Pith. sign in

REVIEW 4 major objections 9 minor 55 references

GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction

T0 review · 4 major / 9 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read GS-2DGS claims that adding monocular depth and normal supervision to 2D Gaussian Splatting makes reflective-object reconstruction comparable to SDF-based methods while training an order of magnitude faster.

desk verdict Useful, reproducible integration of 2DGS with monocular geometry priors for reflective objects; the speed-accuracy claim holds on the tested benchmarks, but the central dependence on untested priors needs a robustness experiment. read the letter →

arxiv 2506.13110 v1 pith:6TUTGZU6 submitted 2025-06-16 cs.CV

Jinguang Tong , Xuesong li , Fahira Afzal Maken , Sundaram Muthu , Lars Petersson , Chuong Nguyen , Hongdong Li This is my paper
classification cs.CV
keywords reflectiveobjectreconstruction2DGaussiansplattingmonoculardepthestimationnormalinverserenderingrelightingdeferredshadingfoundationmodelpriors
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

This paper tries to show that highly reflective objects can be reconstructed accurately and quickly by adding geometric supervision from monocular depth and normal prediction models to 2D Gaussian Splatting. The claim is that the resulting method, GS-2DGS, reconstructs mesh geometry with a Chamfer-L1 error of 0.0068 on the Glossy Blender dataset, beating all compared Gaussian-splatting baselines and approaching the SDF-based NeRO at 0.0042, while training in 0.7 hours instead of 12. The ablation attributes most of the gain to the geometric supervision: without it the Chamfer error is 0.0481, with it 0.0084. A physically based shading pipeline with deferred shading is then added to model reflections and relighting, raising rendering quality without sacrificing the speed advantage. If true, it closes most of the accuracy gap between fast Gaussian methods and slow SDF methods on specular objects, at an order-of-magnitude lower cost.

What carries the argument

The load-bearing object is the normal-and-depth-regularised 2D Gaussian surface: planar Gaussian discs whose plane normals are directly renderable, combined with a scale-invariant depth loss and L1 and cosine normal losses against foundation-model predictions. This is what injects geometric consistency into Gaussian splatting, where unconstrained Gaussians otherwise drift to floaters and bumpy surfaces. A secondary mechanism is deferred shading: geometry, normals, and PBR parameters are first rendered into a G-buffer, and shading is computed once at the visible surface intersection, which the paper shows yields cleaner relighting and a more accurate recovered environment map than shading every Gaussian separately.

What would settle it

Measure reconstruction on a set of highly specular real objects with ground-truth geometry, such as laser-scanned shiny parts, and compare Chamfer distance with and without the foundation-model losses. The central claim would be falsified if, on some object family where the monocular priors are systematically biased by strong reflections, the geometrically supervised variant reproduces that bias and scores worse than unsupervised 2DGS, or if the full method's error exceeds SDF-based methods by more than a small margin.

Watch

Extended reading notes

Core claim

GS-2DGS is a reconstruction framework built on 2D Gaussian Splatting in which each surface element is a flat, oriented planar Gaussian with a well-defined normal. The method renders depth and normal maps from these primitives and supervises them with pseudo ground truth from monocular foundation models: Depth Pro for metric depth and a Marigold variant fine-tuned for normals. Because those models predict geometry from a single image, they are not confused by view-dependent specular appearance. The rendered geometry is therefore pulled toward a smooth, reflection-independent surface, which the paper shows is the main source of accuracy, dropping Chamfer-L1 from 0.0481 to 0.0084. A Cook-Torrance PBR pipeline with a trainable HDR environment map, a split-sum specular approximation, and deferred shading then explains the view-dependent reflections, producing relightable material parameters and improved environment-lighting estimation. On the Glossy Blender benchmark the reported reconstruction error is 0.0068, better than all compared Gaussian-based methods and second only to NeRO among SDF-based methods, at 0.7 hours training time.

Load-bearing premise

The load-bearing premise is that the monocular depth and normal maps produced by Depth Pro and the fine-tuned Marigold model are accurate enough on reflective surfaces to serve as pseudo ground truth; on unseen specular objects where those priors are wrong, the supervised surface will inherit their bias and the reported speed-accuracy advantage may disappear.

Editorial extensions

If this is right

  • If the reported numbers hold, Gaussian-based reconstructions of reflective objects can match SDF-based accuracy for the first time while keeping training time under an hour on a single GPU, making iterative mesh extraction feasible in practical workflows.
  • Relighting becomes usable in practice: at 160 FPS rendering on the Glossy Blender benchmark, the method outperforms all compared Gaussian baselines on PSNR and SSIM under new environment lights, suggesting real-time relighting of captured specular objects is within reach.
  • The ablation isolates geometry supervision as the dominant factor, implying that future Gaussian reconstruction methods should treat monocular depth and normal priors as a primary source of surface constraints rather than a side regularizer.
  • Because the representation is explicitly planar and surface-aligned, downstream tasks such as material decomposition, normal-map extraction, and mesh export inherit a consistent normal field from the same representation.

Reading between the lines

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

  • The method's generalisation is only as strong as the priors: if a real-world specular object lies outside the training distribution of Depth Pro and the fine-tuned Marigold normal model, the geometric supervision will pull the surface toward the model's bias rather than true geometry, and the advertised speed-accuracy trade-off may not transfer.
  • An untested extension implied by the framework is applying the same monocular geometry supervision to transparent objects or strong inter-reflections, where multiview consistency fails; the reliance on single-view priors is exactly what could make that work, provided the priors hold.
  • A reader could test the method's sensitivity by swapping Depth Pro for a weaker or stronger depth predictor and measuring Chamfer distance; the ablation predicts a monotone relationship between prior accuracy and reconstruction quality.
  • The paper implicitly argues that SDF-level smoothness is unnecessary when good single-view priors are available; a fair test is an object with very high-frequency surface detail, where SDF methods over-smooth and the Gaussian representation should show an advantage.
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 / 9 minor

Summary. The paper proposes GS-2DGS, a reconstruction and relighting method for highly reflective objects built on 2D Gaussian Splatting. It augments 2DGS with monocular geometric supervision from two foundation models (Depth Pro for depth, fine-tuned Marigold for normals), adds a physics-based rendering (PBR) pipeline with a trainable environment map, and applies deferred shading to improve lighting estimation. Experiments on the Glossy Blender and StanfordORB datasets report Chamfer distances that outperform existing Gaussian-based methods, approach the SDF-based NeRO baseline, while training in about 0.7 hours compared to 6–12 hours for the SDF methods. The ablation study in Table 4 shows that the foundation-model geometric supervision is the main driver of reconstruction improvement (Chamfer-L1 dropping from 0.0481 to 0.0084), with PBR and deferred shading contributing smaller gains.

Significance. If the reported results are robust, GS-2DGS offers a practically attractive speed-accuracy trade-off for reflective object reconstruction: close to SDF-level geometry at an order-of-magnitude lower training time, and clearly better geometry than previous Gaussian-based PBR methods. The paper includes several strengths: it builds on a well-established baseline (2DGS), provides an ablation isolating the contribution of each component, evaluates on both synthetic and real datasets, and releases code. The main caveats are that the geometric supervision relies on foundation-model priors whose generalization to unseen specular objects is not tested, and that the comparison set omits the closely related VCR-gaus method, which also uses depth-normal regularizers. The claimed contribution of deferred shading is also less novel than stated, since it is explicitly credited to prior work [47] in the related-work section. Overall, the core idea is sound but needs additional validation before the central claims can be fully accepted.

major comments (4)
  1. [§4, Table 1] The paper claims state-of-the-art results among Gaussian-based methods, but the comparison set omits VCR-gaus [8], a Gaussian surface reconstruction method that also uses depth-normal regularization and is cited in Section 3.2. Since the core contribution is geometric supervision from monocular priors, a direct comparison with VCR-gaus is necessary to support the claim that GS-2DGS outperforms existing Gaussian-based techniques. Without this baseline, the SOTA claim among Gaussian methods is not established.
  2. [§4.1, Tables 1 and 2] The reconstruction quality is measured by Chamfer-L1 distance, which requires extracting a mesh from the 2D Gaussian representation. The paper never specifies the mesh extraction procedure (e.g., Poisson reconstruction, TSDF, or splat-based rendering) or its parameters. Different extraction methods can substantially change Chamfer values, so the reported comparisons are not reproducible and may not be fair to baselines. Please describe the extraction step and ideally release the script.
  3. [§4.3, Table 4 and §3.3] The main reconstruction improvement is attributed to geometric supervision from Depth Pro and fine-tuned Marigold, with the paper asserting in Section 3.3 that these priors are 'insusceptible to the reflective surface problem.' However, the only evidence is a binary on/off ablation; there is no experiment that varies the quality of the pseudo ground truth, for example by corrupting the depth/normal predictions, swapping in different foundation models, or evaluating on objects with anisotropic reflections or inter-reflections. If the priors systematically fail on unseen specular objects, the geometry will be biased by the priors and the advertised speed-accuracy advantage may disappear. A robustness test explicitly manipulating prior quality is needed to support the generalization claim.
  4. [§4, Tables 1–4] All quantitative results are reported as single runs without error bars or multiple seeds. While the improvement from adding geometric supervision is large (0.0481 to 0.0084 in Table 4), the subsequent increments from PBR and deferred shading are small (0.0084 to 0.0074 to 0.0068) and could easily fall within run-to-run variation of Gaussian optimization. Please report mean and standard deviation over at least three independent runs for the main tables, or otherwise justify that the differences are significant.
minor comments (9)
  1. [Table 3] The column header reads 'SSIM↓', but SSIM is higher-better; the arrow should be '↑'. The same issue appears in Table S6.
  2. [§1, Contributions] The contribution list states 'We introduce the deferred shading technique', which conflicts with Section 2.2 where the paper credits Ye et al. [47] as the first to introduce deferred shading to Gaussian Splatting. Please rephrase to make clear that the contribution is the adaptation of deferred shading to reflective object reconstruction within the PBR pipeline, not the technique itself.
  3. [§3.3] There is a typo: 'we use the such models' should be 'we use such models'.
  4. [§3.4.1, Eq. (10)] The roughness parameter is introduced as ρ in Eq. (10) but is later denoted as r in the text and in Section 4 (Implementation Details). Please unify the notation.
  5. [§4.2] The sentence 'Our approach achieves the highest relighting quality among all Gaussian-based methods' appears twice in the same paragraph; please remove the duplicate.
  6. [Figure 1] The caption says 'GS-2DSG'; this should be 'GS-2DGS'.
  7. [§4, Implementation Details] The reported training time of 0.7 hours appears to include only optimization, not the inference time of Depth Pro and Marigold on the training images. For a fair speed comparison against prior methods, please report the total wall-clock time including foundation-model prediction, or state that the comparison is optimization-only.
  8. [Supplementary, Table S5] The training scene for each StanfordORB object is described as randomly selected. Since this choice can affect the reconstruction results, please provide the specific split used or release the exact scene indices for reproducibility.
  9. [§2.1] There is a typo '2D Gausians' that should be '2D Gaussians'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is a supervised training pipeline whose geometry supervision comes from external pre-trained foundation models, benchmarked against held-out ground-truth meshes.

full rationale

The paper's derivation chain combines 2DGS rendering, Cook-Torrance PBR, deferred shading, and geometric supervision from Depth Pro and fine-tuned Marigold. None of the reported target quantities (Chamfer-L1 against ground-truth meshes, PSNR/SSIM under unseen environment lights) is used to define the losses or to fit the reconstruction. The depth supervision in Eq. (8) solves per-image scale and shift parameters by least squares, but these only align the rendered depth to the pseudo-depth; they do not make the reconstruction tautological, because the final evaluation is against independent ground-truth geometry, not against the pseudo-depth. The geometric priors are external pre-trained models with released code and weights, not parameters fitted in this paper, so they constitute independent evidence rather than a self-citation chain. The only self-citation is [41] in the related-work discussion of transparent/specular reconstruction and is not load-bearing for any claim. The central speed-accuracy claim is tested on held-out real data (StanfordORB) and with relighting under new environments, which is outward-looking. No equation reduces to its own input, and no claimed prediction is a renamed fit. Therefore no circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper's central contribution is an engineering system built from standard components; the only input beyond published building blocks is the pretrained geometry prior, which is the main source of the quality gain and the main fragility.

free parameters (5)
  • lambda_n = 0.5
    Weight for normal consistency loss Ln, chosen by hand for stage 1; central to the geometric supervision contribution.
  • lambda_d = 0.05
    Weight for depth alignment loss Ld, chosen by hand.
  • lambda_light = 0.002
    Weight for environment lighting gray-world regularization.
  • lambda_a/m/r = 0.05/0.05/0.01
    Weights for PBR parameter smoothness losses (albedo, metallic, roughness); hand-chosen.
  • scale and shift in depth loss = solved per image
    omega and b in Eq. (8) computed by least squares for each depth map; these are per-image parameters used to align rendered depth to the prior, not global, but they allow the loss to discard any global scale bias from the prior.
assumptions (4)
  • domain assumption The monocular depth and normal estimates from Depth Pro and fine-tuned Marigold are accurate enough on reflective objects to serve as pseudo ground truth.
    Section 3.3; the ablation shows this supervision is the main source of improvement, so if false, the central result fails.
  • domain assumption The Cook-Torrance BRDF with split-sum approximation can model the reflectance of the tested objects.
    Section 3.4; standard approximation, but it assumes isotropic GGX and a single envmap.
  • domain assumption 2D Gaussian primitives can represent the surface unambiguously; normals are defined by cross product of tangent vectors (Eq. 4).
    Section 3.2; inherited from 2DGS [20].
  • domain assumption The environment lighting is represented by a trainable HDR cube map without accounting for inter-reflections.
    Section 3.4; this is an approximation used by prior inverse rendering methods.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction." pith.science (2026). https://pith.science/paper/6TUTGZU6

@misc{pith2026250613110,
  author       = {Pith},
  title        = {Pith review of: GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6TUTGZU6}},
  note         = {Machine review of arXiv:2506.13110}
}
read the original abstract

3D modeling of highly reflective objects remains challenging due to strong view-dependent appearances. While previous SDF-based methods can recover high-quality meshes, they are often time-consuming and tend to produce over-smoothed surfaces. In contrast, 3D Gaussian Splatting (3DGS) offers the advantage of high speed and detailed real-time rendering, but extracting surfaces from the Gaussians can be noisy due to the lack of geometric constraints. To bridge the gap between these approaches, we propose a novel reconstruction method called GS-2DGS for reflective objects based on 2D Gaussian Splatting (2DGS). Our approach combines the rapid rendering capabilities of Gaussian Splatting with additional geometric information from foundation models. Experimental results on synthetic and real datasets demonstrate that our method significantly outperforms Gaussian-based techniques in terms of reconstruction and relighting and achieves performance comparable to SDF-based methods while being an order of magnitude faster. Code is available at https://github.com/hirotong/GS2DGS

Figures

Figures reproduced from arXiv: 2506.13110 by the authors.

Figure 1
Figure 1. The overview of our GS-2DSG method for reflective object reconstruction. The foundation models supervise 2DGS reconstruc [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Quality of the normal and depth predictions from the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The comparison between forward shading and deferred [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Reconstruction results on the Glossy Bender dataset. Our method achieves significantly better results compared to other Gaussian [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Comparison of the results on the StanfordORB dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Relighting results on the Glossy Bender dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Comparison of environment lighting estimation quality [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 34 canonical work pages

  1. [47]

    3d gaussian splat- ting with deferred reflection

    Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splat- ting with deferred reflection. InACM SIGGRAPH 2024 Con- ference Papers, pages 1–10, 2024. 5

    work page 2024

  2. [8]

    Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction

    Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yun- song Wang, and Gim Hee Lee. Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction. arXiv preprint arXiv:2406.05774, 2024. 3

    arXiv 2024

  3. [1]

    Rethinking induc- tive biases for surface normal estimation

    Gwangbin Bae and Andrew J Davison. Rethinking induc- tive biases for surface normal estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9535–9545, 2024. 2

    work page 2024

  4. [2]

    Mip-nerf: A multiscale representation for anti-aliasing neu- ral 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 neu- ral radiance fields. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 5855–5864,

  5. [3]

    Depth pro: Sharp monocular metric depth in less than a second.arXiv preprint arXiv:2410.02073, 2024

    Aleksei Bochkovskii, Ama ¨el Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second.arXiv preprint arXiv:2410.02073, 2024. 1, 3, 4, 6

    arXiv 2024

  6. [4]

    High-quality surface splatting on today’s gpus

    Mario Botsch, Alexander Hornung, Matthias Zwicker, and Leif Kobbelt. High-quality surface splatting on today’s gpus. InProceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics, 2005., pages 17–141. IEEE, 2005. 3

    work page 2005

  7. [5]

    Physically-based shading at disney

    Brent Burley and Walt Disney Animation Studios. Physically-based shading at disney. InAcm Siggraph, pages 1–7. vol. 2012, 2012. 4

    work page 2012

  8. [6]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction.arXiv preprint arXiv:2406.06521, 2024

    Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction.arXiv preprint arXiv:2406.06521, 2024. 2

    arXiv 2024

Show all 55 references
  1. [7]

    Neusg: Neural im- plicit surface reconstruction with 3d gaussian splatting guid- ance.arXiv preprint arXiv:2312.00846, 2023

    Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural im- plicit surface reconstruction with 3d gaussian splatting guid- ance.arXiv preprint arXiv:2312.00846, 2023. 3

    2023 arXiv

  2. [9]

    A reflectance model for computer graphics.ACM Siggraph Computer Graphics, 15(3):307–316, 1981

    Robert L Cook and Kenneth E Torrance. A reflectance model for computer graphics.ACM Siggraph Computer Graphics, 15(3):307–316, 1981. 4

    1981

  3. [10]

    High-quality surface recon- struction using gaussian surfels

    Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface recon- struction using gaussian surfels. InACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 1

    2024

  4. [11]

    The triangle processor and normal vector shader: a vlsi system for high performance graphics

    Michael Deering, Stephanie Winner, Bic Schediwy, Chris Duffy, and Neil Hunt. The triangle processor and normal vector shader: a vlsi system for high performance graphics. Acm siggraph computer graphics, 22(4):21–30, 1988. 5

    1988

  5. [12]

    Diffusiondepth: Diffusion denoising approach for monocular depth estima- tion

    Yiquan Duan, Xianda Guo, and Zheng Zhu. Diffusiondepth: Diffusion denoising approach for monocular depth estima- tion. InEuropean Conference on Computer Vision, pages 432–449. Springer, 2025. 2

    2025

  6. [13]

    Omnidata: A scalable pipeline for making multi- task mid-level vision datasets from 3d scans

    Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi- task mid-level vision datasets from 3d scans. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 10786–10796, 2021. 2

    2021

  7. [14]

    Predicting depth, surface nor- mals and semantic labels with a common multi-scale con- volutional architecture

    David Eigen and Rob Fergus. Predicting depth, surface nor- mals and semantic labels with a common multi-scale con- volutional architecture. InProceedings of the IEEE inter- national conference on computer vision, pages 2650–2658,

  8. [15]

    Depth map prediction from a single image using a multi-scale deep net- work.Advances in neural information processing systems, 27, 2014

    David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep net- work.Advances in neural information processing systems, 27, 2014. 2

    2014

  9. [16]

    Geowiz- ard: Unleashing the diffusion priors for 3d geometry esti- mation from a single image

    Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowiz- ard: Unleashing the diffusion priors for 3d geometry esti- mation from a single image. InEuropean Conference on Computer Vision, pages 241–258. Springer, 2025. 1

    2025

  10. [17]

    Planar reflection-aware neural radiance fields.arXiv preprint arXiv:2411.04984, 2024

    Chen Gao, Yipeng Wang, Changil Kim, Jia-Bin Huang, and Johannes Kopf. Planar reflection-aware neural radiance fields.arXiv preprint arXiv:2411.04984, 2024. 2

    2024 arXiv

  11. [18]

    Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray trac- ing.arXiv preprint arXiv:2311.16043, 2023

    Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray trac- ing.arXiv preprint arXiv:2311.16043, 2023. 1, 2, 4, 5, 7

    2023 arXiv

  12. [19]

    Nerfren: Neural radiance fields with reflections

    Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song- Hai Zhang. Nerfren: Neural radiance fields with reflections. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18409–18418, 2022. 2

    2022

  13. [20]

    2d gaussian splatting for geometrically ac- curate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac- curate radiance fields. InACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 1, 2, 5, 7

    2024

  14. [21]

    Transparent and specu- lar object reconstruction

    Ivo Ihrke, Kiriakos N Kutulakos, Hendrik PA Lensch, Mar- cus Magnor, and Wolfgang Heidrich. Transparent and specu- lar object reconstruction. InComputer graphics forum, pages 2400–2426. Wiley Online Library, 2010. 2

    2010

  15. [22]

    Gaussian- shader: 3d gaussian splatting with shading functions for re- flective surfaces

    Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xi- aoxiao Long, Wenping Wang, and Yuexin Ma. Gaussian- shader: 3d gaussian splatting with shading functions for re- flective surfaces. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, pag...

    2024

  16. [23]

    The rendering equation

    James T Kajiya. The rendering equation. InProceedings of the 13th annual conference on Computer graphics and inter- active techniques, pages 143–150, 1986. 1, 4

    1986

  17. [24]

    Real shading in unreal engine 4.Proc

    Brian Karis and Epic Games. Real shading in unreal engine 4.Proc. Physically Based Shading Theory Practice, 4(3):1,

  18. [25]

    Repurpos- ing diffusion-based image generators for monocular depth estimation

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Met- zger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos- ing diffusion-based image generators for monocular depth estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492– ...

    2024

  19. [26]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

  20. [27]

    Stanford-orb: a real-world 3d object inverse rendering benchmark.Advances in Neural Information Processing Systems, 36, 2024

    Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agar- wala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark.Advances in Neural Information Processing Systems, 36, 2024. 6, 7

    2024

  21. [28]

    Tensosdf: Roughness-aware tensorial representation for robust geom- etry and material reconstruction.ACM Transactions on Graphics (TOG), 43(4):1–13, 2024

    Jia Li, Lu Wang, Lei Zhang, and Beibei Wang. Tensosdf: Roughness-aware tensorial representation for robust geom- etry and material reconstruction.ACM Transactions on Graphics (TOG), 43(4):1–13, 2024. 1, 2, 7

    2024

  22. [29]

    Megadepth: Learning single- view depth prediction from internet photos

    Zhengqi Li and Noah Snavely. Megadepth: Learning single- view depth prediction from internet photos. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 2041–2050, 2018. 2

    2018

  23. [30]

    Neuralangelo: High-fidelity neural surface reconstruction

    Zhaoshuo Li, Thomas M ¨uller, Alex Evans, Russell H Tay- lor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 8456–8465, 2023. 2

    2023

  24. [31]

    Gs-ir: 3d gaussian splatting for inverse rendering

    Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21644–21653, 2024. 1, 2, 4, 7

    2024

  25. [32]

    Nero: Neural geometry and brdf reconstruction of reflective objects from multiview images.ACM Transactions on Graphics (TOG), 42(4):1–22, 2023

    Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nero: Neural geometry and brdf reconstruction of reflective objects from multiview images.ACM Transactions on Graphics (TOG), 42(4):1–22, 2023. 1, 2, 5, 6, 7

    2023

  26. [33]

    Stealing stable dif- fusion prior for robust monocular depth estimation.arXiv preprint arXiv:2403.05056, 2024

    Yifan Mao, Jian Liu, and Xianming Liu. Stealing stable dif- fusion prior for robust monocular depth estimation.arXiv preprint arXiv:2403.05056, 2024. 2

    2024 arXiv

  27. [34]

    Fine-tuning image-conditional diffusion models is easier than you think

    Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. InProceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision (WACV),

  28. [35]

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

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

    2021

  29. [36]

    Extracting triangular 3d models, materials, and lighting from images

    Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas M¨uller, and Sanja Fi- dler. Extracting triangular 3d models, materials, and lighting from images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pa...

    2022

  30. [37]

    Ren ´e Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer.IEEE transactions on pattern analysis and machine intelligence, 44(3):1623–1637, 2020. 2, 4

    2020

  31. [38]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. InProceedings of the IEEE/CVF international conference on computer vision, pages 12179–12188, 2021. 2

    2021

  32. [39]

    Make3d: Learning 3d scene structure from a single still image.IEEE transactions on pattern analysis and machine intelligence, 31(5):824–840, 2008

    Ashutosh Saxena, Min Sun, and Andrew Y Ng. Make3d: Learning 3d scene structure from a single still image.IEEE transactions on pattern analysis and machine intelligence, 31(5):824–840, 2008. 2

    2008

  33. [40]

    The surprising effectiveness of diffusion models for optical flow and monocular depth estimation.Advances in Neural Information Processing Systems, 36, 2024

    Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation.Advances in Neural Information Processing Systems, 36, 2024. 2

    2024

  34. [41]

    Seeing through the glass: Neural 3d reconstruction of object inside a transpar- ent container

    Jinguang Tong, Sundaram Muthu, Fahira Afzal Maken, Chuong Nguyen, and Hongdong Li. Seeing through the glass: Neural 3d reconstruction of object inside a transpar- ent container. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12555– 1...

    2023

  35. [42]

    Ref-nerf: Struc- tured view-dependent appearance for neural radiance fields

    Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Struc- tured view-dependent appearance for neural radiance fields. In2022 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 5481–5490. IEE...

    2022

  36. [43]

    Microfacet models for refraction through rough surfaces.Rendering techniques, 2007:18th, 2007

    Bruce Walter, Stephen R Marschner, Hongsong Li, and Ken- neth E Torrance. Microfacet models for refraction through rough surfaces.Rendering techniques, 2007:18th, 2007. 4

    2007

  37. [44]

    Neus: Learning neural im- plicit surfaces by volume rendering for multi-view recon- struction.Advances in Neural Information Processing Sys- tems, 34:27171–27183, 2021

    Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural im- plicit surfaces by volume rendering for multi-view recon- struction.Advances in Neural Information Processing Sys- tems, 34:27171–27183, 2021. 7

    2021

  38. [45]

    Neus2: Fast learning of neural implicit surfaces for multi-view recon- struction

    Yiming Wang, Qin Han, Marc Habermann, Kostas Dani- ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view recon- struction. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 2

    2023

  39. [46]

    Transformer-based attention networks for continuous pixel-wise prediction

    Guanglei Yang, Hao Tang, Mingli Ding, Nicu Sebe, and Elisa Ricci. Transformer-based attention networks for continuous pixel-wise prediction. InProceedings of the IEEE/CVF International Conference on Computer vision, pages 16269–16279, 2021. 2

    2021

  40. [48]

    Geonet: Unsupervised learn- ing of dense depth, optical flow and camera pose

    Zhichao Yin and Jianping Shi. Geonet: Unsupervised learn- ing of dense depth, optical flow and camera pose. InPro- ceedings of the IEEE conference on computer vision and pat- tern recognition, pages 1983–1992, 2018. 2

    1983

  41. [49]

    Plenoctrees for real-time rendering of neural radiance fields

    Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752– 5761, 2021. 2

    2021

  42. [50]

    Gsdf: 3dgs meets sdf for improved rendering and reconstruction.arXiv preprint arXiv:2403.16964, 2024

    Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction.arXiv preprint arXiv:2403.16964, 2024. 1, 2

    2024 arXiv

  43. [51]

    Monovit: Self-supervised monocular depth estimation with a vision transformer

    Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and Ste- fano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. In2022 international conference on 3D vision (3DV), pages 668–678. IEEE...

    2022

  44. [52]

    Unleashing text-to-image diffu- sion models for visual perception

    Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffu- sion models for visual perception. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5729–5739, 2023. 2

    2023

  45. [53]

    Instant-nerf: In- stant on-device neural radiance field training via algorithm- accelerator co-designed near-memory processing

    Yang Katie Zhao, Shang Wu, Jingqun Zhang, Sixu Li, Chaojian Li, and Yingyan Celine Lin. Instant-nerf: In- stant on-device neural radiance field training via algorithm- accelerator co-designed near-memory processing. In2023 60th ACM/IEEE Design Automation Conference (DAC), page...

    2023

  46. [54]

    Representation We implement GS-2DGS mainly based on the original 2DGS

    Additional Implementation Details 1.1. Representation We implement GS-2DGS mainly based on the original 2DGS. The 2DGS uses {x,s,t} and {α,c} to represent its geometric and volumetric appearance properties respec- tively, where x is the position of the Gaussian, s and t are th...

  47. [55]

    Results on Glossy Blender dataset ReconstructionFig

    Additional Experiments 2.1. Results on Glossy Blender dataset ReconstructionFig. S3 shows the additional reconstruc- tion results on the Glossy Blender dataset. Our method achieves the best reconstruction quality among all the Gaussian-based methods. Compared to the SDF-based ...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.