Pith. sign in

REVIEW 5 major objections 7 minor 38 references

I2V-GS: Infrastructure-to-Vehicle View Transformation with Gaussian Splatting for Autonomous Driving Data Generation

T0 review · 5 major / 7 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 I2V-GS synthesizes vehicle-perspective images from fixed roadside cameras by augmenting Gaussian Splatting training with LiDAR-calibrated, diffusion-inpainted pseudo views, claiming state-of-the-art quality on the new RoadSight dataset.

desk verdict A sensible I2V rendering pipeline whose headline numbers are undercut by an evaluation protocol with no vehicle-view ground truth; worth a serious referee but needs major experimental fixes. read the letter →

arxiv 2507.23683 v1 pith:ANTFGH2P submitted 2025-07-31 cs.CV

Jialei Chen , Wuhao Xu , Sipeng He , Baoru Huang , Dongchun Ren This is my paper
classification cs.CV
keywords GaussianSplattinginfrastructure-to-vehicleviewtransformationautonomousdrivingdatagenerationnovelsynthesispseudodiffusioninpaintingsparse-viewreconstructionRoadSightdataset
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

Infrastructure cameras mounted at intersections are inexpensive and always on, but self-driving models must see the road from a vehicle's perspective. This paper tries to bridge that gap by training a Gaussian Splatting scene from a few fixed roadside viewpoints and then rendering novel vehicle-level views, turning one intersection into many vehicle-view training frames. The core proposal is I2V-GS, which the authors claim is the first framework for infrastructure-to-vehicle view transformation in autonomous driving data generation, together with RoadSight, a multi-modality dataset of real intersections. On RoadSight, the method reports state-of-the-art synthesis quality, improving NTA-IoU by 45.7%, NTL-IoU by 34.2%, and FID by 14.9% relative to StreetGaussian. If the approach holds, autonomous driving training data can be synthesized cheaply and at scale from existing roadside sensors rather than dedicated data-collection fleets.

What carries the argument

The load-bearing mechanism is the adaptive depth warp, a forward-warping step that projects rendered 3D points into a nearby pseudo view while bounding pixel displacement by a preset epsilon so that occlusions and geometric errors stay manageable. LiDAR-anchored monocular depth calibration, fitting the parameters c1, c2, and c3 in the depth formula to LiDAR measurements, supplies the geometry that makes warping reliable. A cascade strategy iteratively applies a latent diffusion model to inpaint warped-image holes, using the previous pseudo view as a prior to enforce cross-view consistency, and confidence-guided optimization weights each pseudo view's supervision by W = lambda1 * (1 - L2) + (1 - lambda1) * SSIM between rendered and inpainted pixels. Together these components turn sparse infrastructure images into dense training views that let the Gaussian Splatting model extrapolate to vehicle trajectories.

What would settle it

Set up an intersection with a real vehicle-view camera at exactly the pose the method is asked to synthesize; train I2V-GS using only the roadside infrastructure views, then compare the render to the real vehicle-view image, especially in regions the diffusion model inpainted. If the synthesized content there, such as a traffic light color or a parked vehicle, is systematically invented and does not match ground truth, the claim that pseudo views are reliable training targets is falsified.

Watch

Extended reading notes

Core claim

I2V-GS claims that a 3D Gaussian scene, optimized from sparse fixed infrastructure views, can render photorealistic vehicle-perspective images despite large viewpoint changes when the training set is augmented with dense pseudo views. The pseudo views are generated by forward-warping current renderings to nearby poses using an adaptive depth warp whose displacement is bounded to avoid geometric over- and under-warping, with monocular depth calibrated to LiDAR to anchor geometry. Occlusion holes are then filled by a latent diffusion model, and a cascade strategy chains pseudo views so inpainted content stays consistent across frames. Because the diffusion model can hallucinate, a cross-view confidence weight derived from L2 and SSIM differences between the rendered and inpainted views down-weights unreliable regions during optimization. The paper reports that this pipeline outperforms StreetGaussian on RoadSight by 45.7% in NTA-IoU, 34.2% in NTL-IoU, and 14.9% in FID, and on Waymo by 29.7% in NTA-IoU, 4.2% in NTL-IoU, and a 33.3% FID reduction.

Load-bearing premise

The pipeline treats the diffusion-inpainted pseudo views as ground truth for training; the assumption is that what the diffusion model invents to fill occlusion holes matches the actual scene closely enough that the Gaussian model learns true geometry rather than hallucinated content, and confidence weighting only reduces the influence of mismatched regions instead of repairing an underlying wrong scene structure.

Editorial extensions

If this is right

  • Synthesized vehicle views from fixed roadside cameras can expand a single intersection into many training trajectories, reducing reliance on expensive dedicated vehicle fleets.
  • The improvements reported on RoadSight (NTA-IoU +45.7%, NTL-IoU +34.2%, FID -14.9% vs StreetGaussian) imply that the pseudo-view cascade helps most in large-extrapolation settings where prior 3DGS driving methods degrade.
  • The RoadSight dataset itself, with LiDAR-anchored depth, multi-modality views, and anonymized identities, gives a reproducible testbed for infrastructure-to-vehicle view transformation.
  • Because the method works from fixed cameras with LiDAR calibration, existing highway and intersection monitoring infrastructure becomes a potential source of training data for downstream perception models.

Reading between the lines

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

  • If the pseudo views are geometrically consistent, the cascade strategy should scale to even larger view shifts by chaining more pseudo views, potentially synthesizing full 360-degree vehicle loops from a single intersection; this is not tested in the paper.
  • A natural next test is to feed the synthesized vehicle views into an end-to-end driving planner and measure closed-loop or open-loop performance, since the paper evaluates only image-quality and detection-style metrics, not downstream driving quality.
  • The confidence weight W could double as an automatic data-quality filter: instead of thresholding by hand, one could drop low-confidence pseudo views during training or use them to actively query new infrastructure viewpoints, an extension the authors do not discuss.
  • The dependence on LiDAR for depth calibration may limit deployment to infrastructure already equipped with LiDAR; cameras-only depth calibration would be the natural ablation to test.
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 / 7 minor

Summary. The paper proposes I2V-GS, a 3D Gaussian Splatting framework for synthesizing vehicle-perspective views from infrastructure-mounted cameras and LiDAR. The pipeline uses LiDAR-calibrated monocular depth, an adaptive depth-warping scheme to generate dense pseudo views, a cascade latent-diffusion inpainting strategy to fill disoccluded regions, and a confidence-guided optimization that weights pseudo-view supervision by the agreement between the rendered and inpainted images. The authors also introduce RoadSight, a multi-modal, multi-view infrastructure dataset. Experiments on RoadSight and Waymo report large improvements over DNGaussian, SparseGS, StreetGaussian, and S3Gaussian in NTA-IoU, NTL-IoU, and FID.

Significance. If the quantitative claims survive scrutiny, the paper addresses a genuinely useful scenario: producing driving data from cheaper, already-deployed infrastructure sensors rather than dedicated vehicle fleets. The LiDAR-anchored depth calibration in Sec. 4.2 is well motivated, the adaptive warp bound in Eq. 9 is a conservative and explicit simplification, and the cascade inpainting strategy is a reasonable way to densify sparse fixed viewpoints. The RoadSight dataset also targets an under-served setting. However, the central performance claim is not yet established because the RoadSight evaluation lacks a defined vehicle-view ground truth, the FID reference distribution is ambiguous, and the self-supervised pseudo-view loop has no mechanism for detecting geometrically wrong but self-consistent content. The contribution is promising but requires substantially stronger validation.

major comments (5)
  1. [Sec. 5.2 / Table 2] The RoadSight evaluation protocol is underspecified in a load-bearing way. RoadSight is described in Sec. 5.1 as collected only from infrastructure sensors; no vehicle-mounted camera is mentioned, so the source of ground-truth vehicle-view images is unclear. The metric description says NTA-IoU and NTL-IoU compare detections with ground truth LiDAR 'after projection,' but the paper does not state which LiDAR is used, which camera intrinsics and extrinsics define the vehicle view, how LiDAR returns are projected into an image that was never captured, or how occluded regions without LiDAR returns are handled. Without this information, the headline numbers in Table 2 cannot be reproduced or meaningfully interpreted.
  2. [Sec. 5.2 / Table 2] The FID metric as described is not a valid novel-view fidelity measure on RoadSight. The text states that FID assesses the feature-distribution distance between 'the synthesized novel view and the original view.' If the original views are the infrastructure images, then FID measures similarity to a different camera viewpoint rather than correctness of the synthesized vehicle view; a lower FID could simply mean the synthetic vehicle image is more distributionally similar to infrastructure imagery. The paper must specify the reference distribution and, ideally, evaluate against paired vehicle-view images.
  3. [Sec. 4.3 / Sec. 4.4 / Fig. 6] The confidence-guided optimization does not, by itself, ensure that pseudo views are reliable. The confidence weight in Eq. 13 is computed from L2 and SSIM between the Gaussian render and the inpainted image, and the loss in Eq. 14 is applied to the same pair. If the render and the inpainted image agree on plausible but geometrically incorrect content, the confidence weight is high and the error is reinforced rather than corrected. The red-to-green traffic light change shown in Fig. 6 is exactly such a case: the authors acknowledge the diffusion model can alter semantic content, but the proposed weighting can only downweight disagreement, not detect agreement on wrong content. The claim that cross-view consensus yields 'reliable inpaint content' needs a direct consistency evaluation between pseudo views, not only end-task metrics.
  4. [Sec. 4.1 / Eq. 7] The adaptive depth-warp derivation appears to assume no relative rotation between source and target views. Eq. 5 includes the rotation term R'R^{-1}, but Eq. 7 contains only translation deltas and the source depth z; no rotation terms appear. For infrastructure-to-vehicle transformation, the rotation change is large, and the simplification leading to Eq. 9 is not justified in the text. The authors should either derive the bound with rotation included or state explicitly that R'=R is assumed, and evaluate whether that assumption holds in the reported scenarios.
  5. [Sec. 5.2 / Sec. 5.3 / Table 3] No statistical reliability is reported. The RoadSight results are single numbers over four scenes and the Waymo results over ten sequences, yet the pipeline includes stochastic diffusion inpainting and randomized training; run-to-run variance is expected. Without multiple seeds or confidence intervals, the claimed improvements (e.g., 45.7% NTA-IoU over StreetGaussian) cannot be distinguished from noise. This concern is reinforced by Table 3, where the differences between ablated variants are small (0.786 vs. 0.804 NTA-IoU).
minor comments (7)
  1. [Throughout] The metric is written inconsistently as both 'NTA-Iou' and 'NTA-IoU'; please standardize to 'IoU' throughout, including in the abstract, Table 2, and Table 3.
  2. [Sec. 5.3] There is a typo in the baseline name: 'StreerGaussian' should be 'StreetGaussian.'
  3. [Sec. 4.1 / Eq. 7] Eq. 7 appears with garbled line-break characters in the manuscript and the variables Delta t_x, Delta t_y, Delta t_z are not explicitly defined in the surrounding text; please clean up the equation and define all symbols.
  4. [Sec. 5.1] The dataset section reports 50 scenes of 20 seconds each but the experiments use only 4 representative scenes; please clarify the selection criterion and whether the remaining scenes are used at all.
  5. [Sec. 5.1] No download link, license, or data-access statement is provided for the introduced RoadSight dataset; as a claimed contribution, the dataset should be made available or an access procedure should be described.
  6. [Sec. 5.2] The latent diffusion model is referenced only as 'a latent diffusion model [19]'; for reproducibility, the specific inpainting checkpoint, prompt, and guidance settings should be stated.
  7. [Sec. 5.2] The text says 'we adopt Depth-Anything [33]' but the reference list contains both 'Depth Anything' [32] and 'Depth Anything v2' [33]; please clarify which variant and which pretrained weights were used.

Circularity Check

2 steps flagged · score 6.0 of 10

Vehicle-view training target is the model's own warped/inpainted render; confidence weighting is a self-consistency check, making the headline synthesis-quality claim partially circular despite LiDAR and input-view anchors.

  1. self definitional [Sec. 4.3, Algorithm 1 (pseudo-view generation and training loop)]
    "Algorithm 1: Iwarp, Dwarp, Mwarp ←ψ( ˆI ′, ˆD′ | V′ j−1,V ′ j); I ′, D′, M′ ← D(Iwarp, Dwarp, Mwarp); V ′ j(gt)←I ′, D′, M′ ... Compute loss L ′ ← Lcon( ˆI ′ t, I′ gt)."

    The pseudo-view ground truth is produced by warping the Gaussian's own current render Î′, D̂′ to the next pseudo view and then inpainting holes with a diffusion model: I′ = D(ψ(Î′, D̂′)). This I′ is then stored as V′j(gt) and used in the confidence-guided loss Lcon against the Gaussian render Î′t. Therefore the vehicle-view supervisory signal is a function of the model's own output; the target is defined in terms of the prediction. No vehicle-mounted camera or independent vehicle-view image is used in this loop, so the optimization encourages the Gaussian to reproduce its own warped/inpainted appearance. The claimed novel-vehicle-view quality is thus partially a self-consistency property rather than agreement with an externally observed vehicle view.

  2. self definitional [Sec. 4.4, Eqs. 13-14 (confidence-guided optimization)]
    "Given the inpainted image I ′ and rendered image ˆI ′ under V ′ j, the combined confidence weight is: W=λ 1 ·(1−L 2( ˆI ′, I′)) + (1−λ1)·SSIM( ˆI ′, I′). ... Lcon =E[ W·| ˆI ′, I′| 1 ]."

    The confidence weight that is supposed to assess the reliability of the diffusion inpaint is computed by comparing the inpainted image I′ with the Gaussian's own render Î′ at the same pseudo view. Both quantities come from the same pipeline: I′ is the warped and inpainted version of Î′, so W is high exactly when the render and the inpainted warp agree, regardless of whether that shared content is geometrically or semantically correct (the ablation in Fig. 6 shows a red light changed to green when cascade is removed). The optimization Lcon = E[W·|Î′−I′|1] then mainly enforces agreement between two outputs of the same model, so the 'cross-view information' reduces to a self-consistency check and cannot serve as an external validation of the inpainted content.

full rationale

The derivation chain for the claimed vehicle-view synthesis quality is partially self-referential. The pseudo views that provide dense training supervision are constructed by taking the current Gaussian render at one pseudo view, forward-warping it to the next pseudo view, and inpainting the resulting holes with a pretrained diffusion model (Sec. 4.2-4.3, Alg. 1). This inpainted warp is then labeled 'gt' and used in Lcon to optimize the same Gaussian. Consequently, the target for the vehicle view is a function of the model's own output; the model is trained to reproduce a target that is derived from itself. The confidence weight in Eq. 13 is also an internal consistency check between the Gaussian render and the diffusion-inpainted warp, so it cannot validate the inpainted content against any external vehicle-view observation. This is a genuine but partial circularity: the pipeline is not closed, because real input infrastructure views are used in L_ori during warm-up and LiDAR anchors the monocular depth calibration (Sec. 4.2), and the diffusion model is a fixed external generative prior. The headline NTA/NTL and FID numbers, however, rest on an underspecified evaluation protocol (projected-LiDAR labels with no stated vehicle-view camera, and FID against the 'original view' rather than a matched vehicle frame); that is an evaluation-validity concern rather than a circularity reduction, so it does not by itself raise the circularity score. There is no load-bearing self-citation or imported uniqueness theorem. Overall, the central vehicle-view 'prediction' is trained against a target generated by the model's own rendering, which warrants a 6 rather than 0-2, but the external LiDAR/input-view anchors and pretrained diffusion prior keep it from being fully circular.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The framework leans on standard 3DGS and camera geometry, and on two pre-trained models (monocular depth and diffusion inpainting) whose reliability in this novel setting is not demonstrated. The fitted depth calibration and the user-set warp bound and loss weights are the main free parameters.

free parameters (4)
  • c1, c2, c3 (depth calibration coefficients) = fitted per scene, not reported
    In Eq. 11-12, these parameters map monocular disparity to absolute depth using LiDAR supervision; they are fitted to each scene's LiDAR point cloud.
  • epsilon (warp difference bound) = not reported
    In Eq. 8-9, this user-set parameter controls the maximum pixel displacement in the adaptive depth warp and is chosen by hand.
  • loss weights lambda and lambda_1 = not reported
    lambda weighs RGB vs SSIM in Eq. 4 and lambda_1 weighs L2 vs SSIM in Eq. 13; both are manually set.
  • training schedule hyperparameters = warmup 3000; cascade every 3000; 3 cycles
    These control when pseudo views are generated and how many cycles are used; chosen by hand in Sec. 5.2.
assumptions (6)
  • standard math 3D Gaussian Splatting rendering equations
    Used as background in Sec. 3 without proof.
  • standard math Pinhole camera projection and rigid body motion
    Used in Eqs. 5-6 for warping.
  • domain assumption Monocular depth estimates from Depth-Anything are approximately correct up to scale and offset, and can be calibrated with LiDAR
    Sec. 4.2 relies on this to provide geometry guidance.
  • domain assumption The latent diffusion inpainting model produces plausible and reasonably view-consistent content for large holes in driving scenes
    Sec. 4.3 relies on this for filling warped images.
  • domain assumption LiDAR and cameras are hardware-synchronized and extrinsically calibrated
    Sec. 5.1 states this, and the depth calibration depends on it.
  • ad hoc to paper Setting Δt_z=0 and z=z_min to derive the warp bounds is a conservative and valid simplification
    Sec. 4.1 uses this to derive Eq. 9; it may limit the warp range but avoids distortion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of I2V-GS: Infrastructure-to-Vehicle View Transformation with Gaussian Splatting for Autonomous Driving Data Generation." pith.science (2026). https://pith.science/paper/ANTFGH2P

@misc{pith2026250723683,
  author       = {Pith},
  title        = {Pith review of: I2V-GS: Infrastructure-to-Vehicle View Transformation with Gaussian Splatting for Autonomous Driving Data Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ANTFGH2P}},
  note         = {Machine review of arXiv:2507.23683}
}
read the original abstract

Vast and high-quality data are essential for end-to-end autonomous driving systems. However, current driving data is mainly collected by vehicles, which is expensive and inefficient. A potential solution lies in synthesizing data from real-world images. Recent advancements in 3D reconstruction demonstrate photorealistic novel view synthesis, highlighting the potential of generating driving data from images captured on the road. This paper introduces a novel method, I2V-GS, to transfer the Infrastructure view To the Vehicle view with Gaussian Splatting. Reconstruction from sparse infrastructure viewpoints and rendering under large view transformations is a challenging problem. We adopt the adaptive depth warp to generate dense training views. To further expand the range of views, we employ a cascade strategy to inpaint warped images, which also ensures inpainting content is consistent across views. To further ensure the reliability of the diffusion model, we utilize the cross-view information to perform a confidenceguided optimization. Moreover, we introduce RoadSight, a multi-modality, multi-view dataset from real scenarios in infrastructure views. To our knowledge, I2V-GS is the first framework to generate autonomous driving datasets with infrastructure-vehicle view transformation. Experimental results demonstrate that I2V-GS significantly improves synthesis quality under vehicle view, outperforming StreetGaussian in NTA-Iou, NTL-Iou, and FID by 45.7%, 34.2%, and 14.9%, respectively.

Figures

Figures reproduced from arXiv: 2507.23683 by the authors.

Figure 1
Figure 1. We propose transferring images captured by infrastructure to vehicle views for autonomous driving system training to reduce the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of I2V-GS. Initially, the input views warm up the Gaussian optimization. Then, given the rendered [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The layout of sensors. All sensors are attached to the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Quality comparison of vehicle view rendering with DNGaussian [ [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Quality comparison of novel trajectory rendering with DNGaussian [ [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Ablation study on proposed methods. Original monocular depth is adopted in ‘w/o LiDAR mono depth’ and fixed depth warp is [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 22 canonical work pages

  1. [1]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621–11631, 2020. 1

    work page 2020

  2. [2]

    End-to-end autonomous driving: Challenges and frontiers.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 2024

    Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, An- dreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 2024. 1

    2024

  3. [3]

    Magicdrive: Street view generation with diverse 3d geometry control.arXiv preprint arXiv:2310.02601, 2023

    Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. Magicdrive: Street view generation with diverse 3d geometry control.arXiv preprint arXiv:2310.02601, 2023. 1, 3

    arXiv 2023

  4. [4]

    Vista: A generalizable driving world model with high fidelity and versatile controllability.arXiv preprint arXiv:2405.17398, 2024

    Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability.arXiv preprint arXiv:2405.17398, 2024. 1, 3

    arXiv 2024

  5. [5]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in neural information processing systems, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in neural information processing systems, 30, 2017. 6

    2017

  6. [6]

    Gaia-1: A generative world model for au- tonomous driving.arXiv preprint arXiv:2309.17080, 2023

    Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gian- luca Corrado. Gaia-1: A generative world model for au- tonomous driving.arXiv preprint arXiv:2309.17080, 2023. 1, 3

    arXiv 2023

  7. [7]

    St-p3: End-to-end vision-based au- tonomous driving via spatial-temporal feature learning

    Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based au- tonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533–549. Springer, 2022. 1

    work page 2022

  8. [8]

    Planning-oriented autonomous driving

    Yi Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wen Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 17853–17862, 2022. 1

    work page 2022

Show all 38 references
  1. [9]

    S3gaussian: Self-supervised street gaussians for autonomous driving.arXiv preprint arXiv:2405.20323, 2024

    Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. S3gaussian: Self-supervised street gaussians for autonomous driving.arXiv preprint arXiv:2405.20323, 2024. 1, 2, 7, 8

    2024 arXiv

  2. [10]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023. 2, 3, 4, 5, 6

    2023

  3. [11]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.CoRR, abs/1412.6980, 2014. 6

    2014 arXiv

  4. [12]

    Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normaliza- tion

    Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normaliza- tion. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775–20785, 2024. ...

    2024

  5. [13]

    Learning distilled collaboration graph for multi-agent perception.Advances in Neural Infor- mation Processing Systems, 34:29541–29552, 2021

    Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception.Advances in Neural Infor- mation Processing Systems, 34:29541–29552, 2021. 1

    2021

  6. [14]

    A survey on autonomous driving datasets: Statistics, annotation quality, and a future outlook.IEEE Transactions on Intelligent Vehicles, 2024

    Mingyu Liu, Ekim Yurtsever, Jonathan Fossaert, Xingcheng Zhou, Walter Zimmer, Yuning Cui, Bare Luka Zagar, and Alois C Knoll. A survey on autonomous driving datasets: Statistics, annotation quality, and a future outlook.IEEE Transactions on Intelligent Vehicles, 2024. 1

    2024

  7. [15]

    Deceptive-nerf/3dgs: Diffusion- generated pseudo-observations for high-quality sparse-view reconstruction

    Xinhang Liu, Jiaben Chen, Shiu-Hong Kao, Yu-Wing Tai, and Chi-Keung Tang. Deceptive-nerf/3dgs: Diffusion- generated pseudo-observations for high-quality sparse-view reconstruction. InEuropean Conference on Computer Vi- sion, pages 337–355. Springer, 2024. 2

    2024

  8. [16]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 2

    2021

  9. [17]

    Recondreamer: Crafting world models for driving scene reconstruction via online restora- tion.arXiv preprint arXiv:2411.19548, 2024

    Chaojun Ni, Guosheng Zhao, Xiaofeng Wang, Zheng Zhu, Wenkang Qin, Guan Huang, Chen Liu, Yuyin Chen, Yida Wang, Xueyang Zhang, et al. Recondreamer: Crafting world models for driving scene reconstruction via online restora- tion.arXiv preprint arXiv:2411.19548, 2024. 3

    2024 arXiv

  10. [18]

    Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022. 2

    2022 arXiv

  11. [19]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In2022 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 5

    2022

  12. [20]

    Germ ´an Ros, Laura Sellart, Joanna Materzynska, David V´azquez, and Antonio M. L ´opez. The synthia dataset: A large collection of synthetic images for semantic segmenta- tion of urban scenes.2016 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 3234–3243,

    2016

  13. [21]

    Scalability in per- ception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in per- ception for autonomous driving: Waymo open dataset. In 2024 IEEE/CVF Conference on Computer Vision and Pat...

    2024

  14. [22]

    Street- view image generation from a bird’s-eye view layout.IEEE Robotics and Automation Letters, 9(4):3578–3585, 2024

    Alexander Swerdlow, Runsheng Xu, and Bolei Zhou. Street- view image generation from a bird’s-eye view layout.IEEE Robotics and Automation Letters, 9(4):3578–3585, 2024. 1, 3

    2024

  15. [23]

    Srinivasan, Jonathan T

    Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Prad- han, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Bar- ron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pag...

    2022

  16. [24]

    Zhang, Francesco Ferroni, and Deva Ramanan

    Haithem Turki, Jason Y . Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 12375–12385, 2023. 2

    2023

  17. [25]

    Drivedreamer: Towards real-world- drive world models for autonomous driving

    Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jia- gang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world- drive world models for autonomous driving. InEuropean Conference on Computer Vision, pages 55–72. Springer,

  18. [26]

    Driving into the future: Multiview visual forecasting and planning with world model for au- tonomous driving

    Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for au- tonomous driving. In2024 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 14749–14759,

  19. [27]

    Privacy of autonomous vehicles: Risks, protection methods, and future directions

    Chulin Xie, Zhong Cao, Yunhui Long, Diange Yang, Ding Zhao, and Bo Li. Privacy of autonomous vehicles: Risks, protection methods, and future directions. 2022. 1

    2022

  20. [28]

    Sparsegs: Real- time 360° sparse view synthesis using gaussian splatting

    Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real- time 360° sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023. 1, 2, 7, 8

    2023 arXiv

  21. [29]

    Street gaussians: Modeling dynamic urban scenes with gaussian splatting

    Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. InEuropean Conference on Computer Vision, pages 156–173, 2024. 1, 2, 6, 7, 8

    2024

  22. [30]

    Streetcrafter: Street view synthesis with controllable video diffusion models.arXiv preprint arXiv:2412.13188, 2024

    Yunzhi Yan, Zhen Xu, Haotong Lin, Haian Jin, Haoyu Guo, Yida Wang, Kun Zhan, Xianpeng Lang, Hujun Bao, Xi- aowei Zhou, et al. Streetcrafter: Street view synthesis with controllable video diffusion models.arXiv preprint arXiv:2412.13188, 2024. 3

    2024 arXiv

  23. [31]

    Generalized predictive model for autonomous driving

    Jiazhi Yang, Shenyuan Gao, Yihang Qiu, Li Chen, Tianyu Li, Bo Dai, Kashyap Chitta, Peng Wu, Jia Zeng, Ping Luo, Jun Zhang, Andreas Geiger, Yu Qiao, and Hongyang Li. Generalized predictive model for autonomous driving. In 2024 IEEE/CVF Conference on Computer Vision and Pat- ter...

    2024

  24. [32]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024. 2

    2024

  25. [33]

    Depth any- thing v2.Advances in Neural Information Processing Sys- tems, 37:21875–21911, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2.Advances in Neural Information Processing Sys- tems, 37:21875–21911, 2024. 6

    2024

  26. [34]

    Driving scene synthesis on free-form trajecto- ries with generative prior.arXiv preprint arXiv:2412.01717,

    Zeyu Yang, Zijie Pan, Yuankun Yang, Xiatian Zhu, and Li Zhang. Driving scene synthesis on free-form trajecto- ries with generative prior.arXiv preprint arXiv:2412.01717,

    arXiv

  27. [35]

    Dair-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection

    Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection. In 2022 IEEE/CVF Conference on Computer Vision and Pat-...

    2022

  28. [36]

    V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecast- ing

    Haibao Yu, Wen-Yen Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xuming Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, Juan Song, Jirui Yuan, Ping Luo, and Za- iqing Nie. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and foreca...

    2023

  29. [37]

    Drivedreamer4d: World models are effective data machines for 4d driving scene rep- resentation.arXiv preprint arXiv:2410.13571, 2024

    Guosheng Zhao, Chaojun Ni, Xiaofeng Wang, Zheng Zhu, Xueyang Zhang, Yida Wang, Guan Huang, Xinze Chen, Boyuan Wang, Youyi Zhang, et al. Drivedreamer4d: World models are effective data machines for 4d driving scene rep- resentation.arXiv preprint arXiv:2410.13571, 2024. 1, 3, 6

    arXiv

  30. [38]

    Drivinggaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes

    Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634– 21643, 2...

    2023

Pith tools

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