Pith. sign in

REVIEW 3 major objections 5 minor 46 references

3D Scene-Camera Representation with Joint Camera Photometric Optimization

T0 review · 3 major / 5 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 Jointly optimizing a photometric camera model with the 3D Gaussian scene separates vignetting, sensor non-uniformity, and lens contaminants from the scene radiance field, recovering clean novel views and a per-camera distortion map.

desk verdict A sensible coupling of DSO-style photometric calibration to 3DGS with a useful custom distorted dataset, but the 'separation' claim outruns the evidence because the camera/scene split is not identifiable from the loss alone. read the letter →

arxiv 2506.20979 v1 pith:TZYTXYHL submitted 2025-06-26 cs.CV

Weichen Dai , Kangcheng Ma , Jiaxin Wang , Kecen Pan , Yuhang Ming , Hua Zhang , Wanzeng Kong This is my paper · ORCID
classification cs.CV
keywords 3DGaussiansplattingcameraphotometricmodelnovelviewsynthesisdistortionvignettinglenscontaminationdefocusdepthregularization
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

Multi-view 3D reconstruction assumes the camera records scene radiance faithfully, but real lenses vignette, sensors respond non-uniformly, and lens surfaces accumulate dirt, fingerprints, and water droplets. This paper claims that if the reconstruction pipeline does not account for these camera-side photometric effects, the radiance field will absorb them as if they were scene content, degrading novel views. The proposed method therefore builds a full camera photometric model — an internal multiplicative factor for vignetting and sensor response, an external attenuation-plus-emission factor for lens contaminants, and a defocus-based circular averaging — and fits it as a shallow MLP that is optimized alternately with a 3D Gaussian scene representation. An inverse-depth opacity regularization stops the camera model from stealing scene detail when constraints are sparse. The result, if correct, is a cleaner scene radiance field under imaging degradation, with the camera photometric map recovered as a useful byproduct.

What carries the argument

The load-bearing object is the full photometric image-formation model and its MLP parametrization. Combining the internal model $I_i(x)=M(x)B_i(x)$ with the external model $\hat{B}_i(x)=S_\alpha(x)R(x)+S_\beta(x)$ and the defocus model yields a rendered pixel $I_i(x)=M(x)\sum_{\hat{x}\in \mathrm{CoC}_p}\bigl(S_\alpha(\hat{x})R(\hat{x})+S_\beta(\hat{x})\bigr)/(\pi R_{\mathrm{CoC}}^2)$, a mean convolution of the externally distorted radiance over a circle of confusion. Under the in-focus simplification, the paper fits the spatially varying factors with two shallow MLPs — $\mathrm{mlp}_\alpha$ for the internal multiplier and $\mathrm{mlp}_\beta$, $\mathrm{mlp}_\gamma$ for the external attenuation and emission — whose outputs multiply and convolve the radiance $R(x)$ coming from Gaussian splatting. Because the MLP shares parameters across all pixels, neighboring photometric parameters stay correlated, which is the compactness assumption that lets the camera model absorb smooth vignetting and blurred lens smudges instead of scene structure. The second mechanism is a depth regularization: while fitting the camera MLP, opacity along each ray is reweighted by a Gaussian $G_{\mathrm{inv}}(id)$ in inverse depth, suppressing floating points so that the camera model does not steal scene detail when constraints are sparse.

What would settle it

Render a synthetic scene with known ground-truth sharp color and known depth-dependent defocus blur of varying circle-of-confusion radius, then run the method: if the recovered camera photometric map reproduces the depth-dependent blur pattern while the recovered scene is incorrectly sharpened in blurred regions, the in-focus simplification (from Eq. 6 to Eq. 7) is falsified. A practical version is to capture the same scene twice with the same lens, once with deep depth of field and once with shallow depth of field, and check whether the estimated camera photometric map is identical in both runs.

Watch

Extended reading notes

Core claim

Radiance-field methods usually treat the camera as a perfect pinhole that records scene radiance linearly; this paper argues that real cameras degrade images in two separable ways — inside the body, via vignetting and a spatially non-uniform sensor response collected into a multiplicative factor $M(x)=G(x)V(x)$, and at the lens surface, where contaminants attenuate scene radiance by $S_\alpha(x)$ and add their own radiance $S_\beta(x)$. The paper claims that by representing the full photometric model as a shallow MLP that is optimized alternately with a 3D Gaussian scene, and by guarding the camera optimization with an inverse-depth opacity regularization, the two components separate: the scene radiance field becomes clean and the MLP outputs a camera photometric map as a byproduct. Experiments on a new dataset of real captures with fingerprints, dirt, water droplets, and vignetting, together with public datasets, support the claim that jointly estimating the camera photometric model improves novel-view fidelity under photometric degradation.

Load-bearing premise

The method's core simplification is that every scene point lies inside the depth of field, so the blur of out-of-focus scene content is ignored, and lens contaminants are treated as sitting at one fixed distance with a constant circle-of-confusion radius; if a scene contains genuinely out-of-focus regions, the camera model will absorb that blur as if it were a lens artifact.

Editorial extensions

If this is right

  • On the paper's custom dataset of real captures with fingerprints, dirt, water droplets, and vignetting, the joint representation beats distortion-blind baselines by roughly 1–4 dB PSNR, with the largest gains on the strongest distortions.
  • Because the camera model is fitted per camera rather than per scene, the optimization yields a camera photometric map as a byproduct; the paper shows that two scenes captured by the same camera produce similar internal photometric maps.
  • The ablation attributes gains to all three components — the camera photometric representation, the defocus model, and the depth regularization — and shows that the defocus model and the depth regularization reinforce each other when used together.
  • On clean public data (NeRF-Synthetic and MipNeRF 360), the method roughly matches or exceeds the strongest splatting baselines, indicating that the camera model captures real distortions without overfitting clean images.
  • The recovered photometric parameter maps can be visualized by feeding the camera MLP a uniform maximum-value image, which offers a direct way to inspect what distortion the model attributes to the camera rather than the scene.

Reading between the lines

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

  • The recovered camera photometric map is effectively a per-camera calibration artifact: it could plausibly be transferred across all scenes shot by the same physical camera, or used to pre-correct raw images before feeding other 3D reconstruction pipelines — an application the paper does not explicitly test.
  • Allowing the circle-of-confusion radius to vary spatially instead of fixing it as a hyperparameter would extend the model to contaminants sitting at different distances from the lens and would partially correct genuine out-of-focus scene blur.
  • A practical diagnostic follows directly: the MLP photometric maps could be monitored over time as an indicator of lens contamination or degradation in robots, phones, and long-running capture rigs.
  • Extending the internal model to time-varying exposure and auto-gain settings would let the same joint optimization separate changing exposure from fixed camera photometric parameters, mirroring how direct odometry methods treat affine brightness.
  • A testable extension would be to feed the method images whose distortion is already precisely calibrated, then check whether the recovered camera map converges to the known ground-truth map; the paper does not report such a quantitative calibration check.
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

3 major / 5 minor

Summary. This paper proposes a joint camera photometric model for 3D Gaussian Splatting. The model separates internal photometric distortions (vignetting and sensor non-uniformity) from external distortions (lens contaminants) through a multiplicative MLP and an additive MLP, with a defocus convolution over a circle of confusion. The camera representation is optimized alternately with the scene radiance field, using a depth regularization that suppresses Gaussians away from the surface. Experiments on a custom dataset with fingerprints, dirt, water droplets, and vignetting show consistent PSNR improvements over several NeRF and 3DGS baselines, and additional experiments on public datasets and with three different cameras are reported. The paper claims that the method effectively separates scene-unrelated information from the 3D scene representation.

Significance. The paper's strength lies in its physically motivated full photometric model and the systematic ablation showing contributions from the camera photometric representation, depth regularization, and defocus model. The consistent gains on the custom degraded dataset and the observation that the recovered internal photometric maps are similar across two scenes with the same camera are encouraging. If the separation claim were validated, the method would be a practical tool for robust scene reconstruction under lens contamination and would produce a reusable camera photometric map. However, the absence of ground-truth validation of the recovered camera maps currently prevents the central claim from being accepted.

major comments (3)
  1. [Section 3.3.1, Eq. (10) and Section 4] The decomposition is not identifiable from the photometric loss. The loss only constrains the product M(x)(S_alpha(x)R(x)+S_beta(x)) or its defocus-convolved form, so scaling M by a constant and S_alpha/S_beta by its inverse leaves the objective unchanged. The depth regularization in Eqs. (11)-(13) constrains the opacity distribution along rays, but not the photometric split. The central claim of separating scene-unrelated information is therefore not established. The only evidence is the qualitative similarity in Fig. 9; there is no synthetic experiment with injected known distortions and no quantitative comparison of recovered camera maps to ground truth. Since the unique contribution of the paper is this separation, the authors should add a synthetic validation: apply a known camera distortion to clean renders, run the method, and report the error between the recovered camera maps and the injected ones, as well as the fidelity of the cleaned scene renderings against the clean ground truth.
  2. [Section 3.1.4, Eqs. (6)-(7)] The simplification from Eq. (6) to Eq. (7) is not justified. The sentence "the points in the scene are considered to be within the DoF" does not resolve the inconsistency, because the defocus model in Eqs. (4)-(5) targets contaminants on the lens surface, which are at a very small object distance and are never within the scene's depth of field. If the implementation follows Eq. (9), which includes the convolution, then Eq. (7) is not the model actually used; if Eq. (7) is used, the ablation benefit of the defocus model (Table 4, DM row) is unexplained. The authors should clarify which equation is implemented and either derive Eq. (7) as a true special case or remove it.
  3. [Section 3.3.2, Eq. (12)] The standard deviation sigma_id in the Gaussian function G_inv is never defined. The depth regularization behaves very differently depending on sigma_id: a large value flattens the Gaussian and disables the regularization, while a small value suppresses many Gaussians. This parameter therefore directly affects the optimization and the claims about preventing overfitting. The authors must specify how sigma_id is computed, whether it is fixed or learned, and how it is initialized.
minor comments (5)
  1. [Section 3.2.2] The text says "mlp_beta and mlp_gamma are the outputs of the MLP_e representing the internal photometric distortion," but the second instance should be "external" instead of "internal."
  2. [Section 3.1 and Table 1] There are several typos, including "pipline" in Section 3.1 and the missing space in "3DGS-MCMC27.81" in Table 1.
  3. [Eq. (11)] The subscript "i" in G_inv(id_i) is not defined; the rendering equation previously uses index n for Gaussians.
  4. [Section 3.1.3] Eq. (4) cites reference [16] for the depth-of-field model, but reference [16] appears to be a paper on stress birefringence in optical fibers, which is not clearly a standard DoF reference; the authors should verify the citation.
  5. [Section 4.4] The claim that the photometric parameter maps of two scenes are "similar" is only qualitative; a quantitative metric, such as correlation or mean squared error after alignment, would strengthen the evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the camera photometric model is an explicitly fitted component, not a prediction derived from itself; under-determination of the decomposition is a correctness concern, not a definitional tautology.

full rationale

The paper's derivation chain is: physical photometric model (Eqs. 2-6), the rendering equation used in the implementation (Eqs. 7-9), photometric loss (Eq. 10), depth regularization (Eqs. 11-13), and alternating optimization. The camera MLP outputs are estimated parameters in this chain, not predictions obtained from the same quantity they are supposed to explain. The central claim of 'separating scene-unrelated information' is asserted as an effect of the joint optimization; the paper does not define 'scene-unrelated' as 'whatever the MLP learns' and then present that definition as a result. It offers cross-scene consistency of the recovered photometric maps (Fig. 9) as independent, though qualitative, evidence. The skeptic's identifiability argument is a valid correctness and rigor limitation: Eq. (10) supervises only the composed image, so the split between the scene radiance R and the camera MLP is not uniquely pinned down by the loss, and Eq. (7) simplifies Eq. (6) using an in-DoF assumption that is not consistently enforced in the implementation Eq. (9). Under-determination and modeling inconsistency, however, are not circularity under the stated rules: no fitted parameter is renamed as a prediction that the same equations force, and no conclusion is assumed in its own premise. There are no load-bearing self-citations; the only author self-citation ([14], Z. Gao, W. Dai, Y. Zhang) appears in a related-work list and is not used to justify the method's premise. No uniqueness theorem or ansatz is imported from the authors' prior work. The method is evaluated against external baselines on real captured distorted data and on public datasets. Accordingly, no step reduces to its own input by construction, and the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 6 assumptions · 2 invented entities

The central claim rests on a physically motivated but partly unvalidated camera model: multiplicative internal attenuation, affine external contamination, and a constant-CoC defocus kernel. The main hand-set constants are R_CoC and lambda; the MLP weights are learned from the same images used for evaluation. The scene-camera split is regularized by a depth prior rather than by external photometric calibration, so identifiability is a concern.

free parameters (2)
  • R_CoC (circle of confusion radius) = set via hyperparameter, value not reported
    Controls the defocus blur radius in Eq. (5) and (9); the paper states it is constant for a given scene and set via a hyperparameter.
  • lambda (SSIM weight in photometric loss) = not reported
    Balances L1 and D-SSIM in Eq. (10); chosen by hand.
assumptions (6)
  • domain assumption Pinhole geometric model with radial distortion corrected in preprocessing (Section 3.1)
    The projection pi maps 3D points to pixels assuming an ideal pinhole and pre-corrected radial distortion.
  • domain assumption Internal photometric factorization I_i(x) = G(x) V(x) B_i(x) (Eq. 2)
    Assumes vignetting and sensor non-uniform response act as a per-pixel multiplicative attenuation independent of scene content.
  • domain assumption External contaminant model B_hat = S_alpha R + S_beta (Eq. 3)
    Assumes contaminants attenuate scene radiance multiplicatively and add their own radiance additively; no physical validation is given.
  • domain assumption Constant circle of confusion for lens-surface contaminants (Eq. 4-5)
    Assumes contaminants lie on the lens with fixed object distance so R_CoC is constant.
  • domain assumption All scene points are within the depth of field, collapsing Eq. (6) to Eq. (7)
    The paper drops the CoC convolution with the sentence 'the points in the scene are considered to be within the DoF,' which is not proven and conflicts with the defocus model.
  • domain assumption Opacity along a ray is unimodal and peaked at the object surface (Section 3.3.2)
    The depth regularization assumes max opacity at the surface, which may fail for transparent or multi-layer scenes.
invented entities (2)
  • Internal photometric attenuation field M(x)
    purpose: Represents combined vignetting and sensor non-uniform response as per-pixel multipliers
    The field is estimated only from the target images; no radiometric calibration or ground-truth vignetting map is used to anchor it.
  • External contaminant fields S_alpha(x) and S_beta(x)
    purpose: Represent attenuation and additive emission from lens dirt, smudges, or droplets
    These fields are interior to the MLP and are never compared to measured contaminant properties; the only support is improved reconstruction metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of 3D Scene-Camera Representation with Joint Camera Photometric Optimization." pith.science (2026). https://pith.science/paper/TZYTXYHL

@misc{pith2026250620979,
  author       = {Pith},
  title        = {Pith review of: 3D Scene-Camera Representation with Joint Camera Photometric Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TZYTXYHL}},
  note         = {Machine review of arXiv:2506.20979}
}
read the original abstract

Representing scenes from multi-view images is a crucial task in computer vision with extensive applications. However, inherent photometric distortions in the camera imaging can significantly degrade image quality. Without accounting for these distortions, the 3D scene representation may inadvertently incorporate erroneous information unrelated to the scene, diminishing the quality of the representation. In this paper, we propose a novel 3D scene-camera representation with joint camera photometric optimization. By introducing internal and external photometric model, we propose a full photometric model and corresponding camera representation. Based on simultaneously optimizing the parameters of the camera representation, the proposed method effectively separates scene-unrelated information from the 3D scene representation. Additionally, during the optimization of the photometric parameters, we introduce a depth regularization to prevent the 3D scene representation from fitting scene-unrelated information. By incorporating the camera model as part of the mapping process, the proposed method constructs a complete map that includes both the scene radiance field and the camera photometric model. Experimental results demonstrate that the proposed method can achieve high-quality 3D scene representations, even under conditions of imaging degradation, such as vignetting and dirt.

Figures

Figures reproduced from arXiv: 2506.20979 by the authors.

Figure 1
Figure 1. Comparison of different methods using the images with photometric distortions. The input images contain severe photo￾metric distortions. Our method effectively separates these distortions (The left image is affected by water droplets, and the right image by vignetting), leading to a more accurate reconstruction with higher fidelity. In each group, the bottom-left image visualizes the photometric distortion modeled b… view at source ↗
Figure 2
Figure 2. Comparison of image formation pipeline models. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Vignetting under different aperture settings. It is evident [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: System overview. In the figure, red arrows indicate camera representation training and black arrows indicate scene representation [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Opacity distribution along one ray. The set of purple [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Different mixed photometric distortions a single image with specific camera coordinates and an￾gles. Therefore, the optimization process adopts an itera￾tive training approach to optimize the camera photometric parameters and the scene photometric parameters sequen￾tia…
Figure 7
Figure 7. Figure 7: Comparison of different methods using the images with [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Visualizing the comparison on the public dataset [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Comparison on different scene images captured with the [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 34 canonical work pages

  1. [1]

    Digital image restoration.IEEE signal processing magazine, 14(2):24–41,

    Mark R Banham and Aggelos K Katsaggelos. Digital image restoration.IEEE signal processing magazine, 14(2):24–41,

    work page

  2. [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,

  3. [3]

    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. 2, 6

    work page 2022

  4. [4]

    Nope-nerf: Optimising neu- ral radiance field with no pose prior

    Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neu- ral radiance field with no pose prior. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 2

    work page 2023

  5. [5]

    Past, present, and future of simultaneous localiza- tion and mapping: Toward the robust-perception age.IEEE Transactions on robotics, 32(6):1309–1332, 2016

    Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, Jos ´e Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localiza- tion and mapping: Toward the robust-perception age.IEEE Transactions on robotics, 32(6):1309–1332, 2016. 2

    2016

  6. [6]

    Neurbf: A neural fields repre- sentation with adaptive radial basis functions

    Zhang Chen, Zhong Li, Liangchen Song, Lele Chen, Jingyi Yu, Junsong Yuan, and Yi Xu. Neurbf: A neural fields repre- sentation with adaptive radial basis functions. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 4182–4194, 2023. 7

    work page 2023

  7. [7]

    Inverse depth parametrization for monocular slam.IEEE transactions on robotics, 24(5):932–945, 2008

    Javier Civera, Andrew J Davison, and JM Martinez Mon- tiel. Inverse depth parametrization for monocular slam.IEEE transactions on robotics, 24(5):932–945, 2008. 6

    work page 2008

  8. [8]

    Large-scale direct slam with stereo cameras

    Jakob Engel, J ¨org St¨uckler, and Daniel Cremers. Large-scale direct slam with stereo cameras. In2015 IEEE/RSJ interna- tional conference on intelligent robots and systems (IROS), pages 1935–1942. IEEE, 2015. 3

    work page 1935

Show all 46 references
  1. [9]

    A photometrically calibrated benchmark for monocular visual odometry.arXiv preprint arXiv:1607.02555, 2016

    Jakob Engel, Vladyslav Usenko, and Daniel Cremers. A photometrically calibrated benchmark for monocular visual odometry.arXiv preprint arXiv:1607.02555, 2016. 4

    2016 arXiv

  2. [10]

    Direct sparse odometry.IEEE transactions on pattern analysis and machine intelligence, 40(3):611–625, 2017

    Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry.IEEE transactions on pattern analysis and machine intelligence, 40(3):611–625, 2017. 2, 3

    2017

  3. [11]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2

    2022

  4. [12]

    Efros, and Xiaolong Wang

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splat- ting, 2024. 2

    2024

  5. [13]

    Multi-view stereo: A tutorial.Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1–148, 2015

    Yasutaka Furukawa, Carlos Hern ´andez, et al. Multi-view stereo: A tutorial.Foundations and Trends® in Computer Graphics and Vision, 9(1-2):1–148, 2015. 3

    2015

  6. [14]

    Adaptive posi- tional encoding for bundle-adjusting neural radiance fields

    Zelin Gao, Weichen Dai, and Yu Zhang. Adaptive posi- tional encoding for bundle-adjusting neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3284–3294, 2023. 1

    2023

  7. [15]

    Removing image artifacts due to dirty camera lenses and thin occluders

    Jinwei Gu, Ravi Ramamoorthi, Peter Belhumeur, and Shree Nayar. Removing image artifacts due to dirty camera lenses and thin occluders. InACM SIGGRAPH Asia 2009 papers, pages 1–10. 2009. 1

    2009

  8. [16]

    Stress birefringence analysis of polarization maintaining optical fibers.Optical Fiber Technology, 11(3): 240–254, 2005

    Rongfeng Guan, Fulong Zhu, Zhiyin Gan, Dexiu Huang, and Sheng Liu. Stress birefringence analysis of polarization maintaining optical fibers.Optical Fiber Technology, 11(3): 240–254, 2005. 2, 4

    2005

  9. [17]

    Plenoptic modeling and rendering from image sequences taken by a hand-held camera

    Benno Heigl, Reinhard Koch, Marc Pollefeys, Joachim Den- zler, and L Van Gool. Plenoptic modeling and rendering from image sequences taken by a hand-held camera. InMuster- erkennung 1999, pages 94–101. Springer, 1999. 2

    1999

  10. [18]

    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,

  11. [19]

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

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023. 7

    2023

  12. [20]

    3d gaussian splatting as markov chain monte carlo.arXiv preprint arXiv:2404.09591, 2024

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Wei- wei Sun, Jeff Tseng, Hossam Isack, Abhishek Kar, An- drea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo.arXiv preprint arXiv:2404.09591, 2024. 5, 7

    2024 arXiv

  13. [21]

    Robust radiometric cal- ibration and vignetting correction.IEEE transactions on pattern analysis and machine intelligence, 30(4):562–576,

    Seon Joo Kim and Marc Pollefeys. Robust radiometric cal- ibration and vignetting correction.IEEE transactions on pattern analysis and machine intelligence, 30(4):562–576,

  14. [22]

    Compact 3d gaussian representation for radiance field

    Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719– 21728, 2024. 1

    2024

  15. [23]

    Light field rendering

    Marc Levoy and Pat Hanrahan. Light field rendering. InPro- ceedings of the 23rd annual conference on Computer graph- ics and interactive techniques, pages 31–42, 1996. 2

    1996

  16. [24]

    All-in-one image restoration for unknown corruption

    Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17452– 17462, 2022. 3

    2022

  17. [25]

    Rt-nerf: Real-time on-device neural radiance fields to- wards immersive ar/vr rendering

    Chaojian Li, Sixu Li, Yang Zhao, Wenbo Zhu, and Yingyan Lin. Rt-nerf: Real-time on-device neural radiance fields to- wards immersive ar/vr rendering. InProceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design, pages 1–9, 2022. 1

    2022

  18. [26]

    Gp- nerf: Generalized perception nerf for context-aware 3d scene understanding

    Hao Li, Dingwen Zhang, Yalun Dai, Nian Liu, Lechao Cheng, Jingfeng Li, Jingdong Wang, and Junwei Han. Gp- nerf: Generalized perception nerf for context-aware 3d scene understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2170...

    2024

  19. [27]

    Let’s see clearly: Contaminant artifact removal for moving cam- eras

    Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V Sander. Let’s see clearly: Contaminant artifact removal for moving cam- eras. InProceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 2011–2020, 2021. 2, 3

    2011

  20. [28]

    Swinir: Image restoration us- ing swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration us- ing swin transformer. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 1833–1844,

  21. [29]

    Barf: Bundle-adjusting neural radiance fields

    Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Si- mon Lucey. Barf: Bundle-adjusting neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741–5751, 2021. 1, 2

    2021

  22. [30]

    Object recognition from local scale-invariant features

    David G Lowe. Object recognition from local scale-invariant features. InProceedings of the seventh IEEE interna- tional conference on computer vision, pages 1150–1157. Ieee, 1999. 2

    1999

  23. [31]

    Devignet: High- resolution vignetting removal via a dual aggregated fusion transformer with adaptive channel expansion

    Shenghong Luo, Xuhang Chen, Weiwen Chen, Zinuo Li, Shuqiang Wang, and Chi-Man Pun. Devignet: High- resolution vignetting removal via a dual aggregated fusion transformer with adaptive channel expansion. InProceed- ings of the AAAI Conference on Artificial Intelligence, pages 4...

    2024

  24. [32]

    Gaussian splatting slam

    Hidenobu Matsuki, Riku Murai, Paul HJ Kelly, and An- drew J Davison. Gaussian splatting slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 18039–18048, 2024. 2, 3

    2024

  25. [33]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    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. InEuropean conference on computer vision, pages 405–421. Springer, 2020. 1, 2, 6

    2020

  26. [34]

    Nerf in the dark: High dynamic range view synthesis from noisy raw images

    Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16190–16199, ...

    2022

  27. [35]

    Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM transactions on graphics (TOG), 41(4):1–15, 2022

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM transactions on graphics (TOG), 41(4):1–15, 2022. 2, 7

    2022

  28. [36]

    Orb-slam: a versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

    Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

  29. [37]

    Deep mesh reconstruction from single rgb images via topology modification networks

    Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 9964–9973, 2019. 2

    2019

  30. [38]

    D-nerf: Neural radiance fields for dynamic scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 1

    2021

  31. [39]

    Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild

    Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8931–8940, 2024. 1

    2024

  32. [40]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 4104–4113, 2016. 1, 2

    2016

  33. [41]

    Bad slam: Bundle adjusted direct rgb-d slam

    Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134–144, 2019. 3

    2019

  34. [42]

    Deep patch vi- sual odometry.Advances in Neural Information Processing Systems, 36, 2024

    Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch vi- sual odometry.Advances in Neural Information Processing Systems, 36, 2024. 1

    2024

  35. [43]

    Mip-splatting: Alias-free 3d gaussian splat- ting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splat- ting. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 19447–19456,

  36. [44]

    Nerf-editing: geometry editing of neural radiance fields

    Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: geometry editing of neural radiance fields. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 18353–18364, 2022. 1

    2022

  37. [45]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728–5739,

  38. [46]

    Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis

    Allan Zhou, Moo Jin Kim, Lirui Wang, Pete Florence, and Chelsea Finn. Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907–17917, 2023. 1

    2023

Pith tools

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