Pith. sign in

REVIEW 3 major objections 5 minor 16 references

Road surfaces can be rebuilt as compact Gaussians in one feed-forward pass, without per-scene training.

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 ↗

T0 review · grok-4.5

2026-07-30 13:24 UTC pith:CQVDHNJQ

load-bearing objection Solid systems paper for feed-forward road Gaussians; the AnySplat comparison is partly confounded by inputs, but the RoGS win and road-plane fusion still make it worth engaging. the 3 major comments →

arxiv 2607.23758 v1 pith:CQVDHNJQ submitted 2026-07-26 cs.CV

RoadVGGT: Road-Structure-Aware Feed-Forward Road Surface Reconstruction

Han Jiao , Chen Liu , Jiakai Sun , Zhanjie Zhang , Mengyuan Yang , Yimeng Li , Mofan Zhou , Kun Zhan
show 1 more author
Lei Zhao
This is my paper
classification cs.CV
keywords road surface reconstructionfeed-forward Gaussian splattinggeometric foundation modelsbird's-eye-view mappinggrid fusionautonomous drivingelevation estimation
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.

Road mapping usually means optimizing a new model for every stretch of road, with coverage and hyperparameters tied to that drive. This paper argues that is unnecessary: a geometric foundation model, given multi-view images plus poses and depth, can drive a learned Gaussian head that predicts dense road attributes, which are then aligned into metric world coordinates and fused on a road-plane grid. Structure-aware grouping keeps thin markings and road–sidewalk boundaries from being averaged away. The result is one compact Gaussian road surface that yields RGB and semantic bird’s-eye maps, elevation, and novel views—without test-time scene optimization—and that beats both a generic feed-forward Gaussian baseline and a strong per-scene road optimizer on quality and storage, including zero-shot on a second driving dataset.

Core claim

RoadVGGT shows that dense pixel-aligned Gaussian predictions from a geometric foundation model can be turned into a compact, renderable road surface by metric road-plane grid fusion with category-aware and road–sidewalk junction grouping, eliminating per-scene optimization while improving image quality, semantic mapping, and elevation accuracy relative to prior feed-forward and optimization-based road methods.

What carries the argument

Road-structure-aware XY grid fusion: confidence-weighted merging of dense Gaussians on a metric road-aligned plane, with category-aware groups and road–sidewalk junction protection so only structure-compatible Gaussians fuse inside each cell.

Load-bearing premise

The frozen geometric backbone, given poses and depth, must already be accurate enough that a learned Gaussian head and fixed-grid fusion can recover thin markings and elevation without any per-scene refinement.

What would settle it

On held-out long road sequences, if feed-forward RoadVGGT (no test-time training) fails to match or beat a per-scene road Gaussian method on shared-region PSNR, BEV mIoU, and elevation error while staying more compact—or collapses when backbone depth/pose noise rises—the central claim does not hold.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • New road logs can be mapped by direct inference instead of training a fresh scene model each time.
  • HD-map style products—RGB/semantic BEV, elevation, and novel views—can share one compact static Gaussian surface.
  • Gaussian capacity can scale with physical road extent via a metric plane grid rather than a fixed token budget or generic 3D voxels.
  • Cross-dataset zero-shot use becomes a practical target when poses and depth are available at capture time.
  • Gains in driving-scene geometric foundation models transfer immediately into better feed-forward road maps.

Where Pith is reading between the lines

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

  • If backbone geometry is the bottleneck, joint or lightly adapted geometry training on road manifolds may unlock larger quality jumps than fusion tweaks alone.
  • The same plane-grid plus structure-group idea could extend to other near-manifold outdoor surfaces (parking lots, runways) where vertical extent is limited.
  • Chunk-then-global-fuse inference suggests a natural streaming mapper for continuous fleet collection, if drift across chunks is controlled.

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 / 5 minor

Summary. The paper proposes RoadVGGT, a feed-forward pipeline for large-scale road surface reconstruction. It attaches a DPT-style Gaussian head to a frozen OmniVGGT backbone, which consumes multi-view RGB together with provided camera poses, LiDAR depth observations, and Mask2Former semantic maps. Predicted pixel-aligned 2D Gaussians are aligned into a metric world frame via Umeyama alignment, filtered to road-surface classes, and compressed by confidence-weighted fusion on a 0.05 m road-plane XY grid, with category-aware grouping and road–sidewalk junction protection to avoid destructive averaging of thin markings and adjacent surfaces. On Waymo (trained) and nuScenes (zero-shot), RoadVGGT reports better PSNR/SSIM/LPIPS, BEV mIoU, and Z-RMSE than AnySplat (feed-forward) and RoGS (per-scene optimization), at 112.81 MB storage and 0.84 s rendering (Table 1), with ablations for grid resolution, Gaussian parameterization, and grouping (Table 2). The central claims are (i) a feed-forward method can replace per-scene optimization for road mapping, and (ii) road-structure-aware fusion preserves quality while reducing storage.

Significance. If the comparison protocol is cleaned up, this is a useful and timely contribution: it is among the first works to convert geometric-foundation-model features into a compact, renderable road-surface map without per-scene optimization, and it shows zero-shot transfer from Waymo to nuScenes. Concrete strengths worth crediting: consistent gains over the per-scene-optimized RoGS (which plausibly received comparable privileged inputs) on all five quality metrics plus storage and runtime; a principled, physically meaningful metric XY-grid fusion whose cell size is comparable across trajectories; structure-aware grouping with a measurable elevation benefit; and a reasonably complete ablation suite (grid resolution at four scales, 2D vs 3D Gaussian parameterization, grouping components). The method is practically relevant for scalable HD-map updates. The significance is currently capped by the unverified fairness of the AnySplat comparison, which the abstract leans on.

major comments (3)
  1. [§4.1, Table 1] §4.1/Table 1: the paper never states what inputs the baselines received. RoadVGGT's input is X={I,K,T,D,S} (§3.1): GT poses, LiDAR depth, and Mask2Former maps. AnySplat is a pose-free method 'from unconstrained views'; if it received no poses/depth, the comparison is confounded. Its Waymo Z-RMSE of 4.8173 m vs 0.23/0.26 for the posed methods is the signature of missing metric scale (median Z-offset alignment, §4.1, removes offset but not scale) — yet on nuScenes AnySplat reaches 0.5595, which is inconsistent with a purely up-to-scale output and is never explained. The manuscript must (a) state exactly what each baseline consumed, (b) describe how AnySplat outputs were scaled/aligned before Z-RMSE, and (c) ideally include a matched-input control (e.g., RoadVGGT without depth input, or AnySplat with GT-pose alignment) so the gain attributable to architecture vs. input conditioning is quant
  2. [§3.3, §4.1 (mIoU protocol)] RoadVGGT's semantic labels are not predicted: they are Mask2Former labels inherited per-pixel and majority-voted per grid cell (§3.1, §3.3). The Waymo mIoU gain (0.4692 vs 0.4361/0.2768) therefore largely measures faithful propagation of an external segmentation model. Two things must be clarified: (1) the provenance of the mIoU ground truth — if the reference BEV semantics derive from the same fine-tuned Mask2Former, RoadVGGT's score partially measures self-agreement with the label source; (2) whether AnySplat and RoGS were given the identical semantic maps. If the baselines used different (or no) label sources, the semantic-mapping column of Table 1 compares input pipelines, not reconstruction methods, and this caveat is load-bearing for the abstract's claim of improved 'semantic mapping'.
  3. [§4.1 (evaluation protocol), Table 1 (Storage)] §4.1 states metrics are computed on the intersection of the reconstruction masks of the compared methods. This is fair per-pixel but hides coverage differences that the paper itself acknowledges (§4.3: RoadVGGT retains more visible surface than RoGS's preset lateral range). Two consequences: (i) a method reconstructing only easy central regions could score better on the intersection while being worse as a map; (ii) the storage comparison (112.81 MB vs AnySplat 712.47 / RoGS 178.67) compares representations of different spatial extents, since RoadVGGT's semantic filtering discards everything off-road while AnySplat models the full scene. Please report per-method reconstructed area/coverage, and either normalize storage by covered area or add a matched-coverage comparison. Per-scene variance (error bars or per-scene tables in the supplement) would also strengthen Table 1.
minor comments (5)
  1. [Table 2(a)] Table 2(a): the r=0.05 m default is described as the best balance, but it is not dominant on all metrics — r=0.02 m gives better mIoU (0.4850 vs 0.4692) and r=0.10 m gives better Z-RMSE (0.2040 vs 0.2312). The choice is defensible, but the text should acknowledge the trade-off explicitly rather than implying 0.05 m is uniformly best.
  2. [§3.2 (Umeyama alignment)] Since GT camera poses are provided to OmniVGGT, it is unclear why a full similarity (Umeyama) alignment of predicted vs. input trajectories is needed — the backbone already conditions on the input poses. Please clarify what residual misalignment this step corrects, and whether scale s deviates from 1 in practice.
  3. [§3.5 (chunked inference)] Chunks are non-overlapping and processed independently, with a single global fusion afterward. Are seams or discontinuities visible at chunk boundaries in BEV or elevation maps? A brief comment or figure would help, since boundary artifacts are a natural failure mode of chunked inference.
  4. [Fig. 1, author block, References] Figure 1 contains the typo 'Camera Paramters'; the author/affiliation block has a stray comma line (', cszhl@zju.edu.cn'); references are inconsistently formatted (mix of venue and arXiv citations, 'RoGs' vs 'RoGS'). Please also state whether code, the fine-tuned Mask2Former, and evaluation scripts will be released.
  5. [Table 2(c), §3.5] The junction-protection ablation (Table 2(c)) shows its main effect on Z-RMSE (0.2312 vs 0.2638) with negligible PSNR change; stating this explicitly would help readers understand which component buys which improvement. Similarly, the LPIPS weight λ=0.05 appears only in §4.1; defining it in §3.5 where the loss is introduced would improve readability.

Circularity Check

0 steps flagged

No circular derivation: empirical feed-forward pipeline evaluated on external benchmarks; gains are experimental, not forced by construction.

full rationale

RoadVGGT’s chain is architectural and empirical, not a claimed first-principles derivation. The backbone (OmniVGGT) is frozen; only the Gaussian head is trained with a standard L1+LPIPS rendering loss on camera-view images (§3.5). Gaussian centers come from back-projected backbone depth after Umeyama alignment to provided poses (§3.2); fusion is confidence-weighted XY-grid averaging with optional category/junction grouping (§3.3–3.4)—none of these steps algebraically force the reported PSNR/SSIM/LPIPS, Z-RMSE, or storage numbers. Semantics are explicitly external Mask2Former inputs propagated by majority vote, not predicted attributes (§3.1, §3.3); calling the result “semantic mapping” is a product claim, not a circular prediction of those labels. Evaluation uses held-out Waymo and zero-shot nuScenes against AnySplat and RoGS (Table 1). No load-bearing self-citation uniqueness theorem, no fitted constant renamed as a prediction, and no equation that reduces a target metric to its own fit. Input asymmetry vs. AnySplat (poses/depth/masks) is a fairness/correctness concern, not circularity. Score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central claim rests on standard multi-view geometry, a frozen pretrained backbone, supplied poses/depth/semantics, and several hand-chosen fusion and training hyperparameters. No new physical entities are postulated; the invented pieces are engineering modules whose only evidence is the reported benchmarks.

free parameters (5)
  • road-plane grid resolution r = 0.05 m
    Default 0.05 m chosen by ablation trade-off among quality, noise, and storage (Table 2a); directly controls compactness and detail retention.
  • LPIPS loss weight λ = 0.05
    Set to 0.05 after linear warmup; shapes the appearance objective used to train the Gaussian head.
  • AdamW learning rate and schedule = 4e-5, 8 epochs
    4e-5, weight decay 0.01, 100 warmup steps, cosine schedule, 8 epochs; standard but claim-mediating training knobs.
  • inference chunk length = 6 (Waymo), 5 (nuScenes)
    6 Waymo / 5 nuScenes consecutive timestamps; affects memory and how global fusion sees the trajectory.
  • fine-structure category set L_fine
    Defines which semantic classes keep separate fusion groups; membership is a design choice that changes what survives grid fusion.
axioms (5)
  • domain assumption Provided camera poses and depth observations are accurate enough to serve as geometric conditioning and metric alignment targets.
    Invoked throughout §3.1–3.2; Umeyama alignment and back-projection depend on input trajectory and depth.
  • domain assumption Road surfaces are locally close to a ground-aligned 2D manifold, justifying flattened 2D Gaussians and XY-plane (not 3D voxel) fusion.
    Stated in §3.2–3.3 and ablated in Table 2b; load-bearing for the fusion design.
  • domain assumption External Mask2Former semantic maps are reliable enough for road filtering, majority-vote labels, and structure-aware grouping.
    §3.1 and §4.1 treat segmentation as system input, not a predicted Gaussian attribute.
  • ad hoc to paper Freezing OmniVGGT camera/depth/point heads and training only the Gaussian head transfers sufficient geometry for road mapping.
    §3.5 training protocol; Conclusion admits heavy dependence on backbone geometry quality.
  • ad hoc to paper Confidence-weighted averaging within structure-compatible XY groups preserves appearance and elevation better than uniform or purely volumetric fusion.
    Core of §3.3–3.4; supported by ablations but not derived from first principles.
invented entities (2)
  • Road-structure-aware XY grid fusion (category groups + road–sidewalk junction protection) no independent evidence
    purpose: Compress dense pixel Gaussians while protecting markings and curb boundaries.
    Main methodological construct; evidence is internal ablations and baselines, not an external measurable entity.
  • DPT-style Gaussian head predicting attributes plus fusion confidence on OmniVGGT features no independent evidence
    purpose: Turn multi-view tokens into dense renderable road Gaussians in one pass.
    Standard head-on-backbone pattern specialized to this pipeline; validated only by the paper’s renders and tables.

pith-pipeline@v1.2.0-grok45-kimik3 · 16317 in / 3601 out tokens · 54220 ms · 2026-07-30T13:24:44.029001+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of RoadVGGT: Road-Structure-Aware Feed-Forward Road Surface Reconstruction." pith.science (2026). https://pith.science/paper/CQVDHNJQ

@misc{pith2026260723758,
  author       = {Pith},
  title        = {Pith review of: RoadVGGT: Road-Structure-Aware Feed-Forward Road Surface Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CQVDHNJQ}},
  note         = {Machine review of arXiv:2607.23758}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large-scale road surface reconstruction supports high-definition mapping, autonomous-driving perception, annotation, and simulation. Existing road-specialized optimization methods can produce high-quality road representations, but they typically require per-scene training and scene-dependent coverage design around the driving trajectory, limiting scalable reconstruction over newly collected roads. To address these limitations, we introduce RoadVGGT, a road-structure-aware feed-forward framework that reconstructs compact Gaussian road surfaces without test-time per-scene optimization. RoadVGGT uses a geometric foundation model to exploit multi-view images together with provided pose and depth observations, and predicts dense pixel-aligned Gaussian attributes through a learned Gaussian head. To make these dense predictions usable for large road surfaces, we align them into a consistent metric world coordinate system and fuse redundant Gaussians on the road-aligned XY plane through confidence-weighted grid fusion. Category-aware grouping and road--sidewalk junction protection further control fusion around vulnerable road structures. The resulting representation supports RGB and semantic bird's-eye-view maps, elevation estimation, and novel view synthesis. RoadVGGT eliminates the need for per-scene optimization in prior methods, reconstructs complete road surfaces with a compact Gaussian representation, and improves image quality, semantic mapping, and elevation accuracy. Extensive experiments demonstrate the potential of geometric foundation models for scalable feed-forward road surface reconstruction.

Figures

Figures reproduced from arXiv: 2607.23758 by Chen Liu, Han Jiao, Jiakai Sun, Kun Zhan, Lei Zhao, Mengyuan Yang, Mofan Zhou, Yimeng Li, Zhanjie Zhang.

Figure 1
Figure 1. Figure 1: Overview of RoadVGGT. Given multi-view road images with provided camera poses, depth observations, and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of road-surface reconstruction on Waymo and zero-shot nuScenes. Columns show ground [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of dense prediction and grid fusion at different resolutions. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of 2D and 3D Gaussian [PITH_FULL_IMAGE:figures/full_fig_p007_4.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

16 extracted references · 11 linked inside Pith

  1. [6]

    arXiv:2409.02382

    GGS: Generalizable Gaussian Splatting for Lane Switching in Autonomous Driving. arXiv:2409.02382. Huang, N.; Wei, X.; Zheng, W.; An, P.; Lu, M.; Zhan, W.; Tomizuka,M.;Keutzer,K.;andZhang,S.2024.S 3Gaussian: Self-Supervised Street Gaussians for Autonomous Driving. arXiv:2405.20323. Itkin, R.; Issachar, N.; Keypur, Y.; Chen, X.; Chen, A.; and Benaim, S

    Pith/arXiv arXiv 2024

  2. [7]

    arXiv:2604.15284

    GlobalSplat: Efficient Feed- Forward 3D Gaussian Splatting via Global Scene Tokens. arXiv:2604.15284. Jiang,L.;Mao,Y.;Xu,L.;Lu,T.;Ren,K.;Jin,Y.;Xu,X.;Yu, M.; Pang, J.; Zhao, F.; et al

    Pith/arXiv arXiv

  3. [8]

    arXiv:2407.02598

    AutoSplat: Constrained Gaussian Splatting for Autonomous Driving Scene Reconstruction. arXiv:2407.02598. Li, Y.; Lv, C.; Tang, Z.; Yang, H.; and Huang, D

    Pith/arXiv arXiv

  4. [9]

    arXiv:2603.00697

    TokenSplat: Token-aligned 3D Gaussian Splatting for Feed- forward Pose-free Reconstruction. arXiv:2603.00697. Lin, H.; Chen, S.; Liew, J. H.; Chen, D. Y.; Li, Z.; Shi, G.; Feng, J.; and Kang, B

    arXiv

  5. [10]

    Mei, R.; Sui, W.; Zhang, J.; Qin, X.; Wang, G.; Peng, T.; Chen, T.; and Yang, C

    Depth Anything 3: Re- covering the visual space from any views.arXiv preprint arXiv:2511.10647. Mei, R.; Sui, W.; Zhang, J.; Qin, X.; Wang, G.; Peng, T.; Chen, T.; and Yang, C

    Pith/arXiv arXiv

  6. [11]

    arXiv:2003.08934

    NeRF: Represent- ing Scenes as Neural Radiance Fields for View Synthesis. arXiv:2003.08934. Peng,H.;Li,H.;Dai,Y.;Lan,Y.;Luo,Y.;Qi,T.;Zhang,Z.; Zhan, Y.; Zhang, J.; Xu, W.; and Liu, Z

    Pith/arXiv arXiv 2003

  7. [12]

    arXiv:2511.10560

    OmniVGGT: Omni-Modality Driven Visual Geometry Grounded Trans- former. arXiv:2511.10560. Smart, B.; Zheng, C.; Laina, I.; and Prisacariu, V. A

    arXiv

  8. [13]

    ADGaussian: Generalizable Gaussian Splatting for Au- tonomous Driving with Multi-modal Inputs.arXiv preprint arXiv:2504.00437. Sun, P.; Kretzschmar, H.; Dotiwalla, X.; Chouard, A.; Pat- naik, V.; Tsui, P.; Guo, J.; Zhou, Y.; Chai, Y.; Caine, B.; Vasudevan, V.; Han, W.; Ngiam, J.; Zhao, H.; Timofeev, A.; Ettinger, S.; Krivokon, M.; Gao, A.; Joshi, A.; Zhan...

    arXiv

  9. [14]

    InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Scalability in Perception for Autonomous Driving: Waymo Open Dataset. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Wang, J.; Chen, M.; Karaev, N.; Vedaldi, A.; Rupprecht, C.; and Novotny, D. 2025a. VGGT: Visual Geometry Grounded Transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Wa...

    Pith/arXiv arXiv

  10. [15]

    arXiv:2403.11789

    EMIE-MAP: Large- Scale Road Surface Reconstruction Based on Explicit Mesh and Implicit Encoding. arXiv:2403.11789. Xie, Z.; Zhang, J.; Li, W.; Zhang, F.; and Zhang, L

    Pith/arXiv arXiv

  11. [16]

    Yu,H.;Xiao,K.;Wang,J.;Hao,R.;Hu,G.;Huang,Y.;Qin, H.;Jing,B.;Bo,Y.;andLuo,P.2026.ReconDrive:FastFeed- Forward 4D Gaussian Splatting for Autonomous Driving Scene Reconstruction

    No Pose, No Problem: Sur- prisingly Simple 3D Gaussian Splats from Sparse Unposed Images.arXiv preprint arXiv:2410.24207. Yu,H.;Xiao,K.;Wang,J.;Hao,R.;Hu,G.;Huang,Y.;Qin, H.;Jing,B.;Bo,Y.;andLuo,P.2026.ReconDrive:FastFeed- Forward 4D Gaussian Splatting for Autonomous Driving Scene Reconstruction. Inhttps://arxiv.org/abs/2603.07552. Zhou, H.; Shao, J.; Xu,...

    Pith/arXiv arXiv 2026

  12. [2020]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    nuScenes: A Multimodal Dataset for Autonomous Driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Charatan,D.;Li,S.;Tagliasacchi,A.;andSitzmann,V.2024. pixelSplat:3DGaussianSplatsfromImagePairsforScalable Generalizable 3D Reconstruction. InCVPR. Chen, X.; Xiong, Z.; Chen, Y.; Li, G.; Wang, N.; Luo, H.; Chen...

    2024

  13. [2022]

    Feng,Z.;Wu,W.;Deng,T.;andWang,H.2024

    Masked-attention Mask Transformer for Uni- versal Image Segmentation. Feng,Z.;Wu,W.;Deng,T.;andWang,H.2024. RoGs:Large ScaleRoadSurfaceReconstructionwithMeshgridGaussian. arXiv:2405.14342. Han, H.; Zhou, K.; Long, X.; Wang, Y.; and Xiao, C

    Pith/arXiv arXiv 2024

  14. [2024]

    arXiv preprint arXiv:2403.14627

    MVSplat: Effi- cient3DGaussianSplattingfromSparseMulti-ViewImages. arXiv preprint arXiv:2403.14627. Cheng,B.;Misra,I.;Schwing,A.G.;Kirillov,A.;andGird- har, R

    Pith/arXiv arXiv

  15. [2025]

    arXiv:2512.03004

    DGGT: Feedforward 4D Recon- structionofDynamicDrivingScenesusingUnposedImages. arXiv:2512.03004. Chen, Y.; Xu, H.; Zheng, C.; Zhuang, B.; Pollefeys, M.; Geiger, A.; Cham, T.-J.; and Cai, J

    arXiv

  16. [2026]

    arXiv:2512.04021

    C3G: Learning Compact 3D Representations with 2K Gaussians. arXiv:2512.04021. Caesar, H.; Bankiti, V.; Lang, A. H.; Vora, S.; Liong, V. E.; Xu, Q.; Krishnan, A.; Pan, Y.; Baldan, G.; and Beijbom, O

    Pith/arXiv arXiv