3 and Fig. A1; the method assumes the frozen encoder of each depth baseline yields features that capture platform/environment differences.
  • domain assumption Diagonal Gaussian approximation with Jeffreys divergence adequately measures client distribution similarity.
    Section 3.3, Eq. (3); this is a modeling choice not justified beyond computational convenience.
  • domain assumption Client cluster assignments remain valid for the entire training process.
    Clustering is performed once before training and never updated; if model training shifts distributions, clusters may become stale.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FeDepth: Federated Learning for Depth Estimation under Robot Heterogeneity." pith.science (2026). https://pith.science/paper/2E5VKA3Z

@misc{pith2026260801129,
  author       = {Pith},
  title        = {Pith review of: FeDepth: Federated Learning for Depth Estimation under Robot Heterogeneity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2E5VKA3Z}},
  note         = {Machine review of arXiv:2608.01129}
}
read the original abstract

Although recent robot perception research emphasizes training on data from diverse environments to improve generalization, most existing methods still rely on centralized learning, which is inefficient and difficult to scale across heterogeneous robot platforms. Federated learning (FL) offers an alternative by enabling distributed training without raw data transfer, but it suffers from severe performance degradation under domain shifts caused by heterogeneity across clients. In real robotic deployments, data distributions often overlap across platforms, environments, and sensing conditions, making it difficult to partition clients into clearly separated domains. However, this characteristic breaks the assumption of clearly separable client domains commonly used in clustered FL. To address this gap in robot perception, particularly in depth estimation, we introduce two realistic and unexplored non-IID scenarios that reflect heterogeneity in terms of platform, environment, and depth distribution. We then propose FeDepth, a descriptor-based clustered FL framework that models client relationships through soft clustering. Unlike hard clustering methods that assume clearly separated clusters, FeDepth allows clients to participate in multiple clusters, capturing continuous and ambiguous domain transitions commonly observed in robotic environments. Extensive experiments demonstrate that FeDepth consistently improves robustness over standard FL and clustered FL baselines across multiple depth estimation architectures, providing a practical and effective solution for federated robot perception. Our project page is available at https://vision3d-lab.github.io/fedepth/.

Figures

Figures reproduced from arXiv: 2608.01129 by the authors.

Figure 1
Figure 1. Illustration of FeDepth. The proposed FeDepth, a soft clustering-based FL framework that models complex inter-client relationships for robust training and effective cluster model updates. For example, Clients 3 and 10, which share the same platform (e.g., UGV), are assigned to two overlapping clusters. from diverse environments [23, 44] has been shown to enhance generalization beyond controlled laboratory conditions… view at source ↗
Figure 2
Figure 2. Overview of the two scenarios for robot heterogeneity. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Data characteristics of two scenarios. Depth distribution and UMAP visualization [32]. instance, Client A (Legged robot) may capture data in a ‘Forest’ environment, while Client B (UAV) operates in the same ‘Forest’. While both clients share the ‘Forest’ attribute, their data distributions diverge due to platform-specific char￾acteristics like viewpoint and motion dynamics. This combinatorial nature and the presence… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Overall framework of FeDepth. (a) The Soft Clustering stage is first per￾formed before the initial round. In this stage, each client utilizes a Descriptor Extractor to generate a client descriptor summarizing its local data distribution. Then, the server aggregates the…
Figure 5
Figure 5. Figure 5: Qualitative results for DCDepth on the HPE scenario. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Cluster map and convergence graph. Results for NeWCRFs [46] in the HPE and BMR scenarios. Colors indicate cluster assignments, with each column repre￾senting a client. The convergence graphs report RMSElog over communication rounds. are continuous and overlapping, maki…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 27 canonical work pages

  1. [1]

    Acar,D.A.E.,Zhao,Y.,Navarro,R.M.,Mattina,M.,Whatmough,P.N.,Saligrama, V.: Federated learning based on dynamic regularization. In: Int. Conf. Learn. Rep- resent. (2021)

    work page 2021

  2. [2]

    In: IEEE Conf

    Arandjelovic, R., Gronát, P., Torii, A., Pajdla, T., Sivic, J.: NetVLAD: CNN Ar- chitecture for Weakly Supervised Place Recognition. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5297–5307 (2016).https://doi.org/10.1109/CVPR.2016.572

    doi:10.1109/cvpr.2016.572 2016

  3. [3]

    In: IEEE Conf

    Bhat, S.F., Alhashim, I., Wonka, P.: AdaBins: Depth Estimation Using Adaptive Bins. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 4009–4018 (2021).https: //doi.org/10.1109/CVPR46437.2021.00400

    arXiv 2021

  4. [4]

    Bhat, S.F., Alhashim, I., Wonka, P.: LocalBins: Improving Depth Estimation by Learning Local Distributions. In: Eur. Conf. Comput. Vis. pp. 480–496 (2022). https://doi.org/10.1007/978-3-031-19769-7_28

    doi:10.1007/978-3-031-19769-7_28 2022

  5. [5]

    arXiv preprint arXiv:2302.12288 (2023)

    Bhat, S.F., Birkl, R., Wofk, D., Wonka, P., Müller, M.: ZoeDepth: Zero-shot Trans- fer by Combining Relative and Metric Depth. arXiv preprint arXiv:2302.12288 (2023)

    arXiv 2023

  6. [6]

    O’Reilly (2008)

    Bradski, G.R., Kaehler, A.: Learning OpenCV - computer vision with the OpenCV library: software that sees. O’Reilly (2008)

    work page 2008

  7. [7]

    IEEE Trans

    Cao, Y., Wu, Z., Shen, C.: Estimating depth from monocular images as classifica- tion using deep fully convolutional residual networks. IEEE Trans. Circuit Syst. Video Technol.28(11), 3174–3182 (2018).https://doi.org/10.1109/TCSVT. 2017.2740321

    arXiv 2018

  8. [8]

    In: IEEE Conf

    Chaney, K., Cladera, F., Wang, Z., Bisulco, A., Hsieh, M.A., Korpela, C., Ku- mar, V., Taylor, C.J., Daniilidis, K.: M3ED: Multi-Robot, Multi-Sensor, Multi- Environment Event Dataset. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 4016–4023 (2023).https://doi.org/10.1109/CVPRW59228.2023.00419

    arXiv 2023

Show all 50 references
  1. [9]

    Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Eur. Conf. Comput. Vis. pp. 833–851. Springer-Verlag (2018).https://doi.org/10.1007/ 978-3-030-01234-2_49

    2018

  2. [10]

    Day, W.H.E., Edelsbrunner, H.: Efficient Algorithms for Agglomerative Hierarchi- cal Clustering Methods. J. Classification1, 7–24 (1984).https://doi.org/10. 1007/BF01890115

    1984

  3. [11]

    arXiv preprint arXiv:2003.13461 (2020)

    Deng, Y., Kamani, M.M., Mahdavi, M.: Adaptive personalized federated learning. arXiv preprint arXiv:2003.13461 (2020)

    2003 arXiv

  4. [12]

    Dinh,C.T.,Tran,N.H.,Nguyen,T.D.:Personalizedfederatedlearningwithmoreau envelopes. In: Adv. Neural Inform. Process. Syst. vol. 33, pp. 21394–21405 (2020) 16 G. Leeet al

    2020

  5. [13]

    Eigen, D., Puhrsch, C., Fergus, R.: Depth map prediction from a single image using a multi-scale deep network. In: Adv. Neural Inform. Process. Syst. pp. 2366–2374 (2014)

    2014

  6. [14]

    In: IEEE Conf

    Gehrig, M., Scaramuzza, D.: Recurrent vision transformers for object detection with event cameras. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 13884–13893 (2023).https://doi.org/10.1109/CVPR52729.2023.01334

    2023

  7. [15]

    Geiger, A., Lenz, P., Stiller, C., Urtasun, R.: Vision meets robotics: The KITTI dataset. Int. J. Robot. Res.32(11), 1231–1237 (2013).https://doi.org/10.1177/ 0278364913491297

    2013

  8. [16]

    In: IEEE Conf

    Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? The KITTI vision benchmark suite. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3354–3361 (2012).https://doi.org/10.1109/CVPR.2012.6248074

    2012

  9. [17]

    Ghosh, A., Chung, J., Yin, D., Ramchandran, K.: An efficient framework for clus- tered federated learning. In: Adv. Neural Inform. Process. Syst. (2020)

    2020

  10. [18]

    In: IEEE Conf

    Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., Gaidon, A.: 3D Packing for Self- Supervised Monocular Depth Estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2482–2491 (2020).https://doi.org/10.1109/CVPR42600.2020.00256

    2020

  11. [19]

    Guo, Y., Tang, X., Lin, T.: Enhancing clustered federated learning: Integration of strategies and improved methodologies. In: Int. Conf. Learn. Represent. (2025)

    2025

  12. [20]

    In: Proceedings of the 53rd International Conference on Parallel Process- ing (ICPP ’24)

    Islam, M.S., Javaherian, S., Xu, F., Yuan, X., Chen, L., Tzeng, N.: FedClust: Tack- ling Data Heterogeneity in Federated Learning through Weight-Driven Client Clus- tering. In: Proceedings of the 53rd International Conference on Parallel Process- ing (ICPP ’24). pp. 474–483. A...

    2024 doi

  13. [21]

    Oxford (1998)

    Jeffreys, H.: The theory of probability. Oxford (1998)

    1998

  14. [22]

    Karimireddy, S.P., Kale, S., Mohri, M., Reddi, S.J., Stich, S.U., Suresh, A.T.: SCAFFOLD: Stochastic Controlled Averaging for Federated Learning. In: Int. Conf. Mach. Learn. (2020)

    2020

  15. [23]

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W., Dollár, P., Girshick, R.B.: Segment anything. In: Int. Conf. Comput. Vis. pp. 3992–4003 (2023).https://doi.org/10.1109/ ICCV51070.2023.00371

    2023

  16. [24]

    Kou, W., Zhu, G., Cheng, B., Wang, S., Tang, M., Wu, Y.: FedEMA: Federated Exponential Moving Averaging with Negative Entropy Regularizer in Autonomous Driving.In:IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems (IROS). pp. 10190–10197. IEEE (2025).https://doi.org/...

    2025

  17. [25]

    Krizhevsky, A.: Learning multiple layers of features from tiny images. Tech. rep., University of Toronto (2009)

    2009

  18. [26]

    Laina, I., Rupprecht, C., Belagiannis, V., Tombari, F., Navab, N.: Deeper depth prediction with fully convolutional residual networks. In: Int. Conf. 3D Vis. pp. 239–248 (2016).https://doi.org/10.1109/3DV.2016.32

    2016 doi

  19. [27]

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

    Lee, J., Cho, G., Park, J., Kim, K., Lee, S., Kim, J.H., Jeong, S.G., Joo, K.: Slabins: Fisheye depth estimation using slanted bins on road environments. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8765–8774 (2023)

    2023

  20. [28]

    In: Proceedings of Machine Learning and Systems (MLSys) (2020)

    Li, T., Sahu, A.K., Zaheer, M., Sanjabi, M., Talwalkar, A., Smith, V.: Federated optimization in heterogeneous networks. In: Proceedings of Machine Learning and Systems (MLSys) (2020)

    2020

  21. [29]

    Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of FedAvg on non-IID data. In: Int. Conf. Learn. Represent. (2020) FeDepth 17

    2020

  22. [30]

    In: AAAI

    Liu, Y., Huang, A., Luo, Y., Huang, H., Liu, Y., Chen, Y., Feng, L., Chen, T., Yu, H., Yang, Q.: FedVision: An Online Visual Object Detection Platform Powered by Federated Learning. In: AAAI. pp. 13172–13179 (2020).https://doi.org/10. 1609/AAAI.V34I08.7021

    2020

  23. [31]

    In: IJCAI

    Lucas, B.D., Kanade, T.: An iterative image registration technique with an appli- cation to stereo vision. In: IJCAI. pp. 674–679 (1981)

    1981

  24. [32]

    arXiv preprint arXiv:1802.03426 (2018)

    McInnes, L., Healy, J., Melville, J.: UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. arXiv preprint arXiv:1802.03426 (2018)

    2018 arXiv

  25. [33]

    McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Int. Conf. Artif. Intell. Stat. pp. 1273–1282 (2017)

    2017

  26. [35]

    In: IEEE Conf

    Poggi, M., Tosi, F.: Federated online adaptation for deep stereo. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 20165–20175 (2024).https://doi.org/10.1109/ CVPR52733.2024.01906

    2024

  27. [36]

    IEEE Trans

    Radenović, F., Tolias, G., Chum, O.: Fine-tuning CNN image retrieval with no human annotation. IEEE Trans. Pattern Anal. Mach. Intell.41(7), 1655–1668 (2019)

    2019

  28. [37]

    Ranftl, R., Bochkovskiy, A., Koltun, V.: Vision transformers for dense prediction. In: Int. Conf. Comput. Vis. pp. 12159–12168 (2021).https://doi.org/10.1109/ ICCV48922.2021.01196

    2021

  29. [38]

    Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. In: Adv. Neural Inform. Process. Syst. pp. 91–99 (2015)

    2015

  30. [39]

    Soares, E.F., Campos, C.A.V.: Self-supervised learning of monocular depth estimators in autonomous vehicles with federated learning

    de S. Soares, E.F., Campos, C.A.V.: Self-supervised learning of monocular depth estimators in autonomous vehicles with federated learning. Eng. Appl. Artif. Intell. 151, 110572 (2025).https://doi.org/10.1016/j.engappai.2025.110572

    2025

  31. [40]

    IEEE Transactions on Neural Networks and Learning Systems32(8), 3710–3722 (2021).https://doi

    Sattler, F., Müller, K.R., Samek, W.: Clustered federated learning: Model-agnostic distributed multitask optimization under privacy constraints. IEEE Transactions on Neural Networks and Learning Systems32(8), 3710–3722 (2021).https://doi. org/10.1109/TNNLS.2020.3015958

    2021

  32. [41]

    Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support inference from RGBD images. In: Eur. Conf. Comput. Vis. pp. 746–760. Springer Berlin Heidelberg (2012).https://doi.org/10.1007/978-3-642-33715-4_54

    2012 doi

  33. [42]

    In: AAAI (2023).https://doi

    Vahidian, S., Morafah, M., Wang, W., Kungurtsev, V., Chen, C., Shah, M., Lin, B.: Efficient distribution similarity identification in clustered federated learning via principal angles between client data subspaces. In: AAAI (2023).https://doi. org/10.1609/aaai.v37i8.26197

    2023 doi

  34. [43]

    Wang, K., Yan, Z., Fan, J., Zhu, W., Li, X., Li, J., Yang, J.: DCDepth: Progressive Monocular Depth Estimation in Discrete Cosine Domain. In: Adv. Neural Inform. Process. Syst. (2024)

    2024

  35. [44]

    In: IEEE Conf

    Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Un- leashing the power of large-scale unlabeled data. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 10371–10381 (2024).https://doi.org/10.1109/CVPR52733. 2024.00987

    2024

  36. [45]

    Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., Zhao, H.: Depth anything V2. In: Adv. Neural Inform. Process. Syst. (2024) 18 G. Leeet al

    2024

  37. [46]

    In: IEEE Conf

    Yuan, W., Gu, X., Dai, Z., Zhu, S., Tan, P.: Neural Window Fully-connected CRFs for Monocular Depth Estimation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 3906–3915 (2022).https://doi.org/10.1109/CVPR52688.2022.00389

    2022

  38. [47]

    IEEE Trans

    Zhang, J., Liu, H., Yang, K., Hu, X., Liu, R., Stiefelhagen, R.: CMX: Cross-Modal Fusion for RGB-X Semantic Segmentation With Transformers. IEEE Trans. Intell. Transp.Syst.24(12),14679–14694(2023).https://doi.org/10.1109/TITS.2023. 3300537

    2023 doi

  39. [48]

    In: ICASSP

    Zhang,Y.,Chen,H.,Lin,Z.,Chen,Z.,Zhao,J.:LCFed:AnEfficientClusteredFed- erated Learning Framework for Heterogeneous Data. In: ICASSP. pp. 1–5 (2025). https://doi.org/10.1109/ICASSP49660.2025.10889428

    2025

  40. [49]

    arXiv preprint arXiv:1806.00582 (2018)

    Zhao, Y., Li, M., Lai, L., Suda, N., Civin, D., Chandra, V.: Federated learning with non-IID data. arXiv preprint arXiv:1806.00582 (2018)

    2018 arXiv

  41. [50]

    In: IEEE Conf

    Zheng, H., Hu, Z., Yang, L., Zheng, M., Xu, A., Wang, B.: Fedcalm: Conflict-aware layer-wise mitigation for selective aggregation in deeper personalized federated learning. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 15444–15453 (2025)

    2025

  42. [51]

    In: IEEE Conf

    Zhou, T., Brown, M., Snavely, N., Lowe, D.G.: Unsupervised learning of depth and ego-motion from video. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 6612–6619 (2017).https://doi.org/10.1109/CVPR.2017.700 FeDepth 1 FeDepth: Federated Learning for Depth Estimation under Robot ...

    2017 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.