Pith. sign in

REVIEW 3 major objections 6 minor 55 references

The paper proposes that bird's-eye-view (BEV) segmentation is better cast as a generative problem: instead of predicting one top-down map, a conditional variational autoencoder with normalizing flows produces multiple candidate maps, which

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 ↗

TVB combines a conditional variational autoencoder, normalizing flows, and attention-based fusion to produce bird's-eye-view segmentation from multiple car cameras, reporting small but consistent IoU gains on nuScenes and OPV2V.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A plausible but under-verified BEV segmentation paper whose core variational derivation in Eq. (11) is not a valid ELBO; the empirical story is coherent but the SOTA claim runs ahead of the baselines. the 3 major comments →

arxiv 2607.14710 v1 pith:WNAQHIV5 submitted 2026-07-16 cs.CV cs.SYeess.SY

Variational Inference for Bird's Eye View Segmentation in Autonomous Driving

Jingyue Shi , Huaicheng Li , Junhui Zhao , Yanxiang Jiang This is my paper
classification cs.CV cs.SYeess.SY
keywords bird's eye view segmentationvariational inferenceconditional variational autoencodernormalizing flowsattention fusionmulti-camera fusionautonomous drivingnuScenes
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 argues that bird's-eye-view (BEV) semantic segmentation, the task of producing a top-down map of a vehicle's surroundings from multiple cameras, is better treated as a generative problem than as a deterministic mapping. The proposed model, TVB, learns a conditional distribution of BEV maps through a conditional variational autoencoder whose latent distribution is made more expressive by normalizing flows. At inference it draws several candidate BEV maps and fuses them with an attention module rather than committing to a single prediction. On the nuScenes and OPV2V datasets, the authors report that this approach outperforms prior deterministic cross-view methods such as CVT and LaRa across vehicles, drivable area, and lane categories.

Core claim

TVB's central claim is that sampling multiple candidate BEV maps from a learned latent distribution and adaptively fusing them yields more accurate and robust BEV segmentation than producing a single map. The paper reports top IoU results on nuScenes (43.2 in Setting 1, +1.7 over the previous best; 41.3 for Vehicles and 76.6 for Drivable Area in Setting 2) and on OPV2V (Vehicles 56.8, +3.3; Drivable Area 61.3, +0.6; Lane 48.1, +2.3). The improvement is attributed to the combination of a conditional variational backbone, normalizing flows that enrich the latent distribution, and an attention-based fusion module that lets relatively better candidate maps contribute more. The paper also finds t

What carries the argument

The load-bearing component is CFLOW, a conditional variational autoencoder with an added K-step normalizing flow. A prior network encodes camera-fused BEV features x into a latent distribution pθ(z0|x); a posterior network, used only in training, encodes x together with ground-truth BEV c into qφ(z0|c,x). A KL term aligns the two, and the flow transforms the prior into a more expressive distribution pθ(zK|x). Monte Carlo sampling from this distribution and decoding produces multiple candidate BEV maps. The second component is the BEV-attention fusion (BAF) module, which computes confidence maps for each candidate, weights them with learned attention, and applies spatial attention to the fuse

Load-bearing premise

The inference-time advantage relies on the prior distribution pθ(z0|x) staying aligned with the posterior qφ(z0|c,x) learned with ground-truth supervision; that alignment is enforced only by the KL term during training, and if the input distribution shifts (night, rain, new camera calibration) the sampled candidates may become miscalibrated and attention fusion cannot recover a correct BEV.

What would settle it

If a deterministic baseline with the same cross-attention encoder and an equally strong decoder were trained on the same data and matched or exceeded TVB's IoU on the same benchmarks (or on a held-out night/rain split), then the generative sampling-and-fusion pipeline would not be the cause of the reported improvement.

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

If this is right

  • If the reported gains hold, multi-hypothesis generation with attention fusion becomes a viable alternative to single-shot BEV segmentation, particularly in occluded and long-range regions.
  • The use of normalizing flows in the latent space improves over a plain Gaussian CVAE; ablations report IoU rising from 41.8 to 43.2 (Setting 1) when flow is added.
  • Attention-based fusion (BAF) outperforms simple Monte Carlo averaging of candidates (Vehicles 56.8 vs 53.6 on OPV2V), suggesting that learned weighting of samples is a key contributor to the improvement.
  • The framework's main practical cost is inference speed: with 30 samples and K=4 flow steps, the model runs at 19 FPS on nuScenes, so real-time deployment would require reducing this overhead.

Where Pith is reading between the lines

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

  • Beyond the paper's claims, the spread across candidate BEV maps could double as an uncertainty estimate; regions where sampled maps disagree may indicate where the model is unsure, which could inform downstream planning decisions, though the paper does not test this.
  • The same CVAE-flow machinery could transfer to other structured prediction tasks in driving, such as occupancy flow prediction or future map prediction, where multiple plausible outcomes are natural.
  • A testable extension would be to distill the 30-sample fusion into a single deterministic pass (e.g., using a student network) and measure how much of the IoU gain survives; this would isolate whether the gains come from the generative distribution or from the fusion mechanism.
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

3 major / 6 minor

Summary. The paper proposes TVB, a CVAE-based BEV segmentation network. Multi-camera images are fused by cross-attention into BEV features x; a CVAE with a K-step normalizing flow ('CFLOW') is trained with ground-truth BEV c, using a prior pθ(z0|x) and a posterior qφ(z0|c,x). At inference, multiple latent samples are drawn from the flow-transformed prior, decoded into candidate BEV maps, and fused by an attention module BAF. Experiments on nuScenes and OPV2V report IoU improvements over CVT/LaRa and several cooperative fusion baselines, with ablations on loss terms, flow steps, fusion, and sample count. The paper claims state-of-the-art BEV segmentation, while also acknowledging limitations in inference speed and lane-structure accuracy.

Significance. The idea of treating BEV segmentation as a conditional generative problem and fusing sampled candidates is potentially interesting and timely. The paper has several strengths: it reports systematic ablations (Tables III-VI), visualizes intermediate flow steps (Fig. 8), evaluates on two datasets, and explicitly discusses efficiency and lane-structure limitations in Section IV-E. However, the central derivation is not yet correct, and the empirical SOTA claim is weakened by the absence of recent baselines and uncertainty estimates. If the objective is corrected and the reported gains survive re-running, the contribution would be a solid application of CVAE with normalizing-flow inference to BEV segmentation.

major comments (3)
  1. [Section III-C, Eqs. (8)-(11)] The ELBO is not well defined. In Eq. (11) the KL term is D_KL[q_phi(z0|c,x) || p_theta(zK|x)], but q_phi is a density over z0 and p_theta over zK=T(z0); a KL between densities on different spaces is undefined without a change of variables. Under the 'shared flow' reading of Fig. 4 the Jacobian factors cancel, so the third term in Eq. (11) should vanish; under the 'posterior-only flow' reading the objective is IAF, but then inference must sample from the untransformed prior p_theta(z0|x), not from p_theta(zK|x) as stated. The training objective and inference procedure are therefore mismatched. Table III's +1.4/+1.2 IoU attributed to the flow is not trustworthy until the objective is corrected and experiments are re-run.
  2. [Section III-C, Eqs. (8)-(10)] Eq. (8) treats a deterministic flow chain as a Markov chain of densities; since each T_k is bijective, q(zk|z_{k-1}) is degenerate and the product form is not the density of zK. Eq. (9) has a sign error: the change-of-variables formula requires a minus sign, not plus. Eq. (10) then contradicts Eq. (9). These are not typos: they make the method unreproducible from the equations.
  3. [Section IV-C, Tables I-II] The 'state-of-the-art' claim is not supported by the baseline selection. BEVFormer, CoBEVT, and BEVDiffuser are discussed in Section II-A but omitted from the comparisons; on nuScenes the strongest listed deterministic baselines are CVT and LaRa. No error bars, multiple seeds, or statistical significance tests are reported, so the +1.4 to +1.7 IoU gains over LaRa (Table I) may be within run-to-run variation. On OPV2V, all baselines are CVT-backbone fusion methods, so the comparison does not cover recent non-CVT or transformer-based cooperative methods.
minor comments (6)
  1. [Eq. (16)] The notation for IoU is informal for tensors; use element-wise product and indicator sums, and specify whether the metric is averaged per class or globally.
  2. [Table III] The table uses symbols like X and !; replace with conventional check/cross marks and define L_r and L_KL in the caption.
  3. [Eq. (12)] The hyperparameters lambda1, lambda2, lambda3 in Eq. (12) are never reported; include their values in Section IV-B.
  4. [Section III-B] The sentence 'distribution in the generative model, and p(z|x) is the prior distribution' is duplicated; remove the extra phrase.
  5. [Table I] Complete the missing FPS/Params entries for CVT, and specify which GPU/software versions are used for FPS comparisons.
  6. [Reproducibility] The paper does not provide code or model weights; please include a reproducibility statement or a public code link.

Circularity Check

0 steps flagged

No significant circularity: the derivation is a standard supervised CVAE+flow pipeline, and the empirical gains are measured against external baselines; the only self-citation is introductory and non-load-bearing.

full rationale

The paper's derivation chain starts from the standard conditional ELBO (Eqs. 1-3), then adds a K-step normalizing flow to the latent distributions (Eqs. 4-10) and defines the CFLOW objective (Eqs. 11-12). These equations do not define the target IoU numbers nor the BEV predictions in terms of themselves; the model is trained with ground-truth BEV supervision (posterior q_phi(z0|c,x)) and at inference samples from the prior p_theta(zK|x), followed by learned attention fusion. The reported gains in Tables I-III are empirical comparisons against CVT, LaRa, etc., not quantities that are fitted and then renamed as predictions. The only self-citation is ref. [3] (Zhao, Shi, Zhuo) used in the introduction for the definition/importance of BEV; it is not load-bearing for the method or results. The appended limitation statement in Sec. IV-E acknowledges sampling overhead and limited lane precision; this is a performance caveat, not an admission of circularity. The skeptic's concern that Eq. (11) mixes q_phi(z0) with p_theta(zK) and has a questionable Jacobian sign is a correctness/reproducibility issue, not a circularity pattern: a malformed or mismatched ELBO would make the training objective invalid, but it would not make the predicted BEV map equal to the model's input by construction. No fitted constant is relabeled as a prediction, no uniqueness argument is imported from the authors' prior work, and no external benchmark is replaced by a self-citation. Therefore the circularity score is low (1), reflecting only the minor self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The paper's contribution is an engineering combination of existing components: CVT-style attention, CVAE, and normalizing flows. No new physical or mathematical entities are postulated. The main free parameters are standard hyperparameters (loss weights, flow steps, sample count, latent dimension) chosen by validation-set ablations.

free parameters (4)
  • Loss weights λ1, λ2, λ3 = Not reported
    Chosen by hand to balance reconstruction, KL, and flow Jacobian terms in Eq. (12); values not disclosed, so the exact loss weighting is unspecified.
  • Number of normalizing flow steps K = 4
    Selected by ablation on OPV2V (Table IV) as an accuracy/speed trade-off.
  • Number of Monte Carlo samples = 30
    Selected by ablation (Table VI); performance peaks near 30 samples on both datasets.
  • Latent space dimension = 6
    Model design choice; no sensitivity study is provided.
axioms (5)
  • standard math Evidence lower bound (ELBO) and reparameterization trick as used in VAEs
    Used in Eq. (1)-(3) to derive the conditional variational objective.
  • standard math Normalizing flows are invertible and the change-of-variables formula holds
    Required for Eq. (4)-(7) and the Jacobian terms in the flow transformation.
  • standard math KL divergence properties and Jensen's inequality
    Underpins the ELBO derivation in Eq. (2).
  • domain assumption The fused BEV features x from cross-attention carry sufficient information to infer the BEV map
    The entire framework assumes that the multi-view fusion (Section III-A) produces a feature representation from which the BEV can be generated; if this fails, the variational model cannot compensate.
  • domain assumption Ground-truth BEV c is available during training and is reliable
    The posterior network is trained with ground-truth BEV supervision (Section III-B); the method relies on the correctness of dataset annotations.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Variational Inference for Bird's Eye View Segmentation in Autonomous Driving." pith.science (2026). https://pith.science/paper/WNAQHIV5

@misc{pith2026260714710,
  author       = {Pith},
  title        = {Pith review of: Variational Inference for Bird's Eye View Segmentation in Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WNAQHIV5}},
  note         = {Machine review of arXiv:2607.14710}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The bird's eye view (BEV) has emerged as a pivotal approach for environmental perception in autonomous driving, providing a unified spatial representation for vehicles. Nevertheless, despite BEV's significance in addressing the challenges inherent to autonomous driving, effectively fusing data from multiple camera sensors and operating in complex external driving environments remains a considerable challenge. To mitigate this issue, we recast the BEV segmentation problem within a variational inference framework. In this paper, we propose a novel transformer-based variational flow transformation network for BEV segmentation, denoted as TVB. Our architecture implicitly learns the mapping from multiple camera views to a unified canonical BEV map during training by exploiting posterior BEV supervision. TVB employs a conditional variational auto encoder (CVAE) as its backbone and produces multiple BEV map candidates. To augment the realism of the generated BEV maps, we integrate normalizing flows into the map generation process, enabling the construction of more complex and expressive probability distributions. Furthermore, we design a BEV-attention fusion (BAF) module that harnesses attention mechanisms to adaptively integrate the multiple candidate BEV maps. Experimental results, evaluated on both the nuScenes and OPV2Vdatasets, demonstrate that our proposed method achieves superior performance in multi-camera view BEV segmentation and lane environment perception.

Figures

Figures reproduced from arXiv: 2607.14710 by Huaicheng Li, Jingyue Shi, Junhui Zhao, Yanxiang Jiang.

Figure 1
Figure 1. Figure 1: The motivation and intention of our proposed method. (a) Some [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The architecture of TVB. During the training phase, TVB first fuses multiple camera images from different viewpoints to obtain BEV features [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Multiple views fusion architecture. Multiple camera images from [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The architecture of the CFLOW network. (a) The training process consists of solid and dashed lines: the prior network and posterior network obtain [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The framework of the proposed BEV-attention fusion module. Multiple confidence maps are obtained from various BEV maps, and an attention [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization examples of BEV segmentation results of TVB on the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of the BEV graph generated after each normalizing flow step [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization results of TVB in challenging scenarios. We present the visualization results of TVB in night and rainy scenarios, focusing on two [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 7 linked inside Pith

  1. [1]

    Tigdistill-bev: Multi-view bev 3d object detection via target inner-geometry learning distillation,

    S. Xu, F. Li, P. Huang, Z. Song, and Z.-X. Yang, “Tigdistill-bev: Multi-view bev 3d object detection via target inner-geometry learning distillation,”IEEE Trans. Circuits Syst. Video Technol., vol. 36, no. 1, pp. 846–860, 2026

    2026

  2. [2]

    Ralibev: Radar and lidar bev fusion learning for anchor box free object detection systems,

    Y . Yang, J. Liu, T. Huang, Q.-L. Han, G. Ma, and B. Zhu, “Ralibev: Radar and lidar bev fusion learning for anchor box free object detection systems,”IEEE Trans. Circuits Syst. Video Technol., vol. 35, no. 5, pp. 4130–4143, 2025

    2025

  3. [3]

    Bev perception for autonomous driving: State of the art and future perspectives,

    J. Zhao, J. Shi, and L. Zhuo, “Bev perception for autonomous driving: State of the art and future perspectives,”Expert Systems with Applica- tions, vol. 258, p. 125103, 2024

    2024

  4. [4]

    Drivemoe: Mixture-of-experts for vision-language-action model in end- to-end autonomous driving,

    Z. Yang, Y . Chai, X. Jia, Q. Li, Y . Shao, X. Zhu, H. Su, and J. Yan, “Drivemoe: Mixture-of-experts for vision-language-action model in end- to-end autonomous driving,”arXiv preprint arXiv:2505.16278, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Towards capturing the temporal dynamics for trajectory prediction: a coarse-to- fine approach,

    X. Jia, L. Chen, P. Wu, J. Zeng, J. Yan, H. Li, and Y . Qiao, “Towards capturing the temporal dynamics for trajectory prediction: a coarse-to- fine approach,” inConference on Robot Learning. PMLR, 2023, pp. 910–920

    2023

  6. [6]

    Computation offloading and resource allocation for cloud assisted mobile edge computing in vehicular networks,

    J. Zhao, Q. Li, Y . Gong, and K. Zhang, “Computation offloading and resource allocation for cloud assisted mobile edge computing in vehicular networks,”IEEE Trans. Veh. Technol., vol. 68, no. 8, pp. 7944– 7956, Jun 2019

    2019

  7. [7]

    Edge caching and computation management for real-time internet of vehicles: An online and distributed approach,

    J. Zhao, X. Sun, Q. Li, and X. Ma, “Edge caching and computation management for real-time internet of vehicles: An online and distributed approach,”IEEE Trans. Intell. Transp. Syst., vol. 22, no. 4, pp. 2183– 2197, Aug 2020

    2020

  8. [8]

    Multi- modality sensing and data fusion for multi-vehicle detection,

    D. Roy, Y . Li, T. Jian, P. Tian, K. Chowdhury, and S. Ioannidis, “Multi- modality sensing and data fusion for multi-vehicle detection,”IEEE Trans. Multimedia, vol. 25, pp. 2280–2295, 2023

    2023

  9. [9]

    Toward robust lidar- camera fusion in bev space via mutual deformable attention and temporal aggregation,

    J. Wang, F. Li, Y . An, X. Zhang, and H. Sun, “Toward robust lidar- camera fusion in bev space via mutual deformable attention and temporal aggregation,”IEEE Trans. Circuits Syst. Video Technol., vol. 34, no. 7, pp. 5753–5764, 2024. IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY 12

    2024

  10. [10]

    Adazoom: Towards scale-aware large scene object detection,

    J. Xu, Y .-L. Li, and S. Wang, “Adazoom: Towards scale-aware large scene object detection,”IEEE Transactions on Multimedia, vol. 25, pp. 4598–4609, 2023

    2023

  11. [11]

    Unisparsebev: A multi-task learning framework with unified sparse query for autonomous driving,

    H. Zhou, Y . Zhang, and H. Qi, “Unisparsebev: A multi-task learning framework with unified sparse query for autonomous driving,”IEEE Trans. Circuits Syst. Video Technol., vol. 36, no. 5, pp. 5788–5799, 2026

    2026

  12. [12]

    Inverse perspec- tive mapping simplifies optical flow computation and obstacle detection,

    H. A. Mallot, H. H. B ¨ulthoff, J. Little, and S. Bohrer, “Inverse perspec- tive mapping simplifies optical flow computation and obstacle detection,” Springer Biol. Cybern., vol. 64, no. 3, pp. 177–185, 1991

    1991

  13. [13]

    A sim2real deep learning ap- proach for the transformation of images from multiple vehicle-mounted cameras to a semantically segmented image in bird’s eye view,

    L. Reiher, B. Lampe, and L. Eckstein, “A sim2real deep learning ap- proach for the transformation of images from multiple vehicle-mounted cameras to a semantically segmented image in bird’s eye view,” inProc. IEEE Int. Conf. Intell. Trans. Syst. (ITSC), 2020, pp. 1–7

    2020

  14. [14]

    Local-to-global semantic learning for multi-view 3d object detection from point cloud,

    R. Qiao, H. Ji, Z. Zhu, and W. Zhang, “Local-to-global semantic learning for multi-view 3d object detection from point cloud,”IEEE Trans. Circuits Syst. Video Technol., vol. 34, no. 10, pp. 9371–9385, 2024

    2024

  15. [15]

    Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving,

    Y . Wang, W.-L. Chao, D. Garg, B. Hariharan, M. Campbell, and K. Q. Weinberger, “Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 8445–8453

    2019

  16. [16]

    Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving,

    X. Ma, Z. Wang, H. Li, P. Zhang, W. Ouyang, and X. Fan, “Accurate monocular 3d object detection via color-embedded 3d reconstruction for autonomous driving,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 6851–6860

    2019

  17. [17]

    Monocular se- mantic occupancy grid mapping with convolutional variational encoder– decoder networks,

    C. Lu, M. J. G. van de Molengraft, and G. Dubbelman, “Monocular se- mantic occupancy grid mapping with convolutional variational encoder– decoder networks,”IEEE Robot. Autom. Lett., vol. 4, no. 2, pp. 445–452, 2019

    2019

  18. [18]

    Cross-view semantic segmentation for sensing surroundings,

    B. Pan, J. Sun, H. Y . T. Leung, A. Andonian, and B. Zhou, “Cross-view semantic segmentation for sensing surroundings,”IEEE Robot. Autom. Lett., vol. 5, no. 3, pp. 4867–4873, 2020

    2020

  19. [19]

    Lif-seg: Lidar and camera image fusion for 3d lidar semantic segmentation,

    L. Zhao, H. Zhou, X. Zhu, X. Song, H. Li, and W. Tao, “Lif-seg: Lidar and camera image fusion for 3d lidar semantic segmentation,”IEEE Transactions on Multimedia, vol. 26, pp. 1158–1168, 2024

    2024

  20. [20]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Adv. Neural Inform. Process. Syst. (NeurIPS), vol. 30, 2017

    2017

  21. [21]

    Cross-view transformers for real-time map- view semantic segmentation,

    B. Zhou and P. Kr ¨ahenb¨uhl, “Cross-view transformers for real-time map- view semantic segmentation,” inProc. IEEE/CVF Conf. Comput. vis. Pattern Recognit. (CVPR), 2022, pp. 13 760–13 769

    2022

  22. [22]

    Lara: Latents and rays for multi-camera bird’s-eye-view semantic segmentation,

    F. Bartoccioni, ´E. Zablocki, A. Bursuc, P. P ´erez, M. Cord, and K. Ala- hari, “Lara: Latents and rays for multi-camera bird’s-eye-view semantic segmentation,” inProc. Conf. Robot Learn. (CoRL), 2023, pp. 1663– 1672

    2023

  23. [23]

    Learning structured output representation using deep conditional generative models,

    K. Sohn, H. Lee, and X. Yan, “Learning structured output representation using deep conditional generative models,”Adv. Neural Inform. Process. Syst. (NeurIPS), vol. 28, 2015

    2015

  24. [24]

    Learning discourse-level diversity for neural dialog models using conditional variational autoencoders,

    T. Zhao, R. Zhao, and M. Eskenazi, “Learning discourse-level diversity for neural dialog models using conditional variational autoencoders,” arXiv:1703.10960, 2017

    Pith/arXiv arXiv 2017

  25. [25]

    Normalizing flows for probabilistic modeling and inference,

    G. Papamakarios, E. Nalisnick, D. J. Rezende, S. Mohamed, and B. Lakshminarayanan, “Normalizing flows for probabilistic modeling and inference,”J. Mach. Learn. Res., vol. 22, no. 1, pp. 2617–2680, 2021

    2021

  26. [26]

    Improved variational inference with inverse autoregressive flow,

    D. P. Kingma, T. Salimans, R. Jozefowicz, X. Chen, I. Sutskever, and M. Welling, “Improved variational inference with inverse autoregressive flow,”Adv. Neural Inform. Process. Syst. (NeurIPS), vol. 29, 2016

    2016

  27. [27]

    Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to- vehicle communication,

    R. Xu, H. Xiang, X. Xia, X. Han, J. Li, and J. Ma, “Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to- vehicle communication,” inProc. Int. Conf. Robot. and Autom. (ICRA), 2022, pp. 2583–2589

    2022

  28. [28]

    nuscenes: A multimodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Kr- ishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 11 621–11 631

    2020

  29. [29]

    Automatic dense visual semantic mapping from street-level imagery,

    S. Sengupta, P. Sturgess, P. Torret al., “Automatic dense visual semantic mapping from street-level imagery,” inProc. IEEE Int. Conf. Intell. Robot. Syst. (IROS), pp. 857–862

  30. [30]

    Monocular 3d vehicle detection using uncalibrated traffic cameras through homography,

    M. Zhu, S. Zhang, Y . Zhong, P. Lu, H. Peng, and J. Lenneman, “Monocular 3d vehicle detection using uncalibrated traffic cameras through homography,” inProc. IEEE/RSJ Int. Conf. Intell. Robot. and Syst. (IROS), 2021, pp. 3814–3821

    2021

  31. [31]

    Driving among flatmobiles: Bird-eye-view occupancy grids from a monocular camera for holistic trajectory planning,

    A. Loukkal, Y . Grandvalet, T. Drummond, and Y . Li, “Driving among flatmobiles: Bird-eye-view occupancy grids from a monocular camera for holistic trajectory planning,” inProc. IEEE/CVF Win. Conf. Applicat. Comput. Vis. (WACV), 2021, pp. 51–60

    2021

  32. [32]

    Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,

    J. Philion and S. Fidler, “Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2020, pp. 194–210

    2020

  33. [33]

    Bevdet: High- performance multi-camera 3d object detection in bird-eye-view,

    J. Huang, G. Huang, Z. Zhu, Y . Ye, and D. Du, “Bevdet: High- performance multi-camera 3d object detection in bird-eye-view,” arXiv:2112.11790, 2021

    Pith/arXiv arXiv 2021

  34. [34]

    Flatfusion: Delving into details of sparse transformer-based camera-lidar fusion for autonomous driving,

    Y . Zhu, X. Jia, X. Yang, and J. Yan, “Flatfusion: Delving into details of sparse transformer-based camera-lidar fusion for autonomous driving,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 8581–8588

    2025

  35. [35]

    Bevsegformer: Bird’s eye view semantic segmentation from arbitrary camera rigs,

    L. Peng, Z. Chen, Z. Fu, P. Liang, and E. Cheng, “Bevsegformer: Bird’s eye view semantic segmentation from arbitrary camera rigs,” inProc. IEEE/CVF Win. Conf. Applicat. Comput. Vis. (WACV), 2023, pp. 5935– 5943

    2023

  36. [36]

    Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers,

    Z. Li, W. Wang, H. Li, E. Xie, C. Sima, T. Lu, Y . Qiao, and J. Dai, “Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2022, pp. 1–18

    2022

  37. [37]

    Cobevt: Cooper- ative bird’s eye view semantic segmentation with sparse transformers,

    R. Xu, Z. Tu, H. Xiang, W. Shao, B. Zhou, and J. Ma, “Cobevt: Cooper- ative bird’s eye view semantic segmentation with sparse transformers,” arXiv:2207.02202, 2022

    Pith/arXiv arXiv 2022

  38. [38]

    Bevdiffuser: Plug-and-play diffusion model for bev denoising with ground-truth guid- ance,

    X. Ye, B. Yaman, S. Cheng, F. Tao, A. Mallik, and L. Ren, “Bevdiffuser: Plug-and-play diffusion model for bev denoising with ground-truth guid- ance,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 1495–1504

    2025

  39. [39]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” arXiv:1312.6114, 2013

    Pith/arXiv arXiv 2013

  40. [40]

    Stochastic backpropa- gation and approximate inference in deep generative models,

    D. J. Rezende, S. Mohamed, and D. Wierstra, “Stochastic backpropa- gation and approximate inference in deep generative models,” inProc. Int. Conf. Mach. Learn. (ICML), 2014, pp. 1278–1286

    2014

  41. [41]

    Attribute2image: Conditional image generation from visual attributes,

    X. Yan, J. Yang, K. Sohn, and H. Lee, “Attribute2image: Conditional image generation from visual attributes,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2016, pp. 776–791

    2016

  42. [42]

    A variational u-net for conditional appearance and shape generation,

    P. Esser, E. Sutter, and B. Ommer, “A variational u-net for conditional appearance and shape generation,” inProceedings of the IEEE confer- ence on computer vision and pattern recognition, 2018, pp. 8857–8866

    2018

  43. [43]

    Gaze estimation by attention-induced hierarchical variational auto- encoder,

    G. Huang, J. Shi, J. Xu, J. Li, S. Chen, Y . Du, X. Zhen, and H. Liu, “Gaze estimation by attention-induced hierarchical variational auto- encoder,”IEEE Trans. Cybern., 2023

    2023

  44. [44]

    Bev-vae: Multi-view image generation with spatial consistency for autonomous driving,

    Z. Chen and H. Zhao, “Bev-vae: Multi-view image generation with spatial consistency for autonomous driving,”arXiv preprint arXiv:2507.00707, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Srflow: Learning the super-resolution space with normalizing flow,

    A. Lugmayr, M. Danelljan, L. Van Gool, and R. Timofte, “Srflow: Learning the super-resolution space with normalizing flow,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2020, pp. 715–732

    2020

  46. [46]

    Pointflow: 3d point cloud generation with continuous normalizing flows,

    G. Yang, X. Huang, Z. Hao, M.-Y . Liu, S. Belongie, and B. Hariharan, “Pointflow: 3d point cloud generation with continuous normalizing flows,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 4541–4550

    2019

  47. [47]

    Phiseg: Capturing uncertainty in medical image segmentation,

    C. F. Baumgartner, K. C. Tezcan, K. Chaitanya, A. M. H ¨otker, U. J. Muehlematter, K. Schawkat, A. S. Becker, O. Donati, and E. Konukoglu, “Phiseg: Capturing uncertainty in medical image segmentation,” inProc. Med. Image Comput. Comput. Assist. Int. (MICCAI), 2019, pp. 119–127

    2019

  48. [48]

    A hierarchical prob- abilistic u-net for modeling multi-scale ambiguities,

    S. A. Kohl, B. Romera-Paredes, K. H. Maier-Hein, D. J. Rezende, S. Es- lami, P. Kohli, A. Zisserman, and O. Ronneberger, “A hierarchical prob- abilistic u-net for modeling multi-scale ambiguities,”arXiv:1905.13077, 2019

    Pith/arXiv arXiv 1905

  49. [49]

    Identity mappings in deep residual networks,

    K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual networks,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2016, pp. 630–645

    2016

  50. [50]

    Fiery: Future instance prediction in bird’s- eye view from surround monocular cameras,

    A. Hu, Z. Murez, N. Mohan, S. Dudas, J. Hawke, V . Badrinarayanan, R. Cipolla, and A. Kendall, “Fiery: Future instance prediction in bird’s- eye view from surround monocular cameras,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 15 273–15 282

    2021

  51. [51]

    Enabling spatio- temporal aggregation in birds-eye-view vehicle estimation,

    A. Saha, O. Mendez, C. Russell, and R. Bowden, “Enabling spatio- temporal aggregation in birds-eye-view vehicle estimation,” inProc. IEEE Int. conf. Robot. Autom. (ICRA), 2021, pp. 5133–5139

    2021

  52. [52]

    Simpb: A single model for 2d and 3d object detection from multiple cameras,

    Y . Tang, Z. Meng, G. Chen, and E. Cheng, “Simpb: A single model for 2d and 3d object detection from multiple cameras,” inEuropean conference on computer vision. Springer, 2024, pp. 1–17

    2024

  53. [53]

    F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds,

    Q. Chen, X. Ma, S. Tang, J. Guo, Q. Yang, and S. Fu, “F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds,” inProc. 4th ACM/IEEE Symp. Edge Comput. (SEC), 2019, pp. 88–100

    2019

  54. [54]

    V2vnet: Vehicle-to-vehicle communication for joint perception and prediction,

    T.-H. Wang, S. Manivasagam, M. Liang, B. Yang, W. Zeng, and R. Ur- tasun, “V2vnet: Vehicle-to-vehicle communication for joint perception and prediction,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2020, pp. 605–621. IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY 13

    2020

  55. [55]

    Learning distilled collaboration graph for multi-agent perception,

    Y . Li, S. Ren, P. Wu, S. Chen, C. Feng, and W. Zhang, “Learning distilled collaboration graph for multi-agent perception,”Adv. Neural Inform. Process. Syst. (NeurIPS), vol. 34, pp. 29 541–29 552, 2021

    2021

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.