PaperlayerView exact revision on arXiv

Source-linked AI summary

Gaussian Splatting Underwater: A Controlled Cross-Regime Study

Olaya Álvarez-Tuñón, Stella Graßhof

arXiv:2608.25483v1cs.CV

TL;DR

Underwater Gaussian-splatting methods may produce visually plausible renders while reconstructing incorrect geometry, especially beyond shallow or controlled conditions. This paper compares five public systems across four water regimes under a shared protocol and finds that deployment conditions often matter more than architecture, with geometry frequently contradicting photometric ranking.

  • Problem

    Evidence is limited on whether underwater Gaussian-splatting methods reconstruct correct geometry in deeper, more turbid, and operational conditions rather than only producing plausible renders.

  • Method

    The paper evaluates five public-code systems across four water regimes using shared poses, initialisation, budget, evaluator, and geometric plus photometric measurements.

  • Results

    Deployment conditions often determine outcomes more than architecture: on the operational survey, a restoration pre-pass before vanilla 3DGS achieves the best geometry without improvement from medium-aware systems.

  • Takeaways & Limitations

    Reliable underwater reconstruction comparisons require geometry alongside appearance because photometric rank can coincide with contaminated geometry.

  • Takeaways & Limitations

    The benchmark largely evaluates clear water, so conditions where underwater reconstruction is hardest are underrepresented.

Abstract

from arXiv · show

The underwater environment is challenging for 3D reconstruction, because particles suspended in the water scatter and diffuse light, turbidity varies, absorption depends on wavelength, and illumination is rarely uniform. Methods based on Gaussian splatting have generally been developed for conditions that allow good image quality, and have primarily been tested on relatively shallow water. This paper examines how well Gaussian splatting performs across publicly available underwater datasets representing different degrees of turbidity, loss of illumination, and colour attenuation, together with an industrial survey. Five systems with public code are run under one protocol, with shared poses, initialisation, budget, and evaluator, to establish their relative advantages, disadvantages, and limitations. What these methods can do turns out to depend more on the setup than on the architecture. Water clarity binds upstream of rendering, since structure-from-motion registers 99.5 \% of frames in clear water and 0.0 \% at 12 NTU. Illumination geometry decides whether a medium model helps at all: under an artificial light that moves with the camera, medium-blind splatting beats both medium-aware systems. On the survey the benchmark's photometric leader comes last, beaten on geometry by a restoration pre-pass in front of vanilla 3DGS---and none of it is visible in the scores the field reports. Scene builds, per-run configurations, and evaluation code are released at https://github.com/olayasturias/uw3dgs

I. INTRODUCTION

The paper tests whether underwater Gaussian-splatting methods reconstruct geometry or merely produce plausible renders. A controlled cross-regime protocol shows that outcomes depend strongly on water clarity, illumination, and evaluation choice.

  • Underwater reconstruction supports marine robotics, inspection, autonomous-vehicle localisation, archaeology, and environmental monitoring.
  • Existing underwater Gaussian-splatting methods were developed mainly for shallow or controlled conditions, leaving deeper and more turbid ROV operations uncertain.
  • Photometric metrics cannot distinguish corrected appearance from geometry fabricated to explain the medium.
  • The study compares five code-available systems across four water regimes using shared poses, initialisation, budget, and evaluator.
  • Photometric and geometric quality can move in opposite directions, while the field’s standard metric saturates as reconstruction becomes difficult.

A. Underwater image formation

The underwater image model combines range-dependent attenuation of scene radiance with accumulated backscatter. Because both terms depend on range and attenuation is wavelength-dependent, colour recovery and geometry recovery are coupled.

  • The model represents directly transmitted scene signal and a backscattered veil accumulated along the same path.
  • J_c denotes true scene colour, z denotes range, and B∞_c denotes background-light radiance at infinite range.
  • β_D and β_B differ and both depend on wavelength, making single-coefficient haze models physically inaccurate in water.
  • Every image-formation term depends on range z, coupling colour recovery with geometry recovery.

B. 3D Gaussian Splatting: State of the Art

Underwater Gaussian-splatting methods modify medium representation, supervision, or primitive state to handle range-dependent scattering. The study spans four regimes from controlled benchmarks to realistic deep and industrial surveys.

  • Standard 3DGS stores radiance on explicit primitives, but underwater scattering varies with range and may require semi-transparent water-column primitives.
  • Medium realisation: Medium realisation ranges from image-space restoration and separate volumetric fields to per-Gaussian terms and modified compositing.
  • Additional supervision: Underwater methods add geometric constraints such as monocular depth, density control, distractor masks, and semantic guidance.
  • Per-Gaussian state: SeaSplat uses degree-0 spherical harmonics, while UW-GS and SeaFree-GS store clean and degraded colours per primitive.
  • Scenes: S1–S4 progress from an in-distribution shallow benchmark and controlled turbidity sequences to deep co-moving-light imagery and an industrial survey.

B. Methods

The methods and evaluation protocol isolate medium representation under shared experimental conditions. Five systems use common scene inputs and budgets, while geometry, photometry, pose registration, and run-to-run spread are measured explicitly.

  • Methods: The five evaluated systems represent no medium model, image-space removal, a separate field, or medium terms attached to primitives.
  • Methods: M1 applies fixed gray-world white balance followed by CLAHE, providing a training-free and network-free underwater baseline.
  • Controlled protocol: The protocol shares poses, initialisation, splits, resolution, inputs, and evaluator so remaining differences are attributable to systems.
  • Controlled protocol: At 7 NTU and above, usable triangulation does not exist, so the clear-water cloud seeds every turbidity level.
  • Controlled protocol: Each cell receives 30 k optimisation iterations on one 24 GB GPU, with outcomes reported when defaults exceed budget.
  • Controlled protocol: Repeated runs estimate approximately ±2 mm spread for M0 and an 82 mm spread for M2 at 7 NTU.

D. Metrics

The paper evaluates photometric fidelity, geometry, floating matter, and registration separately, because image quality alone cannot distinguish reconstruction from haze fitting.

  • PSNR, SSIM, and LPIPS measure held-out-view fidelity, structural similarity, and pretrained-feature distance.
  • A geometry-free control uses the per-pixel median training image to measure what reconstruction adds beyond an easy static image.
  • On S2, rendered depth is compared with stereo reference depth, including an opacity-gated render to quantify translucent floating matter.
  • On S4, opacity-gated Gaussian centres are compared with a photogrammetric cloud using accuracy, completeness, and chamfer after ICP alignment.
  • Registration reports the fraction of frames COLMAP registers on S2, isolating turbidity because sequences differ only in turbidity and use external poses.

E. The experiments

The experiments vary turbidity, pose source, scene, and system coverage under a controlled setup, while documenting resource and input constraints.

  • E1 compares five systems on the benchmark and operational survey, while E2 sweeps measured turbidity with trajectory, poses, and initialisation fixed.
  • M1 is excluded from monochrome S2 and S3, while M4 is evaluated only where monocular depth maps were prepared.
  • Table II reports the SOTRUE turbidity sweep and the E3 replacement of encoder poses with COLMAP poses.
  • Fig. 2 plots PSNR above and stereo-referenced surface depth error below across turbidity for M0, M2, and M3.

IV. RESULTS

The controlled results show that turbidity improves photometric scores while degrading geometry, with systems exhibiting distinct geometric failure modes and registration collapsing at high turbidity.

  • 32.0 dB to 35.9 dB: M0’s PSNR improves from clear water to 7 NTU while surface depth error worsens from 99 mm to 848 mm.
  • 24.1 to 32.0 dB: the geometry-free control also improves across the same turbidity range, shrinking M0’s margin from 7.9 to 3.4 dB.
  • M0 saturates near 850 mm, M2 degrades gradually with a 50× primitive-count collapse, and M3 densifies until memory exhaustion or saturation.
  • 0.13 dB: replacing encoder poses with COLMAP poses changes M3 from 28.30 to 28.17 dB at 0 NTU, within run-to-run noise.
  • 99.5 % to 0.0 %: COLMAP registers nearly all clear-water frames but none at 12 NTU, despite a detector tuned to low-contrast features.

C. E1 and E5: Cross-Regime and Deep Water

Cross-regime evaluation reveals that photometric rankings can disagree with geometric quality, while view overlap and simple restoration can outweigh architecture under operational conditions.

  • E1: Cross-Regime: On S1, all four in-medium systems fall within 1.3 dB, but floater mass ranges from 0.077 for M2 to over half of M4’s Gaussians.
  • E1: Cross-Regime: On S4, M3’s benchmark-leading result becomes last geometrically at 369 mm chamfer and 24.7 dB, with 353 minutes of training.
  • E1: Cross-Regime: The S4 floater statistic ranks M3 second while chamfer ranks it last, distinguishing stray haze from opaque geometry in the wrong place.
  • E1: Cross-Regime: M1 reaches 45 mm chamfer after 27 minutes, comparable to M2 at 46 mm and M0 at 58 mm within the 12 mm alignment residual.
  • E5: Deep Water: In deep water with a moving light, vanilla 3DGS leads appearance but carries the most veil, while the sparse dive loses 10.5 dB from reduced view overlap.

D. E4: Component Ablation

The SeaSplat ablation shows that underwater components can improve photometric scores while worsening geometric fidelity, with backscatter primarily affecting brightness and depth supervision encouraging floaters.

  • Backscatter term: 10.4 dB of PSNR is lost when backscatter is removed, while SSIM changes from 0.895 to 0.825 because the render becomes too dark.Backscatter models haze brightness; PSNR penalizes the resulting mean squared error more strongly than SSIM penalizes the structural change.
  • Evaluation: The geometric evaluation for S1 uses GT-free floater mass because the scene has no metric geometry reference.Table IV reports component ablations on S1 using floater mass as the geometric axis.
  • Depth prior: 2.8 dB of PSNR is lost when depth supervision is removed, but floater mass falls from 0.37 to 0.25.The depth prior encourages thin semi-transparent sheets that satisfy rendered-depth losses without committing to the true surface.
  • Total underwater contribution: 6.2 dB of PSNR is lost when the whole medium is removed, compared with 10.4 dB when only backscatter is removed.Attenuation without the veil term darkens the render more than removing the entire medium.

E. E6: Cross-Site Medium Transfer

Cross-site transfer is weak because medium parameters are site-specific, while medium placement creates distinct trade-offs between appearance, geometry, and computational cost. A restoration pre-pass achieves strong geometry without explicitly modeling the medium, whereas attached medium state and distractor handling impose substantial costs or artifacts.

  • Cross-site transfer: −9.0 dB is the penalty when medium networks fitted on S1 are used to render S3, with PSNR falling from 25.9 to 17.0 dB.The fitted parameters describe the training images rather than reusable water properties.
  • Separate medium field: M2 has the cleanest geometry in the study, with 0.077 floater mass on S1, 0.001 on S4, and ≤0.002 across the sweep.Its separate field degrades gradually from 289 to 809 mm, but can explain scene elements that change under moving light as medium structure.
  • Per-Gaussian medium: M3’s full medium model has floater mass 0.368, versus 0.244–0.259 for every single-component removal.The analytic per-Gaussian medium improves appearance but represents the veil as translucent mass when no alternative explanation is available.
  • Image-space pre-pass: M1 achieves 45 mm chamfer on the operational survey, 68 mm on the clear tank, and 0.038 floater mass on the deep vent.Its image-space enhancement restores gradient signal without modeling the medium, but arbitrarily corrected colour makes its photometrics incomparable.
  • Computational cost: Attaching medium state to primitives costs 67 min–13.1 h, reaching 12× the cost of medium-outside systems on S4.M4 does not finish beyond the 21-view benchmark, while M1–M2 remain within the 24–70 minute band across all four scenes.

VI. CONCLUSION

Across four underwater regimes, deployment conditions influence reconstruction outcomes as much as architecture. Geometry must be evaluated alongside appearance because photometric results can favor contaminated geometry.

  • The outcome is set as much by deployment conditions as by architecture across four water regimes.
  • On the operational survey, a fixed restoration pre-pass before vanilla 3DGS produces the best surfaces, without improvement from medium-aware systems.
  • Water clarity removes poses above roughly 7 NTU before rendering, while moving illumination eliminates the medium models’ advantage.
  • Photometric rank can correlate positively with contaminated geometry, so appearance cannot carry comparisons in this domain.
  • Reliable benchmarking requires reporting geometric evaluation alongside photometric evaluation, especially beyond the field’s predominantly clear-water conditions.
Loading 2608.25483v1…