R-Splatting: Underwater 3D Reconstruction

• R-Splatting is a unified underwater 3D reconstruction framework that fuses multiple restored views to address cross-view illumination and intra-view restoration inconsistencies.
• It employs a view-shared neural field and stochastic, uncertainty-aware opacity optimization to robustly manage depth-dependent attenuation and scattering-induced haze.
• Empirical evaluations on SeaThru-NeRF and BlueCoral3D show improved rendering metrics and geometric fidelity while maintaining competitive real-time performance.

R-Splatting is a unified underwater reconstruction framework that couples underwater image restoration (UIR) with 3D Gaussian Splatting (3DGS) in order to improve both rendering quality and geometric fidelity under depth-dependent attenuation, scattering-induced haze, non-uniform illumination, and cross-view appearance inconsistency (Huang et al., 22 Sep 2025). In the precise named sense, it refers to the method introduced in “From Restoration to Reconstruction: Rethinking 3D Gaussian Splatting for Underwater Scenes” (Huang et al., 22 Sep 2025). The broader splatting literature, however, also uses nearby “R”-prefixed names for other directions, including radiative tomography, ray tracing, radio-frequency radiomap extrapolation, and refinement-oriented 2D Gaussian splatting, so the term is context-sensitive (Zha et al., 2024).

1. Terminological scope and historical placement

R-Splatting most specifically denotes an underwater 3DGS framework that integrates multiple restored views produced by diverse UIR models into a single reconstruction pipeline (Huang et al., 22 Sep 2025). Its stated goal is to bridge underwater restoration and reconstruction rather than to treat restoration as a separate preprocessing stage. This positioning distinguishes it from standard 3DGS pipelines trained directly on degraded underwater imagery, and also from underwater methods that rely primarily on simplified physical models of underwater image formation.

The nomenclature nonetheless sits inside a crowded splatting vocabulary. Elsewhere in the literature, “R” and related prefixes have been used for “Rectifying Radiative Gaussian Splatting” in sparse-view CT (Zha et al., 2024), “Ray Tracing based Gaussian Splatting” in RaySplats (Byrski et al., 31 Jan 2025), radio-frequency Gaussian splatting in RadSplatter (Wang et al., 18 Feb 2025), and “2DGS-R” as a refinement strategy for 2D Gaussian Splatting (Ren et al., 19 Oct 2025). This suggests that “R-Splatting” is not a generic family name with one stable meaning across subfields; in current usage, the exact named method is the underwater framework of (Huang et al., 22 Sep 2025).

2. Underwater reconstruction problem formulation

The method targets multi-view underwater 3D reconstruction and novel view synthesis, with emphasis on simultaneously improving visual fidelity and geometry under adverse underwater image formation (Huang et al., 22 Sep 2025). The paper identifies two inconsistency sources as central: cross-view illumination inconsistency, in which the same object can appear differently across viewpoints because of changing orientation relative to the water surface, lighting, reflection, and refraction, and intra-view restoration inconsistency, in which different UIR models produce different enhanced outputs for the same input image.

These effects interact unfavorably with standard 3DGS. The paper states that underwater degradation harms both photometric supervision and geometric consistency, and that standard 3DGS can react to inconsistent supervision by increasing Gaussian opacity to explain view-specific artifacts. The result is a coupled degradation of rendering and structure, rather than a purely appearance-level failure (Huang et al., 22 Sep 2025). Existing underwater NeRF and 3DGS variants are described as relying on simplified physical models that often fail in complex scenes because effects such as forward scattering and light reflection are neglected.

R-Splatting addresses this by constructing training data not from one image stream but from multiple restored versions of each view. If there are $N$ multi-view underwater images and $M$ UIR models, the framework forms $M+1$ image groups,

$\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$

where $m=0$ includes the original image set (Huang et al., 22 Sep 2025). This representation turns restoration diversity into supervision diversity. A plausible implication is that the method treats restoration disagreement as a source of complementary cues rather than purely as noise.

3. Scene representation and illumination-conditioned color modeling

R-Splatting retains the standard 3DGS scene representation by Gaussian means $X_i$, covariances $\Sigma_i$, opacity, and spherical harmonics color, but alters how color is parameterized and supervised (Huang et al., 22 Sep 2025). The covariance factorization is the usual

$\Sigma_i = R_i S_i S_i^{T} R_i^{T},$

with image-space covariance

$\Sigma'_i = J W \Sigma_i W^{T} J^{T}.$

Thus the geometric core remains 3DGS-compatible.

The principal representational change is a conditional, view-shared color model. For each input image $I$, a tiny autoencoder extracts a latent code and decoded feature map,

$M$0

Given Gaussian position $M$1 and projection matrix $M$2, projected coordinates satisfy

$M$3

and a point-wise illumination feature is obtained by bilinear sampling,

$M$4

The sampled feature $M$5, together with a learnable Gaussian embedding $M$6 and position $M$7, is processed by a lightweight MLP,

$M$8

to predict a spherical-harmonics offset. The final Gaussian color for view direction $M$9 is

$M+1$0

This formulation makes the scene geometry explicit while allowing appearance to depend on restoration or illumination style through $M+1$1 (Huang et al., 22 Sep 2025). The paper describes this as view-shared color modeling because the neural field is executed once to compute a shared representation and multiple views can then be rendered by updating only the view direction in the rasterizer. Relative to fully per-view neural color models, this keeps inference close to the explicit 3DGS regime.

At inference, R-Splatting does not require a new input image to provide $M+1$2. Instead, a lightweight illumination generator maps random noise to a latent code with the same shape as $M+1$3, enabling diverse yet coherent restored renderings from the same reconstructed geometry (Huang et al., 22 Sep 2025). The implementation uses five convolutional layers with ReLU activations, and the latent dimension is set to $M+1$4.

4. Contrastive illumination learning and Uncertainty-Aware Opacity Optimization

The framework supplements the conditional color model with a contrastive loss over illumination latents. The stated objective is to make latent codes from the same restoration style close, latent codes from different restoration styles separated, and to include a latent sampled from a random clean scene in the denominator to improve generalization (Huang et al., 22 Sep 2025). Dynamic queues of recent latent codes are maintained for each restoration style, and the temperature is $M+1$5. The paper’s formula is typographically corrupted in the provided text, but the mechanism is explicit: positives are same-style latents from different views, negatives are latents from different restoration styles, and the target is a view-invariant style representation rather than a viewpoint-specific code.

The more distinctive regularizer is Uncertainty-Aware Opacity Optimization (UAOO). Standard 3DGS learns opacity directly. R-Splatting instead treats training-time opacity as stochastic,

$M+1$6

where $M+1$7 is sigmoid, $M+1$8 is the opacity mean parameter, and $M+1$9 is the opacity uncertainty parameter (Huang et al., 22 Sep 2025). The corresponding gradients scale as

$\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$0

The effect is to regularize opacity updates under inconsistent underwater supervision, suppress abrupt gradient responses, and reduce overfitting to view-specific artifacts (Huang et al., 22 Sep 2025).

At test time, the stochastic opacity is replaced by a deterministic approximation to its expectation,

$\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$1

The uncertainty regularizer is

$\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$2

The total loss is

$\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$3

with $\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$4 (Huang et al., 22 Sep 2025). The reconstruction term follows standard 3DGS and combines DSSIM and $\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$5, although the exact weighting is not reproduced in the provided text. An implementation consequence is that periodic opacity resetting is disabled; UAOO is intended to replace rather than coexist with that heuristic.

5. Datasets, baselines, and empirical results

The evaluation uses SeaThru-NeRF and BlueCoral3D (Huang et al., 22 Sep 2025). SeaThru-NeRF is described as a relatively less challenging benchmark with about 25 front-view images per scene. BlueCoral3D is introduced as a new dataset with two underwater scenes, about 500 high-resolution $\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$6 images per scene, full 360-degree sweeps, and naturally dynamic illumination; every 100th image is assigned to the test set and the rest to training.

Baselines include SeaThru-NeRF, 3DGS, WildGaussian, UW-GS, GS-W, and WaterSplatting (Huang et al., 22 Sep 2025). Reported quantitative metrics are PSNR, SSIM, and LPIPS; SeaThru-NeRF tables also include FPS and training time in GPU hours. The abstract states that R-Splatting improves both rendering quality and geometric fidelity, whereas the provided numerical tables emphasize rendering metrics rather than explicit geometric scores.

On SeaThru-NeRF, R-Splatting achieves $\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$7 on Curaçao, $\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$8 on Panama, $\{\{ \mathbf{I}^m_n \}_{n=1}^{N}\}_{m=0}^{M},$9 on IUI3, and $m=0$0 on Japanese Gardens, in PSNR/SSIM/LPIPS order (Huang et al., 22 Sep 2025). On BlueCoral3D, it reaches $m=0$1 on Scene 1 and $m=0$2 on Scene 2. The gains are larger on BlueCoral3D, which the paper attributes to stronger illumination changes and more complex camera motion.

The runtime profile remains substantially closer to 3DGS than to slower neural-field-heavy baselines. On SeaThru-NeRF, R-Splatting runs at 107 FPS with 0.86 GPU hours of training, compared with 163 FPS and 0.38 GPU hours for 3DGS, 30 FPS for UW-GS, 42 FPS for GS-W, 73 FPS for WildGaussian, and less than 1 FPS for SeaThru-NeRF (Huang et al., 22 Sep 2025). All speed and runtime numbers are measured on an NVIDIA RTX 4090.

Ablations isolate the neural field, UAOO, and periodic opacity resetting. Starting from 3DGS at $m=0$3, adding the neural field raises performance to $m=0$4 (Huang et al., 22 Sep 2025). A model with NF + UAOO + POR yields $m=0$5, whereas the full model with NF + UAOO and no POR reaches $m=0$6. A UAOO-only configuration without POR also improves strongly to $m=0$7. The pattern indicates that UAOO is beneficial when periodic opacity resetting is removed, and that the view-shared neural field and stochastic opacity are complementary.

6. Relation to adjacent splatting variants

R-Splatting belongs to a broader pattern in which 3DGS is adapted by modifying the scene prior, renderer, or target signal rather than abandoning explicit Gaussians. Its nearest conceptual distinction is that it does not redefine Gaussian geometry or rasterization fundamentally; instead it injects restoration-conditioned appearance modeling and uncertainty-regularized opacity learning into an underwater reconstruction pipeline (Huang et al., 22 Sep 2025).

This places it apart from RT-Splatting, which factorizes geometric occupancy and optical opacity to jointly model reflection and transmission on thin semi-transparent specular surfaces (Shi et al., 18 May 2026); from RaySplats, which replaces rasterization with ray tracing against Gaussian-derived confidence ellipsoids (Byrski et al., 31 Jan 2025); from $m=0$8-Gaussian, which reformulates Gaussian splatting for sparse-view tomographic reconstruction by rectifying a radiative integration bias (Zha et al., 2024); from RadSplatter, which extends Gaussian splatting to radio-frequency radiomap extrapolation via camera-free projection to beamspace (Wang et al., 18 Feb 2025); and from 2DGS-R, which uses hierarchical refinement and in-place cloning to reconcile geometry and rendering in 2D Gaussian Splatting (Ren et al., 19 Oct 2025). In each of those cases, the “R” label refers to a different modification axis.

A plausible implication is that the specific contribution of underwater R-Splatting is neither a new Gaussian primitive nor a new physically exact underwater image-formation equation. Its contribution is architectural: multiple restored observations are fused into one 3DGS model, appearance variation is encoded through a contrastively stabilized latent space, and opacity is regularized through stochastic uncertainty (Huang et al., 22 Sep 2025). Within the Gaussian splatting literature, it is therefore best understood as a restoration-aware reconstruction framework specialized to underwater scenes rather than as a generic splatting backend for other domains.