Pith. sign in

REVIEW 4 major objections 5 minor 45 references

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

T0 review · 4 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 Swimm3R claims that scattering-degraded underwater video can be reconstructed into dense, restored 3D maps by distilling in-air geometric priors into a medium-aware SfM backbone and rendering with Beta splatting.

desk verdict Credible, well-ablated adaptation of feed-forward SfM to underwater imaging; the headline PSNR gains rest on an unreleased dataset with test-time pose refinement, so treat the magnitudes with caution. read the letter →

arxiv 2608.00950 v1 pith:FX4BNCDL submitted 2026-08-02 cs.CV cs.ROeess.IV

Minseong Kweon , Junaed Sattar This is my paper
classification cs.CVcs.ROeess.IV
keywords underwater3Dreconstructionstructure-from-motionGaussiansplattingBetascatteringmediumvisuallocalizationimagerestorationdataset
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

Swimm3R tries to establish that underwater 3D reconstruction fails not because geometry is absent, but because scattering and wavelength-dependent attenuation break the correspondence and color cues that standard structure-from-motion relies on. The paper's solution is to take a feed-forward SfM model pretrained on in-air images, fine-tune it on underwater images with a physics-based image-formation head, and use the resulting poses and restored dense point clouds to initialize a splatting renderer. The renderer uses Beta kernels and an explicit medium field so backscatter is modeled separately from scene surfaces. If the approach is right, turbid, low-texture underwater video can produce dense maps, accurate camera trajectories, and renderable scenes without per-scene calibration, and those maps directly improve visual localization. The claimed evidence includes better depth and pose accuracy on the FLSea test sequences and a 1.47 dB PSNR gain over WaterSplatting on a new Barbados dataset.

What carries the argument

Two coupled pieces carry the argument. First, medium-aware SfM: LoRA fine-tuning of MASt3R-SfM with feature-map distillation plus an underwater physics head trained with forward, inverse, and saturation losses; the head regresses $\beta_D$, $\beta_B$, and $B_\infty$ and implements $I = J\odot T_D + B_\infty\odot(1-T_B)$ with $T_D=e^{-\beta_D z}$, $T_B=e^{-\beta_B z}$. Second, Underwater Beta Splatting: Beta primitives, an explicit medium field $(\beta_B,\beta_D,B_\infty)(u)=\sigma_{\mathrm{med}}[A_\phi(\mathrm{SH}(d(u)))]$, and Scattering-aware Geometric Gradients, a gradient preconditioner that splits pixel-primitive adjoints into geometry and medium branches and rescales center steps for m

What would settle it

Run Swimm3R, MASt3R-SfM, and WaterSplatting on an underwater scene with known metric poses and depth—for example a tank with calibrated cameras and synthetic scattering or a structured-light scan of a seafloor mock-up—and compare absolute pose error and depth error, with and without the 300-step test-pose refinement. If Swimm3R's advantage disappears under metric ground truth or shrinks when refinement is disabled, the central claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that underwater SfM can be made medium-aware by distilling in-air geometric priors into a feed-forward backbone and coupling it with a splatting renderer that explicitly models the water column. Swimm3R distills the encoder, decoder, and descriptor features of MASt3R-SfM into an underwater student via LoRA, and appends a physics head that predicts per-pixel direct attenuation $\beta_D$, backscatter $\beta_B$, and veiling light $B_\infty$ from raw frames. The head implements the forward formation model $I = J \odot e^{-\beta_D z} + B_\infty \odot (1 - e^{-\beta_B z})$ and its inverse restoration, so the student recovers clean radiance and geometry in one pass; at inferenc

Load-bearing premise

The load-bearing premise is that the Barbados evaluation numbers are trustworthy even though there is no metric ground truth for poses: the reference poses come from GLUEMAP, test poses are refined for 300 steps against the held-out image, and the reported gains could be optimistic if either the reference poses are wrong or the refinement protocol inflates render and localization scores.

Editorial extensions

If this is right

  • Underwater robots could build dense, visually restored maps directly from raw video, without needing reliable sparse keypoints or prior calibration, which matters in turbid and low-texture environments.
  • Any downstream splatting renderer inherits a better initialization from Swimm3R's point clouds; the paper reports that replacing GLUEMAP or MASt3R-SfM initialization with Swimm3R improves rendering across renderers.
  • Explicitly separating backscatter and attenuation from surface radiance should make rendered maps stable under changing water conditions, supporting query-to-render matching for localization.
  • Because the SfM head also outputs optical parameters, the same framework can output scene structure and water-medium estimates simultaneously, useful for monitoring and navigation.

Reading between the lines

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

  • The reported PSNR is measured after fitting each test camera pose to its held-out image, so it is a rendering-after-pose-fitting score, not pure novel-view synthesis; an evaluation without pose refinement would likely lower absolute numbers and could change the ranking.
  • The distillation recipe is not limited to water: any scattering or attenuating medium with a known forward model (fog, smoke, low light) could use the same LoRA-plus-physics-head transfer, provided paired clean-degraded data exist.
  • The GLUEMAP-anchored pose metrics on Barbados compare relative map quality; a controlled dataset with metric ground truth would be needed to know absolute accuracy.
  • The paper's water-type augmentation samples four veiling-light categories; testing on a fifth unseen water type would probe whether the physics head generalizes or merely memorizes the training palette.
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 / 5 minor

Summary. The paper proposes Swimm3R, a unified underwater 3D reconstruction framework that combines a LoRA-fine-tuned MASt3R-SfM backbone with an underwater physics head (predicting attenuation, backscatter, and veiling light) and an Underwater Beta Splatting (UWBS) renderer using Beta kernels, an explicit medium field, and scattering-aware gradient preconditioning. A new four-scene Barbados GoPro dataset is introduced. On the FLSea test sequences, Swimm3R reports improved depth (AbsRel 0.2132 vs. 0.2247 for MASt3R-SfM) and pose accuracy; on Barbados, it reports denser point clouds, better test-frame rendering PSNR (1.47 dB over WaterSplatting under Swimm3R initialization), and improved downstream localization (RRA@15 and RTA@15 gains of 2.0 and 2.4 points). The central claim is that medium-aware SfM plus Beta-based splatting improves underwater geometry, rendering, and localization in scattering conditions.

Significance. If the results hold, the paper addresses a real limitation of underwater 3D reconstruction: the fragility of SfM poses and sparse points under scattering and attenuation. Distilling in-air geometric priors through a physics-based image-formation head is a timely idea, and the external FLSea evaluation provides a non-circular anchor for the SfM component. The explicit underwater medium field, Beta kernels, and SGG preconditioner are inventive, and the ablations are internally consistent. The paper is also admirably candid about the absence of metric ground-truth poses on Barbados and about the comparative nature of the pose numbers. However, the headline Barbados rendering result is measured under a test-time pose-refinement protocol, and the SfM training objective is underspecified with respect to the depth variable, so the empirical support is not yet fully load-bearing.

major comments (4)
  1. [Sec. IV-E, Tab. IV] The reported 1.47 dB PSNR gain over WaterSplatting is obtained after refining each test camera pose for 300 steps against its own held-out image while the map and medium field are fixed. This measures pose-fitting plus rendering, not pure novel-view synthesis, and can differentially favor a renderer whose Beta kernels interact favorably with pose optimization. Since this is the headline rendering claim on a new dataset, please also report PSNR/SSIM/LPIPS without test-pose refinement (or with a fixed common pose), and discuss whether the relative ordering persists. The localization claim in Tab. V is not affected by this issue, but the rendering claim is currently tied to this protocol.
  2. [Sec. III-B, Eqs. (2)-(6)] The physics-head forward and inverse losses require a metric ray depth z_n (Eq. 2), but the manuscript never specifies where z_n comes from during SfM training. If z_n is the student's own predicted depth, then the inverse restoration in Eq. (4) and the forward loss in Eq. (5) are trained with self-generated depth, which is a circularity concern and should be stated. If z_n is obtained from the teacher pointmap, stereo, known geometry, or another source, that source and the gradient flow need to be described. This is load-bearing for interpreting the SfM ablations in Tab. VI.
  3. [Sec. IV-B, Tab. III] The Barbados 'SfM map quality' metrics (epipolar error and 1-px inlier ratio) are computed under each method's own poses, as stated in Sec. IV-B. They are internal-consistency measures, not accuracy relative to a shared reference. With no metric ground-truth geometry on Barbados, the conclusion that Swimm3R yields 'more continuous and geometrically coherent scene structures' is supported only by self-consistent numbers and qualitative figures. Please make this caveat prominent in Sec. IV-D and in the Tab. III caption, and avoid implying that lower self-epipolar error directly translates to metric accuracy.
  4. [Sec. IV-B, Tab. V, Sec. V] On Barbados, GLUEMAP poses serve as the reference for RRA@15 and RTA@15. The paper acknowledges in Sec. V that no metric ground-truth poses are available, but the main-text sentence in Sec. IV-B and the Tab. V caption should explicitly state that all Barbados pose numbers are relative to GLUEMAP and therefore comparative rather than absolute. This is not fatal, but it should be made equally prominent at the point of use, not only in the limitations paragraph.
minor comments (5)
  1. [Sec. III-C, Eq. (10)] The ambient convolutional field A_phi is described as a 1x1 CNN with ReLU hidden activations, but its input/output dimensionalities and the SH degree used are not specified. Please clarify the architecture and how B_inf, beta_B, and beta_D are split from its output.
  2. [Eq. (4)] The notation [tilde J]_0^1 for clipping to [0,1] is used without definition. Please define it explicitly in the text or with a comment after the equation.
  3. [Tab. IV caption] The caption says 'test-frame rendering quality' but the protocol includes 300-step per-test-pose refinement. Please state 'with test-pose refinement' in the caption or table title so that readers are not misled into interpreting the numbers as pure held-out novel-view synthesis.
  4. [Sec. IV-B] The FLSea pose evaluation uses COLMAP on clean sequences as reference. Since Dark3R and MASt3R-SfM are evaluated against the same reference, the comparison is fair; however, a sentence clarifying that FLSea pose numbers are scale-free and reference-relative would help.
  5. [Barbados dataset] The paper establishes a new dataset but no release URL or availability statement is provided. If the dataset is to be a contribution, please include a link or state the intended release plan in the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SfM physics head is trained under paired supervision and the Barbados rendering/localization results are measured against external reference protocols, not derived from the method's own outputs.

full rationale

Swimm3R's claimed contributions are not circular. The medium-aware SfM is initialized from the external MASt3R checkpoint, fine-tuned with LoRA, and trained with a multi-term objective (Lgeom, Lfor, Linv, Lsat) that includes distillation of a fixed teacher's features and explicit forward/inverse supervision against paired clean-underwater data. The inverse restoration in Eq. (4) is a model prediction depending on the network's own estimated depth and medium parameters, but it is trained against ground-truth J and evaluated on held-out FLSea sequences with external COLMAP reference poses, so it is not a fitted input renamed as a prediction. The UWBS renderer is a separate component trained with photometric and regularization losses; the same test-time pose refinement protocol (300 steps on the test image, map fixed) is applied equally to all renderers, so the PSNR comparison is controlled, albeit not a pure novel-view-synthesis test. The Barbados dataset is new, and the paper explicitly states that pose numbers are comparative because metric ground-truth poses are unavailable; this is a disclosed evaluation limitation rather than a self-referential derivation. The authors' self-citations (e.g., Sattar et al. in [9], Kweon & Park in [18]) are background related-work citations and are not load-bearing for any central claim. No equation reduces to a definitional identity, no fitted parameter is presented as an independent prediction, and no uniqueness claim is imported from prior work by the same authors. Therefore the paper's central derivation chain is self-contained and not circular.

Assumptions & free parameters 4 free parameters · 5 assumptions · 3 invented entities

The central reconstruction pipeline rests on a learned physics model with hand-set augmentation priors and unvalidated learned coefficients, plus a standard splatting framework (Beta primitives from UBS, medium modeling from WaterSplatting). The new entities (SGG, ambient field, medium field parameterization) are supported only by internal ablation results and by rendering/localization metrics, not by independent measurements. The Barbados scenes lack ground truth, so geometry claims depend on proxy metrics and reference poses.

free parameters (4)
  • Learned optical coefficients (beta_D, beta_B, B_inf) of the physics head = not reported
    The head regresses these from images and they drive the restoration. No validation against measured optical parameters is given; training-time augmentation ranges (Eq. 9) shape their distribution.
  • Water-type augmentation ranges (Eq. 9) = hand-specified RGB intervals for cyan/teal/green/warm and beta sampling ranges
    These synthesize the training distribution for beta_D, beta_B, B_inf. They are chosen by hand, not derived from data, and directly determine what the network learns to estimate.
  • Test-time per-camera pose codes (3R-GS protocol) = 300 steps of refinement per test frame
    Evaluated test frames get their own pose optimized against the held-out image. This is a per-evaluation fitted quantity that can absorb some test-frame appearance and inflate PSNR.
  • SGG hyperparameters (tau_c, eta, policy f) = tau_c=10,000 steps, eta=0.15
    Chosen by hand; the confidence/depth policy f is not fully specified. They affect geometry optimization and thus the rendered maps.
assumptions (5)
  • domain assumption The Sea-Thru style revised underwater image formation model: I = J * TD + B_inf * (1 - TB).
    Invoked in Eq. (3) and used throughout the physics head and UWBS renderer; it assumes the revised model of Akkaynak and Treibitz is an adequate description of the capture process.
  • domain assumption Direct attenuation beta_D is depth-dependent but backscatter beta_B is depth-independent.
    Stated in Sec. III-B: beta_B is broadcast over pixels. This simplifies the medium field and is a modeling choice that may not hold in real water columns.
  • domain assumption In-air geometric priors (MASt3R encoder/decoder/descriptor features) transfer to underwater imagery via LoRA distillation.
    The whole medium-aware SfM rests on this transfer (Eq. 1). It is plausible and partially supported by FLSea results, but the transfer is not proven, only demonstrated.
  • domain assumption Reference poses from COLMAP (on clean FLSea) and GLUEMAP (on Barbados) are accurate enough for pose evaluation.
    Used in Tabs. II and V; the authors state no metric ground-truth poses exist. Errors in these references directly bias RRA/RTA numbers.
  • domain assumption Alpha compositing with Beta primitives and the fused underwater rasterizer (Eqs. 11-12) is a valid approximation of underwater radiance transfer.
    UWBS rendering quality depends on this compositing; it extends the splatting approximation rather than solving radiative transfer.
invented entities (3)
  • Scattering-aware Geometric Gradients (SGG) preconditioner
    purpose: Filters projected-center gradients so Beta primitives move along geometry rather than explaining backscatter during coarse optimization.
    Introduced in Algorithm 1 with no external validation; its support is internal (Tab. VIII shows LPIPS improves with SGG). It is a new algorithmic object, not a physical entity.
  • Ambient convolutional field A_phi for the underwater medium MLP
    purpose: Predicts per-pixel B_inf, beta_B, beta_D from spherical-harmonic ray directions using a 1x1 conv network.
    A new component of the medium representation; no standalone evidence beyond joint rendering quality.
  • Explicit underwater medium field (B_inf, beta_B, beta_D) in UWBS
    purpose: Factorizes scene appearance into water-column scattering/attenuation and object radiance.
    Inherits the idea from WaterSplatting/SeaThru but with a different parameterization; no direct measurements of the recovered medium are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Swimm3R: Splatting with Medium-aware SfM for Underwater 3D Reconstruction." pith.science (2026). https://pith.science/paper/FX4BNCDL

@misc{pith2026260800950,
  author       = {Pith},
  title        = {Pith review of: Swimm3R: Splatting with Medium-aware SfM for Underwater 3D Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FX4BNCDL}},
  note         = {Machine review of arXiv:2608.00950}
}
abstract

We propose Swimm3R, a unified framework that combines medium-aware structure-from-motion (SfM) with Underwater Beta Splatting to address scattering- and attenuation-induced failures in underwater 3D reconstruction. Swimm3R distills in-air geometric priors into a feed-forward backbone and uses a physics head to regress underwater image-formation parameters, camera poses, and restored point clouds. Additionally, we introduce Underwater Beta Splatting, which extends Gaussian splatting with Beta primitives and scattering-aware geometric gradients for stable underwater geometry representation. We further establish the Barbados underwater video dataset to demonstrate the effectiveness of our method in challenging underwater environments. On this dataset, Swimm3R robustly recovers underwater scene structure under challenging scattering conditions, yielding coherent seafloor geometry. Using these predicted point clouds, the proposed Underwater Beta Splatting improves average PSNR by $1.47$ dB over WaterSplatting while increasing downstream localization performance by $2.0$ and $2.4$ percentage points in RRA@15 and RTA@15, respectively.

Figures

Figures reproduced from arXiv: 2608.00950 by the authors.

Figure 1
Figure 1. Swimm3R achieves medium-aware SfM through in-air [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of Swimm3R. Our model learns medium-aware SfM from raw underwater images via LoRA fine-tuning, whose estimated camera poses and point clouds provide rich geometric priors for underwater Beta splatting. To stabilize restoration and avoid red or cyan saturation, we penalize out-of-range inverse radiance and anchor the veiling light to dim-pixel color statistics: Lsat = 1 N X N n=1  [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 3
Figure 3. Qualitative comparison of underwater 3D reconstruction. Swimm3R yields more continuous and geometrically coherent [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of RGB and depth across SfM initializations (rows) and renderers (columns). UWBS preserves [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative query-to-render feature matching for visual localization. UWBS produces more coherent matches across [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 34 canonical work pages

  1. [1]

    Real-time dense 3D mapping of underwater environments,

    W. Wang, B. Joshi, N. Burgdorfer, K. Batsosc, A. Q. Lid, P. Mordo- haia, and I. Rekleitisb, “Real-time dense 3D mapping of underwater environments,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 5184–5191

    work page 2023

  2. [2]

    Illumination and attenuation correction tech- niques for underwater robotic optical imaging platforms,

    J. W. Kaeli and H. Singh, “Illumination and attenuation correction tech- niques for underwater robotic optical imaging platforms,”Review, Avail- able at: https://www. whoi. edu/cms/files/kaeli joe14 InReview 183164. pdf (accessed 24 August 2018), 2017

    work page 2018

  3. [3]

    A revised underwater image formation model,

    D. Akkaynak and T. Treibitz, “A revised underwater image formation model,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 6723–6732

    work page 2018

  4. [4]

    A computer model for underwater camera systems,

    B. McGlamery, “A computer model for underwater camera systems,” in Ocean Optics VI, vol. 208. SPIE, 1980, pp. 221–231

    work page 1980

  5. [5]

    Computer modeling and the design of optimal underwater imaging systems,

    J. S. Jaffe, “Computer modeling and the design of optimal underwater imaging systems,”IEEE Journal of Oceanic Engineering, vol. 15, no. 2, pp. 101–111, 1990

    work page 1990

  6. [6]

    Single image haze removal using dark channel prior,

    K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, no. 12, pp. 2341–2353, 2010

    2010

  7. [7]

    Automatic red- channel underwater image restoration,

    A. Galdran, D. Pardo, A. Pic ´on, and A. Alvarez-Gila, “Automatic red- channel underwater image restoration,”Journal of Visual Communication and Image Representation, vol. 26, pp. 132–145, 2015

    work page 2015

  8. [8]

    Sea-thru: A method for removing water from underwater images,

    D. Akkaynak and T. Treibitz, “Sea-thru: A method for removing water from underwater images,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 1682–1691

    2019

Show all 45 references
  1. [9]

    Fast underwater image enhancement for improved visual perception,

    M. J. Islam, Y . Xia, and J. Sattar, “Fast underwater image enhancement for improved visual perception,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 3227–3234, 2020

    2020

  2. [10]

    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,” inProceedings of the European Conference on Computer Vision, 2020

    2020

  3. [11]

    3D Gaussian Splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, 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

  4. [12]

    Neural underwater scene representation,

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

    2024

  5. [13]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 56–65

    2023

  6. [14]

    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,”arXiv preprint arXiv:2408.08206, 2024

    2024 arXiv

  7. [15]

    SeaSplat: Representing underwater scenes with 3D Gaussian splatting and a physically grounded image formation model,

    D. Yang, J. J. Leonard, and Y . Girdhar, “SeaSplat: Representing underwater scenes with 3D Gaussian splatting and a physically grounded image formation model,”arXiv preprint arXiv:2409.17345, 2024

    2024 arXiv

  8. [16]

    Plenodium: Underwater 3D scene reconstruction with plenoptic medium representation,

    C. Wu, J. Dong, C. Li, and J. Tang, “Plenodium: Underwater 3D scene reconstruction with plenoptic medium representation,”arXiv preprint arXiv:2505.21258, 2025

    2025 arXiv

  9. [17]

    RestorGS: Depth-aware Gaussian splatting for efficient 3D scene restoration,

    Y . Qiao, M. Shao, L. Meng, and K. Xu, “RestorGS: Depth-aware Gaussian splatting for efficient 3D scene restoration,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 11 177–11 186

    2025

  10. [18]

    OceanSplat: Object-aware Gaussian splatting with trinocular view consistency for underwater scene reconstruction,

    M. Kweon and J. Park, “OceanSplat: Object-aware Gaussian splatting with trinocular view consistency for underwater scene reconstruction,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 7, 2026, pp. 5773–5781

    2026

  11. [19]

    MASt3R-SfM: a fully-integrated solution for unconstrained structure-from-motion,

    B. P. Duisterhof, L. Zust, P. Weinzaepfel, V . Leroy, Y . Cabon, and J. Revaud, “MASt3R-SfM: a fully-integrated solution for unconstrained structure-from-motion,” in2025 International Conference on 3D Vision (3DV). IEEE, 2025, pp. 1–10

    2025

  12. [20]

    LoRA: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “LoRA: Low-rank adaptation of large language models.” Iclr, vol. 1, no. 2, p. 3, 2022

    2022

  13. [21]

    Universal Beta Splatting,

    R. Liu, Z. Gao, B. Planche, M. Chen, V . N. Nguyen, M. Zheng, A. Choudhuri, T. Chen, Y . Wang, A. Fenget al., “Universal Beta Splatting,”arXiv preprint arXiv:2510.03312, 2025

    2025

  14. [22]

    Structure-from-motion revisited,

    J. L. Schonberger and J.-M. Frahm, “Structure-from-motion revisited,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 4104–4113

    2016

  15. [23]

    Pixelwise view selection for unstructured multi-view stereo,

    J. L. Sch ¨onberger, E. Zheng, M. Pollefeys, and J.-M. Frahm, “Pixelwise view selection for unstructured multi-view stereo,” inProceedings of the European Conference on Computer Vision, 2016

    2016

  16. [24]

    3D reconstruction of underwater structures,

    C. Beall, B. J. Lawrence, V . Ila, and F. Dellaert, “3D reconstruction of underwater structures,” in2010 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2010, pp. 4418–4423

    2010

  17. [25]

    Refractive calibration of underwater cameras,

    A. Jordt-Sedlazeck and R. Koch, “Refractive calibration of underwater cameras,” inEuropean conference on computer vision. Springer, 2012, pp. 846–859

    2012

  18. [26]

    Refractive COLMAP: refractive structure-from-motion revisited,

    M. She, F. Seegr ¨aber, D. Nakath, and K. K ¨oser, “Refractive COLMAP: refractive structure-from-motion revisited,” in2024 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 12 816–12 823

    2024

  19. [27]

    Automatic color correction for 3D reconstruction of underwater scenes,

    K. A. Skinner, E. Iscar, and M. Johnson-Roberson, “Automatic color correction for 3D reconstruction of underwater scenes,” inIEEE International Conference on Robotics and Automation. IEEE, 2017, pp. 5140–5147

    2017

  20. [28]

    WaterNeRF: Neural radiance fields for underwater scenes,

    A. V . Sethuraman, M. S. Ramanagopal, and K. A. Skinner, “WaterNeRF: Neural radiance fields for underwater scenes,” inOCEANS 2023- MTS/IEEE US Gulf Coast. IEEE, 2023, pp. 1–7

    2023

  21. [29]

    RecGS: Removing water caustic with recurrent Gaussian Splatting,

    T. Zhang, W. Zhi, B. Meyers, N. Durrant, K. Huang, J. Mangelson, C. Barbalata, and M. Johnson-Roberson, “RecGS: Removing water caustic with recurrent Gaussian Splatting,”IEEE Robotics and Automation Letters, vol. 10, no. 1, pp. 668–675, 2024

    2024

  22. [30]

    DUSt3R: Geometric 3D vision made easy,

    S. Wang, V . Leroy, Y . Cabon, B. Chidlovskii, and J. Revaud, “DUSt3R: Geometric 3D vision made easy,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 20 697– 20 709

    2024

  23. [31]

    Grounding image matching in 3D with MASt3R,

    V . Leroy, Y . Cabon, and J. Revaud, “Grounding image matching in 3D with MASt3R,” inEuropean conference on computer vision. Springer, 2024, pp. 71–91

    2024

  24. [32]

    Dark3R: Learning structure from motion in the dark,

    A. Y . Guo, A. Malik, S. Tedla, Y . Dai, Y . Qin, Z. Salehe, B. Attal, S. Nousias, K. Kutulakos, and D. B. Lindell, “Dark3R: Learning structure from motion in the dark,”arXiv preprint arXiv:2603.05330, 2026

    2026

  25. [33]

    3D reconstruction with spatial memory,

    H. Wang and L. Agapito, “3D reconstruction with spatial memory,” in 2025 International Conference on 3D Vision (3DV). IEEE, 2025, pp. 78–89

    2025

  26. [34]

    VGGT: Visual Geometry Grounded Transformer,

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “VGGT: Visual Geometry Grounded Transformer,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5294–5306

    2025

  27. [35]

    InstantSplat: Sparse-view Gaussian Splatting in seconds,

    Z. Fan, W. Cong, K. Wen, K. Wang, J. Zhang, X. Ding, D. Xu, B. Ivanovic, M. Pavone, G. Pavlakoset al., “InstantSplat: Sparse-view Gaussian Splatting in seconds,”arXiv preprint arXiv:2403.20309, 2024

    2024 arXiv

  28. [36]

    3R-GS: Best practice in optimizing camera poses along with 3DGS,

    Z. Huang, P. Wang, J. Zhang, Y . Liu, X. Li, and W. Wang, “3R-GS: Best practice in optimizing camera poses along with 3DGS,”arXiv preprint arXiv:2504.04294, 2025

    2025 arXiv

  29. [37]

    Global Structure-from-Motion meets feedforward reconstruction,

    L. Pan, J. Sch ¨onberger, and M. Pollefeys, “Global Structure-from-Motion meets feedforward reconstruction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 21 880–21 890

    2026

  30. [38]

    3D Gaussian Splatting as Markov Chain Monte Carlo,

    S. Kheradmand, D. Rebain, G. Sharma, W. Sun, Y .-C. Tseng, H. Isack, A. Kar, A. Tagliasacchi, and K. M. Yi, “3D Gaussian Splatting as Markov Chain Monte Carlo,”Advances in Neural Information Processing Systems, vol. 37, pp. 80 965–80 986, 2024

    2024

  31. [39]

    FLSea: Underwater visual-inertial and stereo- vision forward-looking datasets,

    Y . Randall and T. Treibitz, “FLSea: Underwater visual-inertial and stereo- vision forward-looking datasets,”arXiv preprint arXiv:2302.12772, 2023

    2023 arXiv

  32. [40]

    Human-visual-system-inspired un- derwater image quality measures,

    K. Panetta, C. Gao, and S. Agaian, “Human-visual-system-inspired un- derwater image quality measures,”IEEE journal of oceanic engineering, vol. 41, no. 3, pp. 541–551, 2015

    2015

  33. [41]

    An underwater color image quality evaluation metric,

    M. Yang and A. Sowmya, “An underwater color image quality evaluation metric,”IEEE transactions on image processing, vol. 24, no. 12, pp. 6062–6071, 2015

    2015

  34. [42]

    Underwater ranker: Learn which is better and how to be better,

    C. Guo, R. Wu, X. Jin, L. Han, Z. Chai, W. Zhang, and C. Li, “Underwater ranker: Learn which is better and how to be better,” inProceedings of the AAAI Conference on Artificial Intelligence, 2023

    2023

  35. [43]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 586–595

    2018

  36. [44]

    Camera pose refinement via 3D Gaussian Splatting,

    L. Hao, L. Zhou, Z. Wei, and X. Wang, “Camera pose refinement via 3D Gaussian Splatting,”arXiv preprint arXiv:2508.17876, 2025

    2025 arXiv

  37. [45]

    EPnP: An accurate O( n) solution to the PnP problem,

    V . Lepetit, F. Moreno-Noguer, and P. Fua, “EPnP: An accurate O( n) solution to the PnP problem,”International Journal of Computer Vision, vol. 81, no. 2, pp. 155–166, 2009

    2009

Pith tools

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