Pith. sign in

REVIEW 3 major objections 2 minor 4 cited by

ExtraGS: Geometric-Aware Trajectory Extrapolation with Uncertainty-Guided Generative Priors

T0 review · 3 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read ExtraGS claims that trajectory extrapolation for driving scenes can be made both geometrically consistent and realistic by combining a hybrid Gaussian-SDF road-surface representation, far-field Gaussians, and uncertainty-gated selective use

desk verdict A plausible new combination for trajectory extrapolation, but the uncertainty-gating mechanism has a load-bearing circularity question that the abstract does not answer. read the letter →

arxiv 2508.15529 v2 pith:F7TDPIBJ submitted 2025-08-21 cs.CV

classification cs.CV
keywords drivingscenesimulationtrajectoryextrapolation3DGaussiansplattingsigneddistancefunctionfar-fieldGaussiansself-superviseduncertaintygenerativepriorsnovelviewsynthesis
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

Trajectory extrapolation—synthesizing what a vehicle would see beyond a recorded log—is a hard simulation problem because generative models alone tend to invent geometry and blur details. ExtraGS tries to establish that this can be fixed by anchoring the synthetic scene in explicit geometric structure and letting a generative prior (a pretrained model used as pseudo ground truth) contribute only where extrapolation artifacts occur. Its three coupled mechanisms are a hybrid Gaussian-Signed Distance Function (SDF) representation for road surfaces, far-field Gaussians with learnable scaling for distant objects, and a spherical-harmonics-based uncertainty gate trained without external supervision. If the paper is right, driving simulators can generate extrapolated multi-camera views with better realism and geometric consistency than prior generative-prior approaches, while preserving fidelity along the original trajectory.

What carries the argument

The framework rests on three coordinated pieces. Road Surface Gaussian (RSG): a hybrid Gaussian-Signed Distance Function (SDF) representation—simultaneously a set of Gaussian splats and a signed distance field—that anchors the road as an explicit surface rather than a free-floating volume. Far Field Gaussians (FFG): Gaussian primitives with learnable scale factors, so distant objects do not suffer from a single fixed resolution. Uncertainty gating: an estimator built on spherical harmonics (a standard basis for functions on a sphere), trained self-supervised, that predicts where extrapolation artifacts appear; generative priors are blended in only at those flagged locations. The central oper

What would settle it

On a held-out driving log, synthesize an extrapolated trajectory, render depth from the RSG road surface at those extrapolated positions, and compare it with LiDAR point clouds captured on a second pass through the same road. If the rendered depth error is not smaller than a generative-prior-only baseline, the geometric-consistency claim is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is that the geometric inconsistency and over-smoothing seen in generative-prior extrapolation are not inevitable. Road surfaces are modeled with a hybrid Gaussian-SDF primitive that gives the extrapolated region an explicit surface constraint; far-field Gaussians with learnable scaling stop distant objects from being represented by primitives at the wrong resolution; and a self-supervised spherical-harmonics uncertainty estimator flags where extrapolation artifacts occur, so generative-prior content is integrated only in those regions. Across datasets, multi-camera configurations, and different generative priors, ExtraGS reports improved realism and geomet

Load-bearing premise

The whole approach assumes the generative model's errors are local, not systematic: if a prior keeps making the same wrong prediction in the same place, the self-supervised gate will be trained to call that wrong place trustworthy, and the selective blending will preserve the error.

Editorial extensions

If this is right

  • Driving simulators could extrapolate multi-camera logs into views beyond the recorded trajectory with geometrically consistent road surfaces, not just plausible pixels.
  • The framework should transfer across generative priors, because the uncertainty gate is trained to detect artifacts rather than to match one specific generator.
  • Fidelity along the original trajectory is not sacrificed, since generative content is added only where the uncertainty signal says extrapolation went wrong.
  • Because the representation is explicit—Gaussian primitives plus an SDF—extrapolated views carry usable geometry, which downstream depth or planning modules could consume directly.
  • A single self-supervised uncertainty signal replaces per-scene supervision for deciding where generative priors apply.

Reading between the lines

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

  • The same recipe—an explicit surface anchor plus uncertainty-gated generative completion—could apply to trajectory extrapolation beyond driving, such as indoor navigation or robot camera paths, though the paper does not make this claim.
  • If the spherical-harmonics uncertainty map is accurate, it could be used directly as a per-region confidence layer for generated frames, independent of the rendering pipeline—an output the paper does not mention.
  • The learnable far-field scaling suggests a continuous level-of-detail mechanism; a natural test would be whether extrapolation quality degrades gracefully with distance and whether the learned scales correlate with metric depth.
  • A systematic rather than localized bias in the generative prior would be a stress case: the gate could be trained to trust exactly the wrong artifact regions, a failure mode the paper does not address.
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

3 major / 2 minor

Summary. The paper proposes ExtraGS, a framework for extrapolating novel views along a recorded driving trajectory. It combines geometric and generative priors through three mechanisms: a Road Surface Gaussian (RSG) representation based on a hybrid Gaussian-SDF design, Far Field Gaussians (FFG) with learnable scaling factors, and a self-supervised uncertainty estimation framework based on spherical harmonics (SH) that selectively integrates generative priors only where extrapolation artifacts occur. The abstract claims significant enhancements in realism and geometric consistency across multiple datasets, multi-camera setups, and generative priors, while preserving fidelity along the original trajectory.

Significance. If the claims are correct, ExtraGS would address a known limitation of generative-prior-based trajectory extrapolation: poor geometric consistency and over-smoothing. The proposed hybrid geometric-generative integration with uncertainty gating is structurally plausible and could be a meaningful improvement over naive blending of prior outputs. However, the abstract provides no quantitative evidence, ablations, or formal specifications, so the significance cannot be assessed from the current material. The strength of the contribution depends entirely on details that are not disclosed in the abstract, particularly the design of the uncertainty gate and its training signal.

major comments (3)
  1. [Abstract (central claim)] The abstract asserts that ExtraGS 'significantly enhances the realism and geometric consistency of extrapolated views,' but provides no quantitative metrics, comparisons, error bars, or ablations. This is load-bearing for the paper's central claim. The full text may contain such evidence, but the abstract alone does not support the assertion. Please provide at least one concrete metric (e.g., FID, KID, geometric reprojection error) with baseline comparisons and variance/error bars.
  2. [Abstract (uncertainty gating)] The self-supervised SH-based uncertainty gate is the core mechanism that distinguishes ExtraGS from prior work. The abstract states that generative priors 'often lead to poor geometric consistency and over-smoothed renderings'—i.e., systematic errors—while the gate is supposed to selectively integrate priors 'only where extrapolation artifacts occur.' If the gate's training signal is derived from the same prior outputs (e.g., pseudo ground truth), it may assign low uncertainty to systematic errors, causing the gate to trust the prior precisely where it is wrong. The abstract does not specify the training signal or provide evidence that the gate's uncertainty correlates with actual reconstruction error. This point is central to the claimed geometric-consistency improvement and must be clarified and validated, for example by comparing predicted uncertainty against per-pixel or per-region e
  3. [Abstract (RSG/FFG representation)] The 'hybrid Gaussian-SDF design' and 'learnable scaling factors' for Far Field Gaussians are described only at a high level. No formal definition, optimization procedure, or ablation is given. Without these details, it is impossible to verify that the representation actually addresses the geometric inconsistency of generative priors or that it is not simply a set of free parameters that compensate numerically. Please provide the explicit formulation and an ablation isolating the contribution of each component.
minor comments (2)
  1. [Abstract (specificity)] 'Multiple datasets, diverse multi-camera setups, and various generative priors' should list the datasets, camera configurations, and prior models. This is needed for reproducibility and for the reader to judge the breadth of the evaluation.
  2. [Abstract (language)] 'Significantly enhances' is a strong claim without quantitative anchoring. Consider reporting the effect size or at least naming a specific evaluation metric in the abstract.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified in abstract-only review

full rationale

The provided manuscript is abstract-only, so the derivation chain cannot be inspected beyond the claims stated. The abstract presents a method that integrates geometric and generative priors and uses a self-supervised uncertainty estimation framework based on spherical harmonics to selectively integrate generative priors where artifacts occur. No equations, training-loss definitions, or explicit fitted-vs-predicted quantities are visible. The potential concern that the uncertainty gate may be trained against the same generative priors it is meant to correct is speculative; the abstract does not state that the priors serve as pseudo ground truth for ExtraGS's own training, nor does it define the training signal for the uncertainty estimator. Without concrete text showing a parameter fitted to a target and then renamed as a prediction, or a definition that reduces Equation X to Equation Y by construction, no circular step can be exhibited under the hard rules. There is also no evidence of self-citation or imported uniqueness theorems. Therefore, the appropriate finding is no significant circularity, with a score of 0. If the full text reveals that the uncertainty gate is trained using the same generative prior outputs as ground truth, that would warrant re-evaluation, but such a claim cannot be made from the abstract alone.

Assumptions & free parameters 2 free parameters · 4 assumptions · 2 invented entities

Reconstructed from the abstract only. The framework rests on unproven premises: generative priors usable as pseudo ground truth despite being characterized as geometrically inconsistent; adequacy of the hybrid Gaussian-SDF road surface; and reliability of the SH uncertainty gate. The FFG learnable scaling factors are explicit fitted quantities. The full text is required to check whether any of these are derived, ablated, or externally validated.

free parameters (2)
  • Far Field Gaussian learnable scaling factors = learned during training, values not stated
    The abstract says FFG use 'learnable scaling factors' to handle distant objects; these are fitted on the paper's data and control distant-scene fidelity, but their values, regularization, and per-scene generalization are not described.
  • SH-based uncertainty gate parameters = learned, values not stated
    The self-supervised uncertainty framework is a learned component that decides where extrapolation artifacts occur and hence where generative priors are injected; its network weights and any gating thresholds directly affect the realism vs. fidelity tradeoff claimed in the abstract.
assumptions (4)
  • domain assumption Generative priors are reliable enough to serve as pseudo ground truth for training.
    The abstract states prior methods 'leverage generative priors as pseudo ground truth' and ExtraGS selectively integrates them; if the prior outputs are systematically wrong where geometry fails, the supervision signal is corrupted.
  • domain assumption The hybrid Gaussian-SDF RSG representation faithfully models road surface geometry.
    Road Surface Gaussians are asserted to provide geometric consistency; the representational adequacy of the hybrid design is assumed and not proven in the abstract.
  • domain assumption The evaluation datasets, camera setups, and generative priors are representative and the reported gains generalize.
    The abstract claims 'multiple datasets, diverse multi-camera setups, and various generative priors' without naming datasets, sizes, or statistics.
  • ad hoc to paper The SH-based self-supervised uncertainty signal correctly localizes extrapolation artifacts.
    This is a bespoke mechanism introduced by the paper; its reliability as an artifact detector is internal to the method rather than externally validated.
invented entities (2)
  • Road Surface Gaussian (RSG)
    purpose: Hybrid Gaussian-SDF representation of the road surface to maintain geometric consistency in extrapolated views.
    Introduced by this paper; no external falsifiable handle is offered in the abstract, and its adequacy is demonstrated only through the paper's own experiments.
  • Far Field Gaussians (FFG)
    purpose: Gaussians with learnable scaling factors for efficient rendering of distant objects.
    Introduced by this paper as a representational component; no independent verification mechanism is described in the abstract.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ExtraGS: Geometric-Aware Trajectory Extrapolation with Uncertainty-Guided Generative Priors." pith.science (2026). https://pith.science/paper/F7TDPIBJ

@misc{pith2026250815529,
  author       = {Pith},
  title        = {Pith review of: ExtraGS: Geometric-Aware Trajectory Extrapolation with Uncertainty-Guided Generative Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F7TDPIBJ}},
  note         = {Machine review of arXiv:2508.15529}
}
read the original abstract

Synthesizing extrapolated views from recorded driving logs is critical for simulating driving scenes for autonomous driving vehicles, yet it remains a challenging task. Recent methods leverage generative priors as pseudo ground truth, but often lead to poor geometric consistency and over-smoothed renderings. To address these limitations, we propose ExtraGS, a holistic framework for trajectory extrapolation that integrates both geometric and generative priors. At the core of ExtraGS is a novel Road Surface Gaussian(RSG) representation based on a hybrid Gaussian-Signed Distance Function (SDF) design, and Far Field Gaussians (FFG) that use learnable scaling factors to efficiently handle distant objects. Furthermore, we develop a self-supervised uncertainty estimation framework based on spherical harmonics that enables selective integration of generative priors only where extrapolation artifacts occur. Extensive experiments on multiple datasets, diverse multi-camera setups, and various generative priors demonstrate that ExtraGS significantly enhances the realism and geometric consistency of extrapolated views, while preserving high fidelity along the original trajectory.

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Mind the Gap: Standard 3DGS Evaluation Primarily Measures Near-Trajectory Interpolation

    cs.CV 2026-07 unverdicted novelty 7.0 of 10

    Standard every-Nth-frame holdouts in 3D scene reconstruction primarily measure near-trajectory interpolation, with a consistent 3-12 dB gap to matched-count contiguous spatial holdouts that persists across Gaussian, n...

  2. M$^\text{4}$World: A Multi-view Multimodal Driving World Model for Interactive Object Manipulation and Minute-long Streaming

    cs.CV 2026-07 conditional novelty 6.0 of 10

    M4World is a controllable multi-view camera+LiDAR driving world model with object-level appearance control, four-step causal streaming, and few-clip long-tail adaptation.

  3. Xiaomi Auto World Model: A Joint World Model Integrating Reconstruction and Generation for Autonomous Driving

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    Xiaomi EV World Model integrates WorldRec for sparse-query 3D Gaussian reconstruction and WorldGen for fast causal video generation via bidirectional pretraining and causal fine-tuning to support autonomous driving si...

  4. Xiaomi Auto World Model: A Joint World Model Integrating Reconstruction and Generation for Autonomous Driving

    cs.CV 2026-05 unverdicted novelty 5.0 of 10

    A unified system integrating sparse-query 3D Gaussian reconstruction with multi-stage causal video generation for autonomous driving world models.

Reference graph

Works this paper leans on

42 extracted references · 27 canonical work pages · cited by 3 Pith papers

  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]

    H.; Vora, S.; Liong, V

    Caesar, H.; Bankiti, V.; Lang, A. H.; Vora, S.; Liong, V. E.; Xu, Q.; Krishnan, A.; Pan, Y.; Baldan, G.; and Beijbom, O. 2020. nuScenes: A multimodal dataset for autonomous driving. In CVPR

  4. [4]

    Che, Q.-H.; Nguyen, D.-P.; Pham, M.-Q.; and Lam, D.-K. 2023. TwinLiteNet: An Efficient and Lightweight Model for Driveable Area and Lane Segmentation in Self-Driving Cars. In 2023 International Conference on Multimedia Analysis and Pattern Recognition (MAPR), 1--6

  5. [5]

    Chen, Y.; Gu, C.; Jiang, J.; Zhu, X.; and Zhang, L. 2023. Periodic Vibration Gaussian: Dynamic Urban Scene Reconstruction and Real-time Rendering. arXiv:2311.18561

  6. [6]

    M.; Ivanovic, B.; Litany, O.; Gojcic, Z.; Fidler, S.; Pavone, M.; Song, L.; and Wang, Y

    Chen, Z.; Yang, J.; Huang, J.; de Lutio, R.; Esturo, J. M.; Ivanovic, B.; Litany, O.; Gojcic, Z.; Fidler, S.; Pavone, M.; Song, L.; and Wang, Y. 2024. OmniRe: Omni Urban Scene Reconstruction. arXiv preprint arXiv:2408.16760

  7. [7]

    R.; Zhou, Y.; Yang, Z.; Chouard, A.; Sun, P.; Ngiam, J.; Vasudevan, V.; McCauley, A.; Shlens, J.; and Anguelov, D

    Ettinger, S.; Cheng, S.; Caine, B.; Liu, C.; Zhao, H.; Pradhan, S.; Chai, Y.; Sapp, B.; Qi, C. R.; Zhou, Y.; Yang, Z.; Chouard, A.; Sun, P.; Ngiam, J.; Vasudevan, V.; McCauley, A.; Shlens, J.; and Anguelov, D. 2021. Large Scale Interactive Motion Forecasting for Autonomous Driving: The Waymo Open Motion Dataset. In Proceedings of the IEEE/CVF Internationa...

  8. [8]

    Fan, L.; Zhang, H.; Wang, Q.; Li, H.; and Zhang, Z. 2024. FreeSim: Toward Free-viewpoint Camera Simulation in Driving Scenes. arXiv:2412.03566

Show all 42 references
  1. [9]

    Gao, R.; Chen, K.; Xiao, B.; Hong, L.; Li, Z.; and Xu, Q. 2025. MagicDrive-V2 : High-Resolution Long Video Generation for Autonomous Driving with Adaptive Control. In Proceedings of the IEEE/CVF International Conference on Computer Vision

  2. [10]

    Gropp, A.; Yariv, L.; Haim, N.; Atzmon, M.; and Lipman, Y. 2020. Implicit Geometric Regularization for Learning Shapes. In Proceedings of Machine Learning and Systems 2020, 3569--3579

  3. [11]

    Guo, J.; Deng, N.; Li, X.; Bai, Y.; Shi, B.; Wang, C.; Ding, C.; Wang, D.; and Li, Y. 2023. StreetSurf: Extending Multi-view Implicit Surface Reconstruction to Street Views. arXiv preprint arXiv:2306.04988

  4. [12]

    Hess, G.; Lindstr \"o m, C.; Fatemi, M.; Petersson, C.; and Svensson, L. 2024. SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving. arXiv preprint arXiv:2411.16816

  5. [13]

    Huang, B.; Yu, Z.; Chen, A.; Geiger, A.; and Gao, S. 2024. 2D Gaussian Splatting for Geometrically Accurate Radiance Fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery

  6. [14]

    Jocher, G.; and Qiu, J. 2024. Ultralytics YOLO11

  7. [15]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4)

  8. [16]

    H.; Unberath, M.; Liu, M.-Y.; and Lin, C.-H

    Li, Z.; M\"uller, T.; Evans, A.; Taylor, R. H.; Unberath, M.; Liu, M.-Y.; and Lin, C.-H. 2023. Neuralangelo: High-Fidelity Neural Surface Reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition ( CVPR )

  9. [17]

    Liao, B.; Chen, S.; Zhang, Y.; Jiang, B.; Zhang, Q.; Liu, W.; Huang, C.; and Wang, X. 2024. Maptrv2: An end-to-end framework for online vectorized hd map construction. International Journal of Computer Vision, 1--23

  10. [18]

    Liu, K.; Shao, L.; and Lu, S. 2024. Novel View Extrapolation with Video Diffusion Priors. arXiv preprint arXiv:2411.14208

  11. [19]

    Liu, X.; Zhou, C.; and Huang, S. 2024. 3DGS-Enhancer: Enhancing Unbounded 3D Gaussian Splatting with View-consistent 2D Diffusion Priors. In Globerson, A.; Mackey, L.; Belgrave, D.; Fan, A.; Paquet, U.; Tomczak, J.; and Zhang, C., eds., Advances in Neural Information Processin...

  12. [20]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2020. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV

  13. [21]

    Ni, C.; Zhao, G.; Wang, X.; Zhu, Z.; Qin, W.; Huang, G.; Liu, C.; Chen, Y.; Wang, Y.; Zhang, X.; Zhan, Y.; Zhan, K.; Jia, P.; Lang, X.; Wang, X.; and Mei, W. 2024. ReconDreamer: Crafting World Models for Driving Scene Reconstruction via Online Restoration

  14. [22]

    Ost, J.; Mannan, F.; Thuerey, N.; Knodt, J.; and Heide, F. 2021. Neural Scene Graphs for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2856--2865

  15. [23]

    P.; Mildenhall, B.; Fridovich-Keil, S.; Raghavan, N.; Singhal, U.; Ramamoorthi, R.; Barron, J

    Tancik, M.; Srinivasan, P. P.; Mildenhall, B.; Fridovich-Keil, S.; Raghavan, N.; Singhal, U.; Ramamoorthi, R.; Barron, J. T.; and Ng, R. 2020. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. arXiv:2006.10739

  16. [24]

    Tonderski, A.; Lindstr \"o m, C.; Hess, G.; Ljungbergh, W.; Svensson, L.; and Petersson, C. 2023. NeuRAD: Neural Rendering for Autonomous Driving. arXiv preprint arXiv:2311.15260

  17. [25]

    Y.; Ferroni, F.; and Ramanan, D

    Turki, H.; Zhang, J. Y.; Ferroni, F.; and Ramanan, D. 2023. SUDS: Scalable Urban Dynamic Scenes. In Computer Vision and Pattern Recognition (CVPR)

  18. [26]

    Wang, P.; Liu, L.; Liu, Y.; Theobalt, C.; Komura, T.; and Wang, W. 2021. NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view Reconstruction. arXiv preprint arXiv:2106.10689

  19. [27]

    Wang, Q.; Fan, L.; Wang, Y.; Chen, Y.; and Zhang, Z. 2024. FreeVS: Generative View Synthesis on Free Driving Trajectory. arXiv:2410.18079

  20. [28]

    Z.; Zhang, Y.; Turki, H.; Ren, X.; Gao, J.; Shou, M

    Wu, J. Z.; Zhang, Y.; Turki, H.; Ren, X.; Gao, J.; Shou, M. Z.; Fidler, S.; Gojcic, Z.; and Ling, H. 2025. DIFIX3D+: Improving 3D Reconstructions with Single-Step Diffusion Models. In Proceedings of the Computer Vision and Pattern Recognition Conference, 26024--26035

  21. [29]

    Wu, Z.; Liu, T.; Luo, L.; Zhong, Z.; Chen, J.; Xiao, H.; Hou, C.; Lou, H.; Chen, Y.; Yang, R.; Huang, Y.; Ye, X.; Yan, Z.; Shi, Y.; Liao, Y.; and Zhao, H. 2023. MARS: An Instance-aware, Modular and Realistic Simulator for Autonomous Driving. CICAI

  22. [30]

    Xie, Z.; Zhang, J.; Li, W.; Zhang, F.; and Zhang, L. 2023. S-NeRF: Neural Radiance Fields for Street Views. In International Conference on Learning Representations (ICLR)

  23. [31]

    Yan, Y.; Lin, H.; Zhou, C.; Wang, W.; Sun, H.; Zhan, K.; Lang, X.; Zhou, X.; and Peng, S. 2024. Street Gaussians: Modeling Dynamic Urban Scenes with Gaussian Splatting. In ECCV

  24. [32]

    Yan, Y.; Xu, Z.; Lin, H.; Jin, H.; Guo, H.; Wang, Y.; Zhan, K.; Lang, X.; Bao, H.; Zhou, X.; and Peng, S. 2025. StreetCrafter: Street View Synthesis with Controllable Video Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  25. [33]

    Yang, C.; Chen, Y.; Tian, H.; Tao, C.; Zhu, X.; Zhang, Z.; Huang, G.; Li, H.; Qiao, Y.; Lu, L.; Zhou, J.; and Dai, J. 2022. BEVFormer v2: Adapting Modern Image Backbones to Bird's-Eye-View Recognition via Perspective Supervision. ArXiv

  26. [34]

    W.; Li, B.; Che, T.; Xu, D.; Fidler, S.; Pavone, M.; and Wang, Y

    Yang, J.; Ivanovic, B.; Litany, O.; Weng, X.; Kim, S. W.; Li, B.; Che, T.; Xu, D.; Fidler, S.; Pavone, M.; and Wang, Y. 2023 a . EmerNeRF: Emergent Spatial-Temporal Scene Decomposition via Self-Supervision. arXiv:2311.02077

  27. [35]

    J.; and Urtasun, R

    Yang, Z.; Chen, Y.; Wang, J.; Manivasagam, S.; Ma, W.-C.; Yang, A. J.; and Urtasun, R. 2023 b . UniSim: A Neural Closed-Loop Sensor Simulator. In CVPR

  28. [36]

    Yang, Z.; Pan, Z.; Yang, Y.; Zhu, X.; and Zhang, L. 2024. Driving Scene Synthesis on Free-form Trajectories with Generative Prior. arXiv preprint arXiv:2412.01717

  29. [37]

    Yu, W.; Xing, J.; Yuan, L.; Hu, W.; Li, X.; Huang, Z.; Gao, X.; Wong, T.-T.; Shan, Y.; and Tian, Y. 2024 a . ViewCrafter: Taming Video Diffusion Models for High-fidelity Novel View Synthesis. arXiv preprint arXiv:2409.02048

  30. [38]

    Yu, Z.; Wang, H.; Yang, J.; Wang, H.; Xie, Z.; Cai, Y.; Cao, J.; Ji, Z.; and Sun, M. 2024 b . SGD: Street View Synthesis with Gaussian Splatting and Diffusion Prior. arXiv:2403.20079

  31. [39]

    Zhao, G.; Ni, C.; Wang, X.; Zhu, Z.; Zhang, X.; Wang, Y.; Huang, G.; Chen, X.; Wang, B.; Zhang, Y.; Mei, W.; and Wang, X. 2024. DriveDreamer4D: World Models Are Effective Data Machines for 4D Driving Scene Representation

  32. [40]

    Zhao, G.; Wang, X.; Ni, C.; Zhu, Z.; Qin, W.; Huang, G.; and Wang, X. 2025. ReconDreamer++: Harmonizing Generative and Reconstructive Models for Driving Scene Representation

  33. [41]

    Zhou, H.; Shao, J.; Xu, L.; Bai, D.; Qiu, W.; Liu, B.; Wang, Y.; Geiger, A.; and Liao, Y. 2024 a . HUGS: Holistic Urban 3D Scene Understanding via Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 21336--21345

  34. [42]

    Zhou, X.; Lin, Z.; Shan, X.; Wang, Y.; Sun, D.; and Yang, M.-H. 2024 b . Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21634--21643

Pith tools

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