Pith. sign in

REVIEW 4 major objections 4 minor 46 references

MagicRoad: Semantic-Aware 3D Road Surface Reconstruction via Obstacle Inpainting

T0 review · 4 major / 4 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 Erasing obstacles before rendering lets MagicRoad reconstruct cleaner, more accurate road surfaces from driving video than prior Gaussian splatting baselines.

desk verdict Plausible system, but the PSNR claim is circular (train and evaluate against the same inpainted image) and the geometry evidence is too thin to support the headline. read the letter →

arxiv 2507.23340 v1 pith:BQL5GGLC submitted 2025-07-31 cs.CV

Xingyue Peng , Yuandong Lyu , Lang Zhang , Jian Zhu , Songtao Wang , Jiaxin Deng , Songxin Lu , Weiliang Ma
show 3 more authors
Dangen She Peng Jia XianPeng Lang
This is my paper
classification cs.CV
keywords roadsurfacereconstruction3DGaussiansplatting2DsurfelsvideoinpaintingsemanticsegmentationBEVmappingautonomousdrivingnuScenes
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

MagicRoad claims that clean, geometrically faithful road surfaces can be reconstructed from ordinary multi-camera driving video even when the ground is heavily occluded, by erasing foreground objects before rendering instead of trying to model them. The paper's central bet is that a video inpainting step, guided by semantic segmentation, can remove parked cars, moving vehicles, and pedestrians, and that the resulting 'clean' frames can serve as training targets for a planar Gaussian surface model. On five nuScenes scenes it reports mean PSNR 24.10 versus 23.21 and elevation RMSE 0.050 m versus 0.061 m against the RoGS baseline, all without lidar. If this holds, it would make large-scale HD map and lane-mark production cheaper and more complete in real urban conditions.

What carries the argument

The load-bearing object is the planar-adapted 2D Gaussian surfel, a flattened Gaussian defined in a tangent plane by $G(u) = \exp(-(u^2+v^2)/2)$ with center $p_k$, scales $(s_u, s_v)$, tangents $(t_u, t_v)$, opacity $\alpha$, color $c$, and a semantic attribute $s$. Rendering uses front-to-back $\alpha$ blending of depth-sorted surfels, which gives planar road geometry with few vertical degrees of freedom. Around this, three mechanisms do the work: (1) segmentation-guided video inpainting removes dynamic and static occluders and supplies clean training targets; (2) semantic-aware HSV enhancement adjusts per-class value/saturation statistics across views, $\tilde{V}^v = (\sigma^r_c/\sigma^v_c)(V^v - \mu^v_c) + \mu^r_c$, to equalize appearance; (3) a composite loss $L = \lambda_c L_c + \lambda_d L_d + \lambda_n L_n + \lambda_s L_s$ ties photometric fidelity, depth smoothness, normal consistency, and semantic alignment into one optimization.

What would settle it

Mask out every image region that was occluded by a foreground object in the original camera views and recompute PSNR and elevation RMSE only on never-occluded road patches; if MagicRoad's advantage over RoGS disappears or reverses there, the claimed gains come from matching the inpainter rather than from better geometry. A complementary check is to project reconstructed elevation under parked cars and compare it against lidar scans of the same road captured when the cars were absent.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that occlusion removal and appearance harmonization belong upstream of surface reconstruction, not after it. MagicRoad first segments the video into semantic classes, masks vehicles and pedestrians, and inpaints the masked regions with a diffusion-based video inpainting model so that occluded road texture is plausibly restored in a temporally consistent way. Those inpainted images are then used as the photometric supervision target, and a semantic-aware color enhancement step adjusts value and saturation statistics per class in HSV space to remove lighting and exposure inconsistencies across views. The geometry itself is carried by planar-adapted 2D Gaussian surfels, each carrying a semantic descriptor, which are optimized with a composite loss enforcing photometric match to the inpainted target, depth smoothness, normal alignment to the road plane, and semantic alignment. The paper claims this combination produces BEV maps with RGB appearance, semantic labels, and elevation that are cleaner and more accurate than those of prior road-splatting baselines, including in night and rainy conditions.

Load-bearing premise

The evaluation treats the inpainted image as ground truth for road appearance, so the reported appearance gains may mostly show that the renderer imitates the inpainter rather than that the road surface is truly correct.

Editorial extensions

If this is right

  • Occluded road regions under parked vehicles can be plausibly filled rather than left as holes, so reconstructed BEV maps can show continuous lane markings and curbs.
  • Lidar is not needed for the geometry: the pipeline is fully camera-based, which could lower sensor cost for HD map generation.
  • Because appearance is harmonized across views before rendering, the output is better suited to automatic lane-marking annotation than direct splatting of raw video.
  • The claimed robustness in night and rainy conditions suggests the pipeline can generalize to degraded weather, not just clean daytime scenes.

Reading between the lines

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

  • The paper does not check whether the reported PSNR advantage survives on image regions that were never occluded; that test would separate rendering quality from agreement with the inpainter.
  • If the inpainter invents road texture, such as a false lane marking, the photometric loss would actively teach the Gaussians that fiction; a comparison against later clean observations of the same road would reveal this.
  • The color enhancement is currently a heuristic within the optimization loop; a learnable version with the same semantic priors is a natural next step and would make the module's contribution measurable on held-out scenes.
  • The paper acknowledges reliance on reasonably accurate camera poses; joint pose-and-surface optimization is the obvious extension the authors themselves flag.
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

4 major / 4 minor

Summary. MagicRoad proposes a road-surface reconstruction pipeline that combines planar 2D Gaussian surfels with semantic features, segmentation-guided video inpainting to remove occluders, and semantic-aware HSV color enhancement. The authors evaluate on five nuScenes scenes against the RoGS baseline, reporting PSNR and elevation RMSE, and claim significant improvements in both appearance fidelity and geometric accuracy.

Significance. If the quantitative claims were sound, this would be a practical contribution to BEV road reconstruction and HD mapping. The paper has several genuine strengths: the 2D Gaussian surfel representation with semantic descriptors is well motivated for planar road scenes; the use of video inpainting for temporal consistency is a sensible design choice; and the qualitative figures show plausible restoration of occluded road regions and lane markings. However, the main evidence for appearance superiority is circular, and the geometric evidence is too thinly reported to support the headline claim on its own.

major comments (4)
  1. [Experiments, Metrics / Eq. (7)] The photometric loss in Eq. (7) minimizes ||c(x) - ĉ(x)|| against the 'ground truth image (after inpainting and enhancement)', and the PSNR metric in the Metrics paragraph is computed against the 'ground-truth inpainted image'. This means the reported PSNR measures agreement with the authors' own inpainting output rather than with observed road appearance. For occluded regions, ĉ(x) is generated by DiffuEraser and is hallucinated, not measured, so a high PSNR does not imply reconstruction fidelity. The comparison is also structurally unfair to RoGS: RoGS optimizes against original frames and cannot be expected to reproduce hallucinated targets that MagicRoad is explicitly trained to match. The authors should evaluate appearance on original frames restricted to visible road pixels (excluding the occluder masks), or use independent manual/annotated judgments, and should report results separately for visible and occluded regions.
  2. [Experimental Setup / Table 1] The claim of 'significantly' better performance rests on five scenes with no error bars, no significance tests, and no specification of the elevation ground-truth source or the masks used for RMSE. The authors should expand the evaluation to substantially more scenes, provide standard deviations or confidence intervals, and clearly state whether LiDAR ground truth is used for evaluation and how inpainted regions are treated in the RMSE computation. As written, the mean 0.89 dB PSNR advantage cannot be separated from the circularity identified above, and the 0.011 m RMSE improvement is not established as statistically reliable.
  3. [Obstacle Inpainting] The method description says the segmentation network identifies 'dynamic foreground objects (e.g., vehicles, pedestrians)', but the abstract, introduction, and Figure 4 emphasize removal of static parked vehicles. If static objects are not included in the segmentation mask, the static-occluder claims are unsupported; if they are included, the text should say so explicitly and describe how parked vehicles are distinguished from legitimate road features. This discrepancy directly affects the scope of the claimed contribution.
  4. [Ablations and Analysis] The ablation section states that removing inpainting or color enhancement causes 'noticeable drops in both PSNR and elevation accuracy', but no ablation table or quantitative values are provided. Because the PSNR metric is the circular one described above, this statement cannot be verified. Please provide quantitative ablations with a valid, non-circular metric and per-component numbers for each removal.
minor comments (4)
  1. [Figure 1] Figure 1 contains an untranslated Chinese annotation ('可以再高一点') that should be removed or translated.
  2. [References] The citation formatting is inconsistent, including 'Pollefeys et al. 2004-09-01' in the introduction and 'RoGs' versus 'RoGS' across the text and Table 1.
  3. [Eq. (5)] The semantic-aware color enhancement does not specify how the reference-view statistics μᵣ_c and σᵣ_c are selected or averaged; please provide the exact definition.
  4. [Experimental Setup] The criteria for selecting the five evaluation scenes are not stated; please explain how scenes were chosen and whether they are representative of the full nuScenes dataset.

Circularity Check

1 steps flagged · score 6.0 of 10

The claimed PSNR advantage is circular: the metric's 'ground-truth' is the same inpainted image used as the photometric-loss training target (Eq. 7), so Table 1's appearance comparison is, by construction, a fit to the authors' own inpainting output.

  1. fitted input called prediction [Experiments, Metrics; Eq. (7); Obstacle Inpainting]
    "Peak Signal-to-Noise Ratio (PSNR) is used to measure the photometric consistency between the rendered view and the ground-truth inpainted image. ... Lc = Σ ∥c(x) − ˆc(x)∥1 (7) ... where c(x) is the predicted color from front-to-back Gaussian blending and ˆc(x) is the corresponding ground truth image (after inpainting and enhancement)."

    Eq. (7) trains the Gaussian renderer to reproduce ĉ(x), the inpainted/enhanced image, and the PSNR metric then scores the renderer against exactly this same ĉ(x) labeled as 'ground truth'. For occluded regions, ĉ(x) is hallucinated by the DiffuEraser video inpainting model, not measured by any sensor. Thus the reported PSNR advantage over RoGS, which is not trained on these inpainted targets, is a measure of how well the renderer matches the authors' own inpainting output, not of appearance fidelity to the true road surface. The appearance half of Table 1 therefore reduces to a self-consistency check rather than an independent comparison.

full rationale

The central appearance claim is circular in the specific sense required: the evaluation reference and the training target are the same inpainted image. Eq. (7) minimizes L1 distance between rendered color and the inpainted target, and the PSNR metric compares the rendered view to the 'ground-truth inpainted image'. The baseline RoGS optimizes against original frames, so it cannot be expected to reproduce hallucinated inpainted content, making the 0.89 dB mean PSNR gap an artifact of the evaluation protocol rather than evidence of superior reconstruction. The elevation RMSE comparison is independent in principle, since it uses ground-truth lidar-derived elevation, and the paper reports lower RMSE for MagicRoad on all five scenes; this provides non-circular evidence for the geometric part of the claim. However, the paper's headline claim of 'significantly outperforming' in both appearance and geometry rests substantially on the circular PSNR comparison, and the geometry result is reported on only five scenes without significance testing. No other circular steps were found: the planar Gaussian representation, semantic enhancement, and inpainting modules are presented as contributions and do not reduce to self-citations or renamed known results. Score 6 reflects partial circularity: the appearance prediction is forced by construction, while the geometry claim retains independent content.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method's success depends on assumptions that the inpainting masks are correct, the inpainting output is a faithful restoration of the road, the camera poses are accurate, and the road is planar. These are reasonable engineering priors, but the first two also contaminate the evaluation because the same inpainted images serve as both training target and PSNR reference.

free parameters (2)
  • Loss weights lambda_c, lambda_d, lambda_n, lambda_s
    These weights balance photometric, depth smoothness, normal consistency, and semantic alignment losses in Eq. 6. Their values are not reported, yet they control how strongly the renderer is pulled toward hallucinated inpainted content.
  • Reference view statistics mu_r_c, sigma_r_c
    Eq. 5 normalizes each view's HSV channels using either a chosen reference view or averaged statistics. The selection rule for the reference view is not given, and it materially affects the color harmonization output.
assumptions (4)
  • domain assumption Mask2Former segmentation masks are accurate for all dynamic and static foreground objects
    The inpainting stage relies on these masks to erase occluders; missed objects or over-segmentation corrupt the supervision targets. Referenced in the Obstacle Inpainting section.
  • ad hoc to paper DiffuEraser inpainted pixels are treated as ground truth road appearance
    Eq. 7 and the PSNR metric both use the inpainted image as the reference; under static occluders this content is generated, not measured.
  • domain assumption Camera poses and intrinsics from nuScenes are sufficiently accurate
    Projection of 3D points into views for HSV statistics and Gaussian optimization assumes correct calibration; the conclusion notes the method relies on reasonably accurate camera poses.
  • domain assumption Road surface is locally planar with a dominant normal
    Eq. 9 penalizes deviation from the dominant plane normal N, which is violated on slopes, curbs, and intersections.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MagicRoad: Semantic-Aware 3D Road Surface Reconstruction via Obstacle Inpainting." pith.science (2026). https://pith.science/paper/BQL5GGLC

@misc{pith2026250723340,
  author       = {Pith},
  title        = {Pith review of: MagicRoad: Semantic-Aware 3D Road Surface Reconstruction via Obstacle Inpainting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BQL5GGLC}},
  note         = {Machine review of arXiv:2507.23340}
}
read the original abstract

Road surface reconstruction is essential for autonomous driving, supporting centimeter-accurate lane perception and high-definition mapping in complex urban environments.While recent methods based on mesh rendering or 3D Gaussian splatting (3DGS) achieve promising results under clean and static conditions, they remain vulnerable to occlusions from dynamic agents, visual clutter from static obstacles, and appearance degradation caused by lighting and weather changes. We present a robust reconstruction framework that integrates occlusion-aware 2D Gaussian surfels with semantic-guided color enhancement to recover clean, consistent road surfaces. Our method leverages a planar-adapted Gaussian representation for efficient large-scale modeling, employs segmentation-guided video inpainting to remove both dynamic and static foreground objects, and enhances color coherence via semantic-aware correction in HSV space. Extensive experiments on urban-scale datasets demonstrate that our framework produces visually coherent and geometrically faithful reconstructions, significantly outperforming prior methods under real-world conditions.

Figures

Figures reproduced from arXiv: 2507.23340 by the authors.

Figure 1
Figure 1. We introduce MagicRoad, an innovative frame [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of MagicRoad. Given in-car camera videos, we perform temporal inpainting to remove dynamic objects, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Video inpainting leverages temporal attention [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison results of our methods. Our method effectively fills reconstruction holes caused by parked [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Semantic and elevation comparison. Our method recovers richer semantic details such as fine-grained lane markings. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Results under night and rainy conditions. Our method remains robust in challenging environments, effectively re [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Ablation study on inpainting and color enhance [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 28 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bertalmio, M.; Sapiro, G.; Caselles, V.; and Ballester, C. 2000. Image inpainting. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, 417--424

    work page 2000

  4. [4]

    H.; Vora, S.; Liong, V

    Caesar, H.; Bankiti, V.; Lang, A. H.; Vora, S.; Liong, V. E.; Xu, Q.; Krishnan, A.; Pan, Y.; Baldan, G.; and Beijbom, O. 2020. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11621--11631

    2020

  5. [5]

    Cao, C.; Dong, Q.; and Fu, Y. 2023. Zits++: Image inpainting by improving the incremental transformer on structural priors. IEEE transactions on pattern analysis and machine intelligence, 45(10): 12667--12684

    2023

  6. [6]

    Chang, Y.-L.; Yu Liu, Z.; and Hsu, W. 2019. Vornet: Spatio-temporally consistent video inpainting for object removal. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 0--0

    work page 2019

  7. [7]

    G.; Kirillov, A.; and Girdhar, R

    Cheng, B.; Misra, I.; Schwing, A. G.; Kirillov, A.; and Girdhar, R. 2022. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1290--1299

    2022

  8. [8]

    Fan, R.; Ai, X.; and Dahnoun, N. 2018. Road surface 3D reconstruction based on dense subpixel disparity map estimation. IEEE Transactions on Image Processing, 27(6): 3025--3035

    work page 2018

Show all 46 references
  1. [9]

    Feng, Z.; Wu, W.; and Wang, H. 2024. Rogs: Large scale road surface reconstruction based on 2d gaussian splatting. arXiv e-prints, arXiv--2405

    2024

  2. [10]

    Guo, Q.; Gao, S.; Zhang, X.; Yin, Y.; and Zhang, C. 2017. Patch-based image inpainting via two-stage low rank approximation. IEEE transactions on visualization and computer graphics, 24(6): 2023--2036

    2017

  3. [11]

    Hartley, R.; and Zisserman, A. 2003. Multiple view geometry in computer vision. Cambridge university press

    2003

  4. [12]

    Huang, B.; Yu, Z.; Chen, A.; Geiger, A.; and Gao, S. 2024. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 conference papers, 1--11

    2024

  5. [13]

    B.; Ahuja, N.; and Kopf, J

    Huang, J.-B.; Kang, S. B.; Ahuja, N.; and Kopf, J. 2014. Image completion using planar structure guidance. ACM Transactions on graphics (TOG), 33(4): 1--10

    2014

  6. [14]

    Ke, L.; Tai, Y.-W.; and Tang, C.-K. 2021. Occlusion-aware video object inpainting. In Proceedings of the IEEE/CVF international conference on computer vision, 14468--14478

    2021

  7. [15]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4): 139--1

    2023

  8. [16]

    Kim, D.; Woo, S.; Lee, J.-Y.; and Kweon, I. S. 2019. Deep video inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 5792--5801

    2019

  9. [17]

    Li, X.; Xue, H.; Ren, P.; and Bo, L. 2025. DiffuEraser: A Diffusion Model for Video Inpainting. arXiv preprint arXiv:2501.10018

    2025 arXiv

  10. [18]

    Lugmayr, A.; Danelljan, M.; Romero, A.; Yu, F.; Timofte, R.; and Van Gool, L. 2022. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11461--11471

    2022

  11. [19]

    Mei, R.; Sui, W.; Zhang, J.; Qin, X.; Wang, G.; Peng, T.; Chen, T.; and Yang, C. 2024. Rome: Towards large scale road surface reconstruction via mesh representation. IEEE Transactions on Intelligent Vehicles

    2024

  12. [20]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

    2021

  13. [21]

    G.; Kelly, J.; Brubaker, M

    Mirzaei, A.; Aumentado-Armstrong, T.; Derpanis, K. G.; Kelly, J.; Brubaker, M. A.; Gilitschenski, I.; and Levinshtein, A. 2023. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision...

    2023

  14. [22]

    Newson, A.; Almansa, A.; Fradet, M.; Gousseau, Y.; and P \'e rez, P. 2014. Video inpainting of complex scenes. Siam journal on imaging sciences, 7(4): 1993--2019

    2014

  15. [23]

    Pathak, D.; Krahenbuhl, P.; Donahue, J.; Darrell, T.; and Efros, A. A. 2016. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2536--2544

    2016

  16. [24]

    2004-09-01

    Pollefeys, M.; Gool, L.; Vergauwen, M.; Verbiest, F.; Cornelis, K.; Tops, J.; and Koch, R. 2004-09-01. Visual modeling with a hand-held camera

    2004

  17. [25]

    Ren, J.; Zheng, Q.; Zhao, Y.; Xu, X.; and Li, C. 2022. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3511--3520

    2022

  18. [26]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684--10695

    2022

  19. [27]

    Saharia, C.; Chan, W.; Chang, H.; Lee, C.; Ho, J.; Salimans, T.; Fleet, D.; and Norouzi, M. 2022. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, 1--10

    2022

  20. [28]

    L.; and Frahm, J.-M

    Schonberger, J. L.; and Frahm, J.-M. 2016. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, 4104--4113

    2016

  21. [29]

    L.; Zheng, E.; Frahm, J.-M.; and Pollefeys, M

    Sch \"o nberger, J. L.; Zheng, E.; Frahm, J.-M.; and Pollefeys, M. 2016. 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, 501--518. Springer

    2016

  22. [30]

    P.; Barron, J

    Tancik, M.; Casser, V.; Yan, X.; Pradhan, S.; Mildenhall, B.; Srinivasan, P. P.; Barron, J. T.; and Kretzschmar, H. 2022. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 8248--8258

    2022

  23. [31]

    Turki, H.; Ramanan, D.; and Satyanarayanan, M. 2022. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12922--12931

    2022

  24. [32]

    Wan, Z.; Zhang, J.; Chen, D.; and Liao, J. 2021. High-fidelity pluralistic image completion with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 4692--4701

    2021

  25. [33]

    Weber, E.; Holynski, A.; Jampani, V.; Saxena, S.; Snavely, N.; Kar, A.; and Kanazawa, A. 2024. Nerfiller: Completing scenes via generative 3d inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20731--20741

    2024

  26. [34]

    Wexler, Y.; Shechtman, E.; and Irani, M. 2007. Space-time completion of video. IEEE Transactions on pattern analysis and machine intelligence, 29(3): 463--476

    2007

  27. [35]

    Xia, D.; Zhang, W.; Liu, X.; Zhang, W.; Gong, C.; Huang, J.; Yang, M.; and Yang, D. 2024. DuMapNet: An End-to-End Vectorization System for City-Scale Lane-Level Map Generation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 6015--6024

    2024

  28. [36]

    Xia, D.; Zhang, W.; Liu, X.; Zhang, W.; Gong, C.; Tan, X.; Huang, J.; Yang, M.; and Yang, D. 2025. LDMapNet-U: An End-to-End System for City-Scale Lane-Level Map Updating. arXiv preprint arXiv:2501.02763

    2025 arXiv

  29. [37]

    Xu, R.; Li, X.; Zhou, B.; and Loy, C. C. 2019. Deep flow-guided video inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 3723--3732

    2019

  30. [38]

    Yan, Y.; Lin, H.; Zhou, C.; Wang, W.; Sun, H.; Zhan, K.; Lang, X.; Zhou, X.; and Peng, S. 2024. Street gaussians: Modeling dynamic urban scenes with gaussian splatting. In European Conference on Computer Vision, 156--173. Springer

    2024

  31. [39]

    Yu, J.; Lin, Z.; Yang, J.; Shen, X.; Lu, X.; and Huang, T. S. 2018. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, 5505--5514

    2018

  32. [40]

    Yu, J.; Lin, Z.; Yang, J.; Shen, X.; Lu, X.; and Huang, T. S. 2019. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, 4471--4480

    2019

  33. [41]

    Zhang, K.; Fu, J.; and Liu, D. 2022. Flow-guided transformer for video inpainting. In European conference on computer vision, 74--90. Springer

    2022

  34. [42]

    I.; and Xu, Y

    Zhao, S.; Cui, J.; Sheng, Y.; Dong, Y.; Liang, X.; Chang, E. I.; and Xu, Y. 2021. Large scale image completion via co-modulated generative adversarial networks. arXiv preprint arXiv:2103.10428

    2021 arXiv

  35. [43]

    Zhou, H.; Shao, J.; Xu, L.; Bai, D.; Qiu, W.; Liu, B.; Wang, Y.; Geiger, A.; and Liao, Y. 2024 a . Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 21336--21345

    2024

  36. [44]

    Zhou, L.; Zhang, Z.; Jiang, H.; Sun, H.; Bao, H.; and Zhang, G. 2021. DP-MVS: Detail preserving multi-view surface reconstruction of large-scale scenes. Remote Sensing, 13(22): 4569

    2021

  37. [45]

    Zhou, S.; Chang, H.; Jiang, S.; Fan, Z.; Zhu, Z.; Xu, D.; Chari, P.; You, S.; Wang, Z.; and Kadambi, A. 2024 b . Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog...

    2024

  38. [46]

    C.; and Loy, C

    Zhou, S.; Li, C.; Chan, K. C.; and Loy, C. C. 2023. Propainter: Improving propagation and transformer for video inpainting. In Proceedings of the IEEE/CVF international conference on computer vision, 10477--10486

    2023

Pith tools

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