Pith. sign in

REVIEW 3 major objections 5 minor 44 references

RoGS reconstructs large-scale road surface maps with comparable or better RGB, semantic, and elevation quality than mesh-based methods while running roughly 53x faster in a single training epoch.

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 · deepseek-v4-flash

2026-08-02 00:18 UTC pith:6T7RHNRY

load-bearing objection Solid efficiency win with a self-built benchmark; quality gains not yet established. the 3 major comments →

arxiv 2607.15048 v1 pith:6T7RHNRY submitted 2026-07-16 cs.CV

RoGS: Adaptive Meshgrid Gaussian for Large-Scale Road Surface Mapping

Tianchen Deng , Zhiheng Feng , Wenhua Wu , Ziming Li , Siting Zhu , Hesheng Wang This is my paper
classification cs.CV
keywords road surface reconstruction3D Gaussian splattingmeshgrid Gaussianadaptive meshingsemantic mappingelevation mapautonomous drivinglarge-scale mapping
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 paper argues that representing a road surface as 2D Gaussian surfels arranged on an adaptive meshgrid captures the thin, locally flat structure of roads better than dense 3D Gaussians or fixed meshes. It shows that this representation, together with a road-structure-aware refinement that concentrates surfels at lane markings, boundaries, and height changes, produces RGB, semantic, and elevation maps on KITTI and nuScenes with quality at least comparable to prior mesh-based reconstruction, while cutting optimization time from hours to minutes. The paper also introduces a pose-robust initialization that averages elevation priors from multiple nearby vehicle poses and down-weights inconsistent ones, reducing sensitivity to local trajectory noise. If these claims hold, road surface maps for HD mapping and annotation could be generated at a fraction of the previous computational cost.

Core claim

On its own terms, the paper's central discovery is that a road surface can be parameterized as a meshgrid of 2D Gaussian surfels that each carry color, opacity, semantics, scale, and rotation, and that this representation can be optimized through differentiable rendering to jointly reconstruct RGB, semantic, and elevation maps. The authors show that the meshgrid layout reduces primitive overlap and redundancy relative to both volumetric 3D Gaussians and fixed meshes, while the adaptive refinement places additional surfels where structure is complex. They report that, with one epoch of training, this achieves a 53x speedup over a mesh-based baseline with comparable PSNR and better mIoU and el

What carries the argument

The central object is the adaptive meshgrid Gaussian surfel: a 2D Gaussian flattened along the road normal, placed at vertices of a base grid generated from the vehicle trajectory, with density adaptively increased in cells that show high RGB gradient, semantic boundary, or elevation variation. The meshgrid layout acts as a structured prior that limits overlap and redundant primitives, and the pose-consensus initialization uses a weighted average of elevation predictions from neighboring vehicle poses with a confidence weight derived from their variance, which then moderates the height regularization loss during optimization.

Load-bearing premise

The evaluation relies on pseudo ground-truth BEV maps built from LiDAR point clouds transformed by the same vehicle poses and labeled with the same semantic pseudo-labels that supervise the network, so the benchmark may not be fully independent of the method being tested.

What would settle it

Take a surveyed road surface with high-accuracy control points (e.g., total station or RTK-GPS surveyed lane markings and curbs) and compare RoGS's reconstructed elevation and semantic boundaries against that surveyed truth, rather than against accumulated LiDAR; if the elevation RMSE and mIoU gains over the mesh baseline shrink or reverse, the reported quality gains are artifacts of the self-built reference.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If RoGS's speedup generalizes, road surface map generation for entire city-scale sequences could drop from hours to minutes, enabling near-real-time HD map updates from vehicle fleets.
  • Because the same primitives store color, semantics, and geometry, downstream tasks (BEV perception, lane-level annotation, road topology) could consume one unified representation instead of separate RGB/semantic/elevation layers.
  • The pose-robust initialization makes the method usable with noisy consumer-grade poses, potentially removing the need for expensive post-processed LiDAR-inertial odometry for road mapping.
  • When LiDAR is available, elevation supervision further improves mIoU and elevation error, suggesting that sensor fusion remains valuable even with a strong geometric prior.

Where Pith is reading between the lines

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

  • If the evaluation pseudo-ground truth is partially circular (built from the same poses and semantic pseudo-labels), the reported mIoU and elevation gains over the mesh baseline may partly reflect how closely the representation fits the reference construction, not absolute reconstruction accuracy; an independent benchmark with surveyed ground truth would be needed to confirm.
  • The adaptive meshgrid idea could transfer to other planar surface reconstruction tasks such as building facades, floors, or runways, where a dominant local flatness assumption holds.
  • One could test the pose-robustness claim directly by injecting synthetic pose noise into the input and measuring whether RoGS's elevation error degrades more gracefully than a nearest-pose baseline; the paper's variance-based confidence suggests a monotone relationship.
  • The claim that 2D surfels reduce overlap could be quantified by counting effective overlapping primitives per ray or measuring gradient sparsity, which would isolate the representation's benefit from the adaptive allocation.

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 RoGS, a road-surface-mapping framework that represents the road as 2D Gaussian surfels placed on an adaptive meshgrid. Each surfel stores color, semantics, and geometry, and is optimized through differentiable rendering with RGB, semantic, smoothness, pose-consistency, and optional LiDAR elevation losses. The authors report that RoGS reconstructs high-quality RGB, semantic, and elevation maps on nuScenes and KITTI, with speedups of 53x (one epoch) and 27x (two epochs) over the mesh-based RoMe baseline.

Significance. If the reported performance is reliable, the efficiency gain is practically important for large-scale HD-map generation: reducing optimization from hours to minutes is a substantial step. The adaptive meshgrid idea is well motivated, and the use of 2D Gaussian surfels is a sensible match to the thin-surface geometry of roads. The speedup claim is based on concrete runtime measurements and is therefore testable. The manuscript also includes ablations for each proposed component and states that code will be open-sourced. However, the quality claims currently rest on an evaluation reference constructed from the same poses and semantic pseudo-labels used in training, so the mIoU and elevation improvements are not yet independently established.

major comments (3)
  1. [Section IV-B (Evaluation Metrics)] The pseudo-ground-truth BEV maps are built by transforming each LiDAR frame with the same vehicle poses that initialize RoGS and that enter the trajectory-consistency regularization L_traj (Eq. 18). Semantic pseudo-labels come from Mask2Former and are used both as training supervision (Eq. 16) and as evaluation reference. The elevation reference is the same pose-transformed LiDAR used for optional supervision (Eq. 19). Thus low elevation RMSE and high mIoU partly measure how faithfully RoGS reproduces the priors it was given, not independent reconstruction accuracy. Please provide validation against an independent reference (e.g., surveyed/RTK ground truth, manual annotation, or at least a pose-perturbation test showing that results degrade gracefully and that the reported gains over RoMe are not artifacts of shared pseudo-labels).
  2. [Table I and Section IV-C] The quantitative evaluation covers only five nuScenes scenes, with a single run per scene and no variance or repeated trials. The only baseline is RoMe; no comparison to EMIE-MAP or a 3DGS-based variant is reported. KITTI is mentioned as an experiment dataset, but no quantitative KITTI results appear in Table I or elsewhere. This is insufficient to support the general claim of 'high-quality RGB, semantic, and elevation map reconstruction.' At minimum, report mean and standard deviation over multiple runs, add at least one more strong baseline, and provide quantitative KITTI results, especially for the speedup claim on the 3724 m sequence mentioned in Fig. 6.
  3. [Table I and Fig. 6] The speedup ratio depends critically on the units of the 'Time' column and on matched experimental conditions. Table I reports RoMe 1630 and RoGS 30.7/61.1 but does not state the units, the scene duration, or the definition of one epoch. Fig. 6 states RoMe takes 2.5 hours and RoGS 4 minutes for KITTI sequence 00, a ratio of ~37x, which differs from the abstract's 53x/27x. Please clarify the units, the hardware/software setup, the exact iteration counts, and whether RoMe was run with the same number of epochs/iterations and the same evaluation code. Without this, the reader cannot verify that the speedup is not due to looser convergence or different implementation details.
minor comments (5)
  1. [Abstract and Title] The title says 'RoGS' but the abstract refers to 'ROADGS-T'. Use one consistent name throughout.
  2. [Fig. 6 caption] The caption uses 'RoGs' instead of 'RoGS'.
  3. [Section IV-C first sentence] 'Fig. I reports the quantitative comparison' should be 'Table I'.
  4. [Eq. (4)] The weights omega_g, omega_s, omega_z and threshold tau_a are introduced without a sensitivity analysis. Since the adaptive refinement is a core contribution, a brief ablation varying these values would strengthen the paper.
  5. [Section V-E (Limitations)] The limitation that reconstruction quality depends on pose accuracy is acknowledged, but the suggested alternative of using COLMAP point clouds for LiDAR-free vehicles is not experimentally validated. Please add at least one experiment with COLMAP-derived point clouds or explicitly label this as future work.

Circularity Check

2 steps flagged

Quality metrics are partially self-referential: the BEV pseudo-ground-truth is built from the same pose-transformed LiDAR and Mask2Former labels that supervise RoGS; the speedup claim is independently measurable.

specific steps
  1. fitted input called prediction [Section IV-B 'Evaluation Metrics' and Eq. (16)]
    "Since BEV ground-truth road maps are not directly provided by nuScenes or KITTI, we construct pseudo ground truth using concatenated LiDAR point clouds. Specifically, each LiDAR frame is transformed into the global coordinate system using the vehicle pose. Then, each point is projected onto the nearest camera images, where its color and semantic label are determined by the corresponding projected image pixel."

    These projected semantic labels come from the same Mask2Former predictions that define the training target \bar{Sem}_i in the semantic loss (Eq. 16). The reported mIoU therefore measures how well RoGS reproduces its own training label source, not an independent semantic ground truth. Because RoMe is not given the same LiDAR/multi-pose priors, the comparison is asymmetric; at minimum, the semantic-quality claim is self-referential rather than externally verified.

  2. fitted input called prediction [Section IV-B 'Evaluation Metrics'; Eqs. (8)-(12), (18)-(19)]
    "For geometric evaluation, we use the elevation of the accumulated LiDAR point cloud as the reference. For each reconstructed Gaussian surfel center, we query the nearest LiDAR point within a radius of 0.1 m on the xy-plane and compute the elevation error. ... Lz = Σ_i |z_i − \bar{z}_i|_2^2, where \bar{z}_i is obtained by querying the nearest point-cloud elevation in the xy-plane."

    The RMSE reference is accumulated LiDAR transformed by the same vehicle poses that initialize the surfel elevations (Eqs. 8-10), define the trajectory-consistency target \hat{z}_i in L_traj (Eq. 18), and, when LiDAR is enabled, supervise z_i in L_z (Eq. 19). Low elevation RMSE is therefore partly a measure of staying near the method's own input priors, not of independently verified road geometry. The paper's Limitations admit road height from color/semantics is hard and depends on pose accuracy, reinforcing that the evaluation does not break this dependence.

full rationale

This is a systems paper whose method is an explicit representation plus loss design, not a theorem derivation. The core speedup (53x/27x) is an externally measurable runtime property, and the representation choices (2D Gaussian surfels on an adaptive meshgrid, trajectory-consistency weighting) follow from stated geometric priors rather than from a self-citation chain. No load-bearing uniqueness theorem is imported from the authors' prior work; the many self-citations in Related Works are contextual. The only substantive circularity is in the evaluation of quality: Section IV-B builds the BEV pseudo-GT from LiDAR transformed by the same vehicle poses that initialize and regularize the surface, and colors/labels it from the same camera images and Mask2Former outputs used in the training losses (Eqs. 15-19). Hence mIoU and elevation RMSE partly measure agreement with the method's own supervisory inputs, not independent ground truth; the paper's own Limitations concede pose dependence and difficulty of height recovery. This raises the score above 2 but does not invalidate the efficiency claim, so score 4 rather than 6+.

Axiom & Free-Parameter Ledger

7 free parameters · 7 axioms · 0 invented entities

The paper contributes a representation-and-loss system, not a first-principles derivation. The main ad hoc elements are the adaptive meshgrid subdivision score (Eq. 4), the confidence weighting (Eq. 12), and several un-reported thresholds; none are derived from sensor physics. No new physical entities are introduced.

free parameters (7)
  • base BEV resolution = 0.05 m/pixel
    Chosen by a resolution sweep on nuScenes scene-0655 (Fig. 5); sets surfel density and directly affects reconstruction quality and cost.
  • structure-complexity weights (omega_g, omega_s, omega_z)
    Hand-set weights in Eq. (4) that determine adaptive refinement; no values or sensitivity analysis are reported.
  • adaptive refinement threshold tau_a
    Indicator threshold in Eq. (5) controlling which grid cells are subdivided; value not reported.
  • trajectory-consistency parameters (K, sigma_d, tau_z)
    K (number of neighboring poses), sigma_d (spatial weight scale, Eq. 9), and tau_z (confidence sensitivity, Eq. 12) are user-set but no values or robustness study are provided.
  • maximum refinement level = 1
    Set to one for efficiency; caps detail in complex regions.
  • optimizer learning rates = 1e-4 (opacity/scale/rotation); 1.6e-4 to 1.6e-6 (elevation); 0.008 (color); 0.1 (semantic); 0.001 (exposure)
    Hand-set optimization schedule, including a scene-size-proportional elevation rate; not derived.
  • elevation evaluation matching radius = 0.1 m
    In Section IV-B, elevation RMSE only considers the nearest LiDAR point within 0.1 m in the xy-plane; the metric depends on this choice.
axioms (7)
  • standard math 3D Gaussian splatting's differentiable rendering and covariance parameterization (Eqs. 1-2, 13) are correct.
    Paper builds on [12] without re-deriving; accepted background.
  • domain assumption Setting the normal-direction scale s_z = 0 yields a valid 2D Gaussian surfel that matches road geometry.
    Section III-A; assumes roads are perfectly thin surfaces and 2D Gaussian projection models appearance faithfully.
  • domain assumption The vehicle pose is locally parallel to the road surface, so Eq. (8) predicts surfel elevation from a neighboring pose using a planar road model.
    Section III-B; load-bearing for height initialization and trajectory-consistency prior; can fail on slopes and bumps.
  • domain assumption The pseudo ground-truth BEV maps from LiDAR accumulation (Section IV-B) are accurate references for RGB, semantic, and elevation evaluation.
    Evaluation is only as good as this reference; it uses the same poses and semantic pseudo-labels as training.
  • domain assumption Mask2Former segmentation provides reliable semantic pseudo-labels for supervision and evaluation.
    Used in losses (Eq. 16) and in pseudo-GT construction (Section IV-A/B); errors propagate to both training and evaluation.
  • ad hoc to paper The structure-complexity score kappa_m in Eq. (4) is a sufficient proxy for road complexity, with hand-set weights.
    No evidence that RGB gradient + semantic boundary + height variation optimally predicts where surfels should be added.
  • ad hoc to paper Trajectory-consistency confidence q_i = exp(-sigma^2_z,i / tau^2_z) in Eq. (12) is a reliable measure of pose-prior reliability.
    Assumes disagreement among neighboring pose predictions is Gaussian-like and calibrates confidence; no validation against actual pose error.

pith-pipeline@v1.3.0-alltime-deepseek · 14027 in / 16918 out tokens · 169524 ms · 2026-08-02T00:18:25.317454+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of RoGS: Adaptive Meshgrid Gaussian for Large-Scale Road Surface Mapping." pith.science (2026). https://pith.science/paper/6T7RHNRY

@misc{pith2026260715048,
  author       = {Pith},
  title        = {Pith review of: RoGS: Adaptive Meshgrid Gaussian for Large-Scale Road Surface Mapping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6T7RHNRY}},
  note         = {Machine review of arXiv:2607.15048}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Road surface mapping plays a crucial role in autonomous driving, supporting high-definition map generation, lane-level perception, and automatic road annotation. Recent mesh-based road surface reconstruction methods have shown promising results, but they still suffer from limited reconstruction quality and high optimization cost, especially in large-scale driving scenarios. To address these limitations, we propose ROADGS-T, a robust and efficient large-scale road surface mapping framework based on adaptive meshgrid Gaussian representation. Specifically, we model the road surface by placing 2D Gaussian surfels on a meshgrid, where each surfel explicitly stores color, semantic, and geometric information. Compared with conventional mesh-based representations and 3D Gaussian primitives, the proposed meshgrid Gaussian representation better matches the thin-surface property of roads while significantly reducing redundant primitives and overlap during optimization. To further improve representation efficiency and structural fidelity, we introduce a road-structure-aware adaptive meshgrid strategy, which allocates denser Gaussian surfels to geometrically or semantically complex regions, such as lane markings, road boundaries, and height discontinuities, while maintaining a compact representation in flat road areas. Moreover, instead of relying on a single nearest vehicle pose, we design a trajectory-consistency-guided pose-robust refinement strategy, which estimates local surface priors from multiple neighboring poses and adaptively weights pose-guided height regularization according to their geometric consistency.

Figures

Figures reproduced from arXiv: 2607.15048 by Hesheng Wang, Siting Zhu, Tianchen Deng, Wenhua Wu, Zhiheng Feng, Ziming Li.

Figure 1
Figure 1. Figure 1: We propose RoGS, a robust and efficient large-scale road surface mapping framework based on adaptive meshgrid [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of RoGS. The left side shows the road representation with meshgrid Gaussian. The blue curve indicates the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of the two layouts. Layout-1 covers the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pose-based initialization. For each Gaussian surfel, its [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results at different resolutions. 0.05m/pixel resolution [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Road surface reconstruction results (KITTI odometry sequence-00) using our proposed RoGs, covering a trajectory [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: RGB results of road surface reconstruction for Scene-0655. Because RoMe is in a different reference to ours, the ground [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Semantic and elevation results of road surface reconstruction for Scene-0655. For the elevation error map, brighter [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of road reconstruction results for night and [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Road surface reconstructions of Scene-1 and Scene-2 [PITH_FULL_IMAGE:figures/full_fig_p010_10.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

44 extracted references · 13 linked inside Pith

  1. [1]

    Emie-map: Large-scale road surface reconstruction based on explicit mesh and implicit encoding.arXiv preprint arXiv:2403.11789, 2024

    Wenhua Wu, Qi Wang, Guangming Wang, Junping Wang, Tiankun Zhao, Yang Liu, Dongchao Gao, Zhe Liu, and Hesheng Wang. Emie-map: Large-scale road surface reconstruction based on explicit mesh and implicit encoding.arXiv preprint arXiv:2403.11789, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    Rome: Towards large scale road surface reconstruction via mesh representation.IEEE Transactions on Intelligent V ehicles, 2024

    Ruohong Mei, Wei Sui, Jiaxin Zhang, Xue Qin, Gang Wang, Tao Peng, Tao Chen, and Cong Yang. Rome: Towards large scale road surface reconstruction via mesh representation.IEEE Transactions on Intelligent V ehicles, 2024

    2024

  3. [3]

    Roadbev: Road surface reconstruction in bird’s eye view.arXiv preprint arXiv:2404.06605, 2024

    Tong Zhao, Lei Yang, Yichen Xie, Mingyu Ding, Masayoshi Tomizuka, and Yintao Wei. Roadbev: Road surface reconstruction in bird’s eye view.arXiv preprint arXiv:2404.06605, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Rsrd: A road surface reconstruction dataset and benchmark for safe and comfortable autonomous driving.arXiv preprint arXiv:2310.02262, 2023

    Tong Zhao, Chenfeng Xu, Mingyu Ding, Masayoshi Tomizuka, Wei Zhan, and Yintao Wei. Rsrd: A road surface reconstruction dataset and benchmark for safe and comfortable autonomous driving.arXiv preprint arXiv:2310.02262, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    A survey of structure from motion*.Acta Numerica, 26:305–364, 2017

    Onur Özye¸ sil, Vladislav V oroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*.Acta Numerica, 26:305–364, 2017

    2017

  6. [6]

    Dp-mvs: Detail preserving multi-view surface reconstruction of large-scale scenes.Remote Sensing, 13(22):4569, 2021

    Liyang Zhou, Zhuang Zhang, Hanqing Jiang, Han Sun, Hujun Bao, and Guofeng Zhang. Dp-mvs: Detail preserving multi-view surface reconstruction of large-scale scenes.Remote Sensing, 13(22):4569, 2021

    2021

  7. [7]

    Block-nerf: Scalable large scene neural view synthesis

    Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kret- zschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022

    2022

  8. [8]

    Mega- nerf: Scalable construction of large-scale nerfs for virtual fly-throughs

    Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega- nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022

    2022

  9. [9]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022

    2022

  10. [10]

    Sc-nerf: Self-correcting neural radiance field with sparse views

    Liang Song, Guangming Wang, Jiuming Liu, Zhenyang Fu, Yanzi Miao, et al. Sc-nerf: Self-correcting neural radiance field with sparse views. arXiv preprint arXiv:2309.05028, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Sni-slam: Semantic neural implicit slam.arXiv preprint arXiv:2311.11016, 2023

    Siting Zhu, Guangming Wang, Hermann Blum, Jiuming Liu, Liang Song, Marc Pollefeys, and Hesheng Wang. Sni-slam: Semantic neural implicit slam.arXiv preprint arXiv:2311.11016, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1–14, 2023

    2023

  13. [13]

    A survey on 3d gaussian splatting

    Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Structure-from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4104–4113, 2016

    2016

  15. [15]

    Photo tourism: exploring photo collections in 3d

    Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. InACM siggraph 2006 papers, pages 835–846. 2006

    2006

  16. [16]

    Scene reconstruction and visualization from internet photo collections: A survey.IPSJ Transactions on Computer Vision and Applications, 3:44–66, 2011

    Noah Snavely. Scene reconstruction and visualization from internet photo collections: A survey.IPSJ Transactions on Computer Vision and Applications, 3:44–66, 2011

    2011

  17. [17]

    Towards linear-time incremental structure from motion

    Changchang Wu. Towards linear-time incremental structure from motion. In2013 International Conference on 3D Vision-3DV 2013, pages 127–

    2013

  18. [18]

    Adaptive structure from motion with a contrario model estimation

    Pierre Moulon, Pascal Monasse, and Renaud Marlet. Adaptive structure from motion with a contrario model estimation. InComputer Vision– ACCV 2012: 11th Asian Conference on Computer Vision, Daejeon, Korea, November 5-9, 2012, Revised Selected Papers, Part IV 11, pages 257–270. Springer, 2013

    2012

  19. [19]

    Regformer: an efficient projection-aware transformer network for large-scale point cloud registration

    Jiuming Liu, Guangming Wang, Zhe Liu, Chaokang Jiang, Marc Pollefeys, and Hesheng Wang. Regformer: an efficient projection-aware transformer network for large-scale point cloud registration. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8451– 8460, 2023

    2023

  20. [20]

    Pixelwise view selection for unstructured multi-view stereo

    Johannes L Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501–518. Springer, 2016

    2016

  21. [21]

    Building rome in a day

    Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54(10):105–112, 2011

    2011

  22. [22]

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

    2021

  23. [23]

    Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021

    2021

  24. [24]

    Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam

    Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Co-slam: Joint coordinate and sparse parametric encodings for neural real-time slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13293–13302, 2023

    2023

  25. [25]

    Dvn- slam: Dynamic visual neural slam based on local-global encoding

    Wenhua Wu, Guangming Wang, Ting Deng, Sebastian Ægidiu, Stuart Shanks, Valerio Modugno, Dimitrios Kanoulas, and Hesheng Wang. Dvn- slam: Dynamic visual neural slam based on local-global encoding. In 2025 IEEE International Conference on Robotics and Automation (ICRA), pages 14564–14571. IEEE, 2025

    2025

  26. [26]

    Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering

    Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. InEuropean conference on computer vision, pages 106–122. Springer, 2022

    2022

  27. [27]

    Prosgnerf: Progressive dynamic neural scene graph with frequency modulated auto-encoder in urban scenes.arXiv preprint arXiv:2312.09076, 2023

    Tianchen Deng, Siyang Liu, Xuan Wang, Yejia Liu, Danwei Wang, and Weidong Chen. Prosgnerf: Progressive dynamic neural scene graph with frequency modulated auto-encoder in urban scenes.arXiv preprint arXiv:2312.09076, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    Suds: Scalable urban dynamic scenes

    Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375– 12385, 2023

    2023

  29. [29]

    S-nerf: Neural radiance fields for street views.arXiv preprint arXiv:2303.00749, 2023

    Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views.arXiv preprint arXiv:2303.00749, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    What is the best 3d scene representation for robotics? from geometric to foundation models.arXiv preprint arXiv:2512.03422, 2025

    Tianchen Deng, Yue Pan, Shenghai Yuan, Dong Li, Chen Wang, Mingrui Li, Long Chen, Lihua Xie, Danwei Wang, Jingchuan Wang, Javier Civera, Hesheng Wang, and Weidong Chen. What is the best 3d scene representation for robotics? from geometric to foundation models.arXiv preprint arXiv:2512.03422, 2025

    arXiv 2025

  31. [31]

    Semgauss-slam: Dense semantic gaussian splatting slam.arXiv preprint arXiv:2403.07494, 2024

    Siting Zhu, Renjie Qin, Guangming Wang, Jiuming Liu, and Hesheng Wang. Semgauss-slam: Dense semantic gaussian splatting slam.arXiv preprint arXiv:2403.07494, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    Compact 3d gaussian splatting for dense visual slam.arXiv preprint arXiv:2403.11247, 2024

    Tianchen Deng, Yaohui Chen, Leyan Zhang, Jianfei Yang, Shenghai Yuan, Danwei Wang, and Weidong Chen. Compact 3d gaussian splatting for dense visual slam.arXiv preprint arXiv:2403.11247, 2024

    Pith/arXiv arXiv 2024

  33. [33]

    Vpgs-slam: IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 12 V oxel-based progressive 3d gaussian slam in large-scale scenes.arXiv preprint arXiv:2505.18992, 2025

    Tianchen Deng, Wenhua Wu, Junjie He, Yue Pan, Xirui Jiang, Shenghai Yuan, Danwei Wang, Hesheng Wang, and Weidong Chen. Vpgs-slam: IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 12 V oxel-based progressive 3d gaussian slam in large-scale scenes.arXiv preprint arXiv:2505.18992, 2025

    arXiv 2025

  34. [34]

    Neurogauss4d-pci: 4d neural fields and gaussian deformation fields for point cloud interpolation.arXiv preprint arXiv:2405.14241, 2024

    Chaokang Jiang, Dalong Du, Jiuming Liu, Siting Zhu, Zhenqiang Liu, Zhuang Ma, Zhujin Liang, and Jie Zhou. Neurogauss4d-pci: 4d neural fields and gaussian deformation fields for point cloud interpolation.arXiv preprint arXiv:2405.14241, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Road surface 3d reconstruction based on dense subpixel disparity map estimation.IEEE Transactions on Image Processing, 27(6):3025–3035, 2018

    Rui Fan, Xiao Ai, and Naim Dahnoun. Road surface 3d reconstruction based on dense subpixel disparity map estimation.IEEE Transactions on Image Processing, 27(6):3025–3035, 2018

    2018

  36. [36]

    3d reconstruction of road surfaces using an integrated multi-sensory approach.Optics and lasers in engineering, 45(7):808–818, 2007

    Si-Jie Yu, Sreenivas R Sukumar, Andreas F Koschan, David L Page, and Mongi A Abidi. 3d reconstruction of road surfaces using an integrated multi-sensory approach.Optics and lasers in engineering, 45(7):808–818, 2007

    2007

  37. [37]

    Road surface reconstruction by stereo vision.PFG–Journal of Photogrammetry, Remote Sensing and Geoinformation Science, 88(6):433–448, 2020

    Hauke Brunken and Clemens Gühmann. Road surface reconstruction by stereo vision.PFG–Journal of Photogrammetry, Remote Sensing and Geoinformation Science, 88(6):433–448, 2020

    2020

  38. [38]

    Automatic reconstruction of road surface features by using terrestrial mobile lidar.Automation in Construction, 58:165–175, 2015

    Jenny Guo, Meng-Ju Tsai, and Jen-Yu Han. Automatic reconstruction of road surface features by using terrestrial mobile lidar.Automation in Construction, 58:165–175, 2015

    2015

  39. [39]

    Rui Fan, Umar Ozgunalp, Yuan Wang, Ming Liu, and Ioannis Pitas. Rethinking road surface 3-d reconstruction and pothole detection: From perspective transformation to disparity map segmentation.IEEE Transactions on Cybernetics, 52(7):5799–5808, 2021

    2021

  40. [40]

    High-quality surface reconstruction using gaussian surfels

    Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. arXiv preprint arXiv:2404.17774, 2024

    Pith/arXiv arXiv 2024

  41. [41]

    2d gaussian splatting for geometrically accurate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. arXiv preprint arXiv:2403.17888, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Masked-attention mask transformer for universal image segmentation

    Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022

    2022

  43. [43]

    nuscenes: A multimodal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020

    2020

  44. [44]

    Vision meets robotics: The kitti dataset.The International Journal of Robotics Research, 32(11):1231–1237, 2013

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset.The International Journal of Robotics Research, 32(11):1231–1237, 2013

    2013