Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Plenodium: UnderWater 3D Scene Reconstruction with Plenoptic Medium Representation

T0 review · 4 major / 4 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 Modeling underwater scattering with a trilinearly interpolated spherical-harmonics grid, rather than a direction-only MLP, reconstructs degraded scenes more accurately and 47% faster.

desk verdict A useful SH-grid medium representation for underwater 3DGS, but the depth ranking loss is mathematically inverted and needs a re-derivation and re-ablation. read the letter →

arxiv 2505.21258 v1 pith:G4G5MVNF submitted 2025-05-27 cs.CV

Changguanng Wu , Jiangxin Dong , Chengjian Li , Jinhui Tang This is my paper · ORCID
classification cs.CV
keywords underwater3DreconstructionGaussiansplattingparticipatingmediasphericalharmonicsplenopticmediumrepresentationpseudo-depthcomplementationdepthrankinglossimagerestoration
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

Plenodium tries to establish that the water between camera and scene behaves as a spatially varying scattering medium, and that modeling it with explicit spherical harmonics on a 3D grid reconstructs underwater scenes more accurately and much faster than the direction-only MLP used by prior work. The proposed plenoptic medium representation evaluates the medium's color, attenuation, and backscatter as spherical harmonics of the ray direction, with coefficients trilinearly interpolated from grid vertices, so the medium depends on both where the ray comes from and where it is in space. Combined with a pseudo-depth initialization step that densifies COLMAP point clouds using monocular depth, and a depth-ranking loss that enforces ordinal consistency, the method reports a PSNR gain of at least 0.872 dB and a 47% rendering speedup over the best prior approach on real underwater scenes. If correct, this makes high-quality underwater reconstruction practical at interactive rates and yields restored, dehazed images as a byproduct of separating object radiance from backscatter.

What carries the argument

The load-bearing object is the plenoptic medium representation (Eqs. 6–7): the medium fields $c_{med}$, $\sigma_{att}$, and $\sigma_{bs}$ are each evaluated as spherical harmonics of the ray direction $d$, with the SH coefficient sets stored at the eight vertices of a 3D grid cell and combined by trilinear interpolation to give coefficients at any position $(x,y,z)$. This turns medium-parameter retrieval into a cheap lookup-plus-interpolation instead of an MLP forward pass, while making scattering depend on both direction and spatial position. The two supporting mechanisms are the pseudo-depth Gaussian complementation (Eqs. 9–10), which affinely calibrates monocular depth against COLMAP-initialized regions and inserts new Gaussian primitives in near, high-transmittance regions, and the depth ranking regularized loss (Eq. 12), which penalizes pairwise order violations between downsampled rendered and pseudo depths. Together they carry the argument that medium modeling and geometry initialization, not just photometric fitting, drive the reported gains.

What would settle it

One concrete test: on the paper's simulated dataset, where depth is known, compute the calibrated pseudo-depth of Eqs. 9–10, then measure how often the pairwise order in Eq. 12 agrees with ground truth in regions outside the COLMAP-initialized set; if the depth-ranking-trained model is not closer to true ordinal depth there than the same model without Eq. 12, the claimed geometric benefit of the pseudo-depth prior is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a scattering medium should be represented as a position- and direction-dependent field, and that an explicit spherical-harmonics grid is the right parameterization. Concretely, Plenodium stores SH coefficients for medium color $c_{med}$, object attenuation $\sigma_{att}$, and backscatter $\sigma_{bs}$ at the vertices of a 3D grid, interpolates them trilinearly to any query point, and evaluates them along the ray direction (Eqs. 6–7). This replaces the direction-only MLP medium used in SeaThru-NeRF and WaterSplatting and adds spatial awareness that the MLP lacks. On the four real SeaThru-NeRF scenes, the full system raises average PSNR from 29.600 to 30.472 dB (at least 0.872 dB over the best prior), improves SSIM, and renders at 249 FPS versus 169 FPS. The ablations attribute the gain to the SH encoding, the positional interpolation, the pseudo-depth Gaussian complementation, and the multi-scale SSIM plus depth-ranking losses.

Load-bearing premise

The load-bearing premise is that the pseudo-depth maps from Depth Anything, after the affine calibration of Eqs. 9–10, preserve the true relative ordering of depths in degraded underwater scenes, because that ordering is used both to decide where to insert new Gaussians and to supervise rendered depth through Eq. 12, yet the paper never validates it against ground truth on the real scenes.

Editorial extensions

If this is right

  • Underwater 3D reconstruction and restoration can run at interactive frame rates (249 FPS on a consumer GPU) while improving PSNR by about 0.87 dB over the best prior method.
  • The explicit SH-grid medium is not tied to a particular scene geometry, so it can be added to other splatting pipelines that need scattering media such as fog, smoke, or turbid water.
  • Because medium parameters are accessible per voxel, backscatter can be subtracted explicitly, yielding restored, dehazed images alongside the reconstructed geometry.
  • The pseudo-depth complementation offers a practical remedy for structure-from-motion failure in low-visibility scenes, a bottleneck for all 3D Gaussian splatting methods in underwater environments.
  • The depth ranking loss provides scale-invariant geometric supervision from monocular priors, reducing the method's dependence on exact depth values while improving ordinal stability.

Reading between the lines

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

  • Inference: the SH-grid medium should generalize to other participating media such as smoke, fog, and estuaries, since the paper's mechanism treats scattering as a position- and direction-dependent volumetric field rather than a water-specific phenomenon.
  • Inference: the measured speedup comes from replacing MLP evaluation with SH lookup plus trilinear interpolation, so similar explicit encodings could accelerate other per-ray MLP fields in inverse rendering beyond underwater scenes.
  • Inference: a controlled comparison against a position-and-direction MLP of matched capacity would separate the contribution of positional information from the inductive bias of the SH grid; the paper's ablations compare SH against direction-only MLP, not against an equally expressive position-aware MLP.
  • Inference: because the ranking loss is applied to downsampled depth and never checked against true depth on real scenes, the robustness claim would be strengthened by a synthetic experiment in which pseudo-depth is deliberately corrupted to test how much ordinal error the pipeline tolerates.
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. The paper proposes Plenodium, a 3D Gaussian Splatting-based framework for underwater scene reconstruction that jointly models scene objects and a participating medium. Its main contributions are: (1) an explicit plenoptic medium representation in which the medium color and attenuation/backscatter coefficients are encoded with spherical harmonics evaluated from a trilinearly interpolated 3D coefficient grid, replacing the MLP used by prior work; (2) a pseudo-depth Gaussian complementation (PDGC) that augments COLMAP point clouds with Depth Anything pseudo-depth maps to improve initialization; and (3) a depth-ranking regularized loss intended to enforce ordinal consistency between rendered and pseudo-depths. The method is evaluated on the real SeaThru-NeRF dataset and on a new simulated underwater dataset with controllable scattering, reporting improvements in PSNR/SSIM over WaterSplatting and a 47% rendering speedup.

Significance. If the results hold, the explicit SH-grid medium representation is a practically useful contribution: it removes the MLP inference bottleneck of prior medium models while adding spatial variation, and the reported speed and quality gains on the SeaThru-NeRF benchmark are meaningful. The release of code and a simulated dataset is also a strength. However, the central claim about the depth-ranking loss is undermined by a sign error in Eq. (12), and the PDGC component lacks sufficient specification for reproducibility. The paper is within the scope of the venue and the core idea is worth pursuing, but the current manuscript requires substantive revision before the claims can be accepted.

major comments (4)
  1. [Section 4.3, Eq. (12)] The depth-ranking regularized loss as written does not penalize ordinal violations. For a pair with pseudo-depth difference a = D~_i - D~_j and rendered-depth difference b = D̂_i - D̂_j, the term min(-(a·b), 0) equals 0 when a and b have opposite signs (a violation) and equals a negative value when they have the same sign (an agreement). Thus the loss is zero for the pairs it claims to penalize and is instead an unbounded-below reward that pushes already-correct orderings further apart. It also gives zero gradient for violating pairs. The text states that this loss "penalizes violations of ordinal relationships" and Table 6 credits it with a PSNR improvement, so this is an internal inconsistency. Please correct the sign (e.g., use max(-(a·b), 0)), re-derive the gradient, and re-run the ablation in Table 6.
  2. [Section 4.2, Eqs. (9)-(10) and insertion rule] The pseudo-depth Gaussian complementation does not specify which depth map is used to position the newly inserted Gaussians. The insertion domain Ω_n is defined using the uncalibrated pseudo-depth D~, while Eq. (10) calibrates D~ into D~' using least squares. The paper states that D~' "exhibits significantly mitigated scale ambiguity and reduced offset biases" but never says whether the inserted primitives are positioned with D~, D~', or the rendered depth D̂. Since this is one of the three main contributions, please specify exactly how the 3D positions of the new Gaussians are computed from the (calibrated or uncalibrated) pseudo-depth, and state which depth map enters the Ω_n/Ω_p conditions.
  3. [Section 3, Eqs. (3)-(5)] The notation for the medium extinction coefficient is inconsistent. Eq. (3) defines T_med(z) = e^{-σ_med z}, Eq. (4) integrates with σ_med, and Eq. (5) suddenly uses σ_bs in the medium backscatter terms and σ_att in the object term, without ever relating σ_med to σ_bs. Since Eq. (5) is the core rendering equation, please define all coefficients explicitly and unify the notation (e.g., state that σ_med = σ_bs for the backscatter path, or replace σ_med with σ_bs in Eqs. (3)-(4)). This is necessary for reproducibility and for readers to verify the derivation.
  4. [Section 1 and Section 5.2] The claimed average SSIM improvement of 0.047 over the best-competing method is not supported by Table 1. Computing the per-scene SSIM differences between Plenodium and WaterSplatting gives (0.006, 0.005, 0.004, 0.004), averaging to 0.00475, not 0.047. Please correct the number or clarify which baseline and scenes are included in the average; as written, this headline claim is inaccurate.
minor comments (4)
  1. [Section 4.1, Eq. (6)-(7)] The paper stores "eight fundamental SH coefficients" at each grid vertex, but later states that the maximum SH degree is 3, which requires 16 coefficients per scalar channel. Please clarify the exact number of stored coefficients per vertex and per channel, and reconcile this with the stated SH degree.
  2. [Caption of Figure 3] The caption states "The pseudo-depth for ground truth is estimated using the Depth Any Model for reference purposes," which is confusing because ground truth does not need pseudo-depth estimation. Please clarify what is being shown (e.g., the pseudo-depth of the GT image as an external reference).
  3. [Section 5.1] The phrase "The patch number N is set to 16" is ambiguous: it likely means the downsampled depth maps in Eq. (12) have size 16×16, but "patch number" is not standard. Please state the resolution explicitly and define N consistently in Eq. (12).
  4. [Tables 2 and 5] Table 2 has a misaligned header (the scene names and metric labels are interleaved), and Table 5 contains a typo "COLAMP" for "COLMAP." Please fix these formatting issues.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are evaluated on held-out views against external baselines, and no load-bearing step reduces to the paper's own inputs.

full rationale

The derivation chain is self-contained and externally benchmarked. The medium field (cmed, sigma_att, sigma_bs) is a per-scene free parameter optimized against training views via Eq. 5 and evaluated on held-out views; the headline comparison (Tab. 1, Fig. 1) is against external baselines (WaterSplatting, SeaThru-NeRF, 3DGS, Zip-NeRF) using standard PSNR/SSIM/LPIPS on test images, so the central claim is not a restatement of an input. The pseudo-depth prior comes from the external Depth Anything model, not trained by this paper; using it both to initialize Gaussians and to provide ordinal depth supervision is a design coupling, not a definitional equivalence, because the final outputs are photometric reconstructions scored on held-out views and on a simulated dataset with ground truth. No load-bearing step reduces by construction to a fitted parameter or to a self-citation: references [1,2,40] supply the physical scattering model externally, and no uniqueness theorem or self-authored result is invoked to force the representation choice. One non-circular concern: Eq. 12, as written, is not an ordinal-violation penalty, since min(-ab,0) is negative for agreeing pairs and zero for violating pairs; this is a correctness/reproducibility bug, but it is not circularity because it does not make a claimed prediction equal to the paper's input.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

No new physical entities are postulated; the SH medium grid is a parameterization of an existing physical model. The main unstated ingredients are the strength of the smoothness assumption and the implicit trust in Depth Anything priors.

free parameters (8)
  • lambda_L1 = 0.8
    Empirically set loss weight in Eq. 13 (Sec. 5.1), tuned for the reported results.
  • lambda_ssim = 0.2
    Empirically set loss weight in Eq. 13; the choice of multi-scale SSIM over single-scale SSIM is validated only by ablations.
  • lambda_depth = 5
    Empirically set depth ranking loss weight; a 5x weight relative to photometric losses is chosen by hand.
  • tau_w = 0.99
    Threshold for well-initialized regions in pseudo-depth affine calibration and insertion regions (Sec. 5.1).
  • tau_near = 0.5
    Near-camera threshold for pseudo-depth Gaussian insertion (Sec. 5.1).
  • patch_size_N = 16
    Downsampled depth map size for the depth ranking loss (Sec. 5.1).
  • SH_degree = 3
    Maximum SH degree for medium and Gaussian primitives; a representational capacity knob not derived from data.
  • grid_resolution = not specified
    Resolution of the 3D spatial grid for SH coefficients is not reported in the paper, leaving an important hyperparameter unspecified.
assumptions (4)
  • domain assumption Image formation follows Eq. 2: color is a linear combination of attenuated object radiance and backscattered medium color with exponential depth factors.
    Adopted from Akkaynak-Treibitz [40] and used in Eq. 5; if this model is wrong, all medium-reconstruction methods in the comparison inherit the error.
  • ad hoc to paper The medium is smooth enough to be represented by trilinearly interpolated SH coefficients on a coarse 3D grid with degree 3.
    Sec. 4.1 introduces the grid-based SH representation without reporting grid resolution or validating smoothness; this is the key representational assumption of the proposed novelty.
  • domain assumption COLMAP provides valid camera poses and rough geometry in underwater scenes.
    Sec. 4.2 assumes COLMAP initialization is usable; the paper proposes to complement, not replace, it.
  • domain assumption Depth Anything pseudo-depth has correct ordinal structure in underwater scenes after affine calibration.
    Sec. 4.2-4.3 relies on pseudo-depth for both initialization and the depth ranking loss, but no quantitative validation of pseudo-depth on underwater data is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Plenodium: UnderWater 3D Scene Reconstruction with Plenoptic Medium Representation." pith.science (2026). https://pith.science/paper/G4G5MVNF

@misc{pith2026250521258,
  author       = {Pith},
  title        = {Pith review of: Plenodium: UnderWater 3D Scene Reconstruction with Plenoptic Medium Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G4G5MVNF}},
  note         = {Machine review of arXiv:2505.21258}
}
read the original abstract

We present Plenodium (plenoptic medium), an effective and efficient 3D representation framework capable of jointly modeling both objects and participating media. In contrast to existing medium representations that rely solely on view-dependent modeling, our novel plenoptic medium representation incorporates both directional and positional information through spherical harmonics encoding, enabling highly accurate underwater scene reconstruction. To address the initialization challenge in degraded underwater environments, we propose the pseudo-depth Gaussian complementation to augment COLMAP-derived point clouds with robust depth priors. In addition, a depth ranking regularized loss is developed to optimize the geometry of the scene and improve the ordinal consistency of the depth maps. Extensive experiments on real-world underwater datasets demonstrate that our method achieves significant improvements in 3D reconstruction. Furthermore, we conduct a simulated dataset with ground truth and the controllable scattering medium to demonstrate the restoration capability of our method in underwater scenarios. Our code and dataset are available at https://plenodium.github.io/.

Figures

Figures reproduced from arXiv: 2505.21258 by the authors.

Figure 1
Figure 1. Comparison of Plenodium, Wa￾terSplatting [1], and SeaThru-NeRF [2] on reconstruction and restoration perfor￾mance (PSRN, SSIM), as well as effi￾ciency (FPS, training time). Underwater imaging plays a critical role in diverse ap￾plications, including underwater construction, marine sciences, etc. However, its efficacy is significantly hin￾dered by the complex optical properties of the aquatic environment. These prope… view at source ↗
Figure 2
Figure 2. Overview of our Plenodium. We first employ the pseudo-depth Gaussian complementation to enrich the primitives initialized by COLMAP. Then we utilize the plenoptic medium representation to estimate the medium parameter and render the underwater images following Eqn. 5. Our network is optimized with the loss function in Eqn. 13, including a new depth ranking regularized loss. to accurately model the scattering medium … view at source ↗
Figure 3
Figure 3. Rendering performance comparison of our Plenodium against existing methods on the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Restoration performance comparison of our Plenodium against existing methods on the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Restoration performance comparison of our Plenodium against existing methods on “Beach” [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Efficiency comparison of our plenoptic medium representation against MLP-based rep￾resentation in forward and backward. step = 1000 step = 2000 step = 15000 w/o PDGC w/ PDGC [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Swimm3R: Splatting with Medium-aware SfM for Underwater 3D Reconstruction

    cs.CV 2026-08 conditional novelty 6.0 of 10

    Swimm3R couples a scattering-aware, feed-forward structure-from-motion backbone with underwater Beta splatting to reconstruct and render 3D scenes from turbid underwater video, improving rendering PSNR and localizatio...

Reference graph

Works this paper leans on

50 extracted references · 41 canonical work pages · cited by 1 Pith paper

  1. [1]

    WaterSplatting: Fast underwater 3D scene reconstruction using gaussian splatting,

    H. Li, W. Song, T. Xu, A. Elsig, and J. Kulhanek, “WaterSplatting: Fast underwater 3D scene reconstruction using gaussian splatting,”International Conference on 3D Vision, 2025

    work page 2025

  2. [2]

    Seathru- nerf: Neural radiance fields in scattering media,

    D. Levy, A. Peleg, N. Pearl, D. Rosenbaum, D. Akkaynak, S. Korman, and T. Treibitz, “Seathru- nerf: Neural radiance fields in scattering media,” inIEEE Conference on Computer Vision and Pattern Recognition, 2023, pp. 56–65

    work page 2023

  3. [3]

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

    J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, and P. P. Srinivasan, “Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields,” inInternational Conference on Computer Vision, 2021, pp. 5855–5864

    work page 2021

  4. [4]

    Zip-nerf: Anti-aliased grid-based neural radiance fields,

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Zip-nerf: Anti-aliased grid-based neural radiance fields,” inInternational Conference on Computer Vision, 2023, pp. 19 697–19 705

    work page 2023

  5. [5]

    Mini-splatting: Representing scenes with a constrained number of gaussians,

    G. Fang and B. Wang, “Mini-splatting: Representing scenes with a constrained number of gaussians,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 165–181

    2024

  6. [6]

    Taming 3dgs: High-quality radiance fields with limited resources,

    S. S. Mallick, R. Goel, B. Kerbl, M. Steinberger, F. V . Carrasco, and F. De La Torre, “Taming 3dgs: High-quality radiance fields with limited resources,” inACM SIGGRAPH Asia, 2024, pp. 1–11

    work page 2024

  7. [7]

    Instant neural graphics primitives with a multiresolution hash encoding,

    T. Müller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,”ACM Transactions on Graphics, vol. 41, no. 4, pp. 102:1– 102:15, Jul. 2022

    work page 2022

  8. [8]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

Show all 50 references
  1. [9]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering.”ACM Transactions on Graphics, vol. 42, no. 4, pp. 139–1, 2023

    2023

  2. [10]

    Structure-from-motion revisited,

    J. L. Schönberger and J.-M. Frahm, “Structure-from-motion revisited,” inIEEE Conference on Computer Vision and Pattern Recognition, 2016

    2016

  3. [11]

    Pixelwise view selection for unstructured multi-view stereo,

    J. L. Schönberger, E. Zheng, M. Pollefeys, and J.-M. Frahm, “Pixelwise view selection for unstructured multi-view stereo,” inEuropean Conference on Computer Vision, 2016

    2016

  4. [12]

    Depth anything: Unleashing the power of large-scale unlabeled data,

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth anything: Unleashing the power of large-scale unlabeled data,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 10 371–10 381

    2024

  5. [13]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,” Conference on Neural Information Processing Systems, vol. 37, pp. 21 875–21 911, 2025

    2025

  6. [14]

    Plenoxels: Ra- diance fields without neural networks,

    S. Fridovich-Keil, A. Yu, M. Tancik, Q. Chen, B. Recht, and A. Kanazawa, “Plenoxels: Ra- diance fields without neural networks,” inIEEE Conference on Computer Vision and Pattern Recognition, 2022, pp. 5501–5510

    2022

  7. [15]

    Distwar: Fast differentiable rendering on raster-based rendering pipelines,

    S. Durvasula, A. Zhao, F. Chen, R. Liang, P. K. Sanjaya, and N. Vijaykumar, “Distwar: Fast differentiable rendering on raster-based rendering pipelines,”arXiv preprint arXiv:2401.05345, 2023

    2023 arXiv

  8. [16]

    gsplat: An open-source library for gaussian splatting,

    V . Ye, R. Li, J. Kerr, M. Turkulainen, B. Yi, Z. Pan, O. Seiskari, J. Ye, J. Hu, M. Tancik, and A. Kanazawa, “gsplat: An open-source library for gaussian splatting,”Journal of Machine Learning Research, vol. 26, no. 34, pp. 1–17, 2025

    2025

  9. [17]

    Hugs: Human gaussian splats,

    M. Kocabas, J.-H. R. Chang, J. Gabriel, O. Tuzel, and A. Ranjan, “Hugs: Human gaussian splats,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 505–515. 10

    2024

  10. [18]

    Gaussiana- vatars: Photorealistic head avatars with rigged 3d gaussians,

    S. Qian, T. Kirschstein, L. Schoneveld, D. Davoli, S. Giebenhain, and M. Nießner, “Gaussiana- vatars: Photorealistic head avatars with rigged 3d gaussians,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 299–20 309

    2024

  11. [19]

    Gaussian- dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models,

    T. Yi, J. Fang, J. Wang, G. Wu, L. Xie, X. Zhang, W. Liu, Q. Tian, and X. Wang, “Gaussian- dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models,” in IEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 6796–6807

    2024

  12. [20]

    Text-to-3d using gaussian splatting,

    Z. Chen, F. Wang, Y . Wang, and H. Liu, “Text-to-3d using gaussian splatting,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 401–21 412

    2024

  13. [21]

    Re- paint123: Fast and high-quality one image to 3d generation with progressive controllable repainting,

    J. Zhang, Z. Tang, Y . Pang, X. Cheng, P. Jin, Y . Wei, X. Zhou, M. Ning, and L. Yuan, “Re- paint123: Fast and high-quality one image to 3d generation with progressive controllable repainting,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 303–320

    2024

  14. [22]

    Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes,

    X. Zhou, Z. Lin, X. Shan, Y . Wang, D. Sun, and M.-H. Yang, “Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 634–21 643

    2024

  15. [23]

    Sgs-slam: Semantic gaussian splatting for neural dense slam,

    M. Li, S. Liu, H. Zhou, G. Zhu, N. Cheng, T. Deng, and H. Wang, “Sgs-slam: Semantic gaussian splatting for neural dense slam,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 163–179

    2024

  16. [24]

    S2gaussian: Sparse-view super-resolution 3d gaussian splatting,

    Y . Wan, Y . Cheng, M. Shao, and W. Zuo, “S2gaussian: Sparse-view super-resolution 3d gaussian splatting,”IEEE Conference on Computer Vision and Pattern Recognition, 2025

    2025

  17. [25]

    Bad-gaussians: Bundle adjusted deblur gaussian splatting,

    L. Zhao, P. Wang, and P. Liu, “Bad-gaussians: Bundle adjusted deblur gaussian splatting,” in European Conference on Computer Vision. Springer, 2024, pp. 233–250

    2024

  18. [26]

    Gaussian splatting on the move: Blur and rolling shutter compensation for natural camera motion,

    O. Seiskari, J. Ylilammi, V . Kaatrasalo, P. Rantalankila, M. Turkulainen, J. Kannala, E. Rahtu, and A. Solin, “Gaussian splatting on the move: Blur and rolling shutter compensation for natural camera motion,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 160–177

    2024

  19. [27]

    Deraings: Gaussian splatting for enhanced scene reconstruction in rainy environments,

    S. Liu, X. Chen, H. Chen, Q. Xu, and M. Li, “Deraings: Gaussian splatting for enhanced scene reconstruction in rainy environments,” inAssociation for the Advancement of Artificial Intelligence, vol. 39, no. 5, 2025, pp. 5558–5566

    2025

  20. [28]

    Lighting every darkness with 3dgs: Fast training and real-time rendering for hdr view synthesis,

    X. Jin, P. Jiao, Z.-P. Duan, X. Yang, C. Li, C.-L. Guo, and B. Ren, “Lighting every darkness with 3dgs: Fast training and real-time rendering for hdr view synthesis,”Conference on Neural Information Processing Systems, vol. 37, pp. 80 191–80 219, 2025

    2025

  21. [29]

    Neural underwater scene representation,

    Y . Tang, C. Zhu, R. Wan, C. Xu, and B. Shi, “Neural underwater scene representation,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 11 780–11 789

    2024

  22. [30]

    Deblur-nerf: Neural radiance fields from blurry images,

    L. Ma, X. Li, J. Liao, Q. Zhang, X. Wang, J. Wang, and P. V . Sander, “Deblur-nerf: Neural radiance fields from blurry images,” inIEEE Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 861–12 870

    2022

  23. [31]

    Hdr-nerf: High dynamic range neural radiance fields,

    X. Huang, Q. Zhang, Y . Feng, H. Li, X. Wang, and Q. Wang, “Hdr-nerf: High dynamic range neural radiance fields,” inIEEE Conference on Computer Vision and Pattern Recognition, 2022, pp. 18 398–18 408

    2022

  24. [32]

    Nan: Noise-aware nerfs for burst-denoising,

    N. Pearl, T. Treibitz, and S. Korman, “Nan: Noise-aware nerfs for burst-denoising,” inIEEE Conference on Computer Vision and Pattern Recognition, 2022, pp. 12 672–12 681

    2022

  25. [33]

    Nerf-sr: High quality neural radiance fields using supersampling,

    C. Wang, X. Wu, Y .-C. Guo, S.-H. Zhang, Y .-W. Tai, and S.-M. Hu, “Nerf-sr: High quality neural radiance fields using supersampling,” inACM International Conference on Multimedia, 2022, pp. 6445–6454

    2022

  26. [34]

    Underwater single image color restoration using haze-lines and a new quantitative dataset,

    D. Berman, D. Levy, S. Avidan, and T. Treibitz, “Underwater single image color restoration using haze-lines and a new quantitative dataset,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 8, pp. 2822–2837, 2020. 11

    2020

  27. [35]

    Generalization of the dark channel prior for single image restoration,

    Y .-T. Peng, K. Cao, and P. C. Cosman, “Generalization of the dark channel prior for single image restoration,”IEEE Transactions on Image Processing, vol. 27, no. 6, pp. 2856–2868, 2018

    2018

  28. [36]

    Underwater image restoration based on image blurriness and light absorption,

    Y .-T. Peng and P. C. Cosman, “Underwater image restoration based on image blurriness and light absorption,”IEEE Transactions on Image Processing, vol. 26, no. 4, pp. 1579–1594, 2017

    2017

  29. [37]

    Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration,

    C. Zhao, W. Cai, C. Dong, and C. Hu, “Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 8281–8291

    2024

  30. [38]

    Uveb: A large-scale bench- mark and baseline towards real-world underwater video enhancement,

    Y . Xie, L. Kong, K. Chen, Z. Zheng, X. Yu, Z. Yu, and B. Zheng, “Uveb: A large-scale bench- mark and baseline towards real-world underwater video enhancement,” inIEEE Conference on Computer Vision and Pattern Recognition, 2024, pp. 22 358–22 367

    2024

  31. [39]

    An underwater image enhancement benchmark dataset and beyond,

    C. Li, C. Guo, W. Ren, R. Cong, J. Hou, S. Kwong, and D. Tao, “An underwater image enhancement benchmark dataset and beyond,”IEEE transactions on image processing, vol. 29, pp. 4376–4389, 2019

    2019

  32. [40]

    A revised underwater image formation model,

    D. Akkaynak and T. Treibitz, “A revised underwater image formation model,” inIEEE Confer- ence on Computer Vision and Pattern Recognition, 2018, pp. 6723–6732

    2018

  33. [41]

    Osmosis: Rgbd diffusion prior for underwater image restoration,

    O. B. Nathan, D. Levy, T. Treibitz, and D. Rosenbaum, “Osmosis: Rgbd diffusion prior for underwater image restoration,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 302–319

    2024

  34. [42]

    Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time,

    H. Zeng, J. Cai, L. Li, Z. Cao, and L. Zhang, “Learning image-adaptive 3d lookup tables for high performance photo enhancement in real-time,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 04, pp. 2058–2073, 2022

    2022

  35. [43]

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

    B. Mildenhall, P. Hedman, R. Martin-Brualla, P. P. Srinivasan, and J. T. Barron, “Nerf in the dark: High dynamic range view synthesis from noisy raw images,” inIEEE Conference on Computer Vision and Pattern Recognition, 2022, pp. 16 190–16 199

    2022

  36. [44]

    Multiscale structural similarity for image quality assessment,

    Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multiscale structural similarity for image quality assessment,” inThe Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, vol. 2. Ieee, 2003, pp. 1398–1402

    2003

  37. [45]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  38. [46]

    Structure-guided ranking loss for single image depth prediction,

    K. Xian, J. Zhang, O. Wang, L. Mai, Z. Lin, and Z. Cao, “Structure-guided ranking loss for single image depth prediction,” inIEEE Conference on Computer Vision and Pattern Recognition, 2020, pp. 611–620

    2020

  39. [47]

    Sparsenerf: Distilling depth ranking for few-shot novel view synthesis,

    G. Wang, Z. Chen, C. C. Loy, and Z. Liu, “Sparsenerf: Distilling depth ranking for few-shot novel view synthesis,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 9065–9076

    2023

  40. [48]

    Nerfstudio: A modular framework for neural radiance field development,

    M. Tancik, E. Weber, E. Ng, R. Li, B. Yi, J. Kerr, T. Wang, A. Kristoffersen, J. Austin, K. Salahi, A. Ahuja, D. McAllister, and A. Kanazawa, “Nerfstudio: A modular framework for neural radiance field development,” inACM SIGGRAPH, 2023

    2023

  41. [49]

    Absgs: Recovering fine details in 3d gaussian splatting,

    Z. Ye, W. Li, S. Liu, P. Qiao, and Y . Dou, “Absgs: Recovering fine details in 3d gaussian splatting,” inACM International Conference on Multimedia, 2024, pp. 1053–1061

    2024

  42. [50]

    Instant dehazing of images using polarization,

    Y . Y . Schechner, S. G. Narasimhan, and S. K. Nayar, “Instant dehazing of images using polarization,” inProceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, vol. 1. IEEE, 2001, pp. I–I. 12

    2001

Pith tools

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