Gaussian Splatting Underwater: A Controlled Cross-Regime Study

Olaya Álvarez-Tuñón; Stella Graßhof

cs.CVAug 26, 2026

Gaussian Splatting Underwater: A Controlled Cross-Regime Study

Authors: Olaya Álvarez-Tuñón, Stella Graßhof

Abstract

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

Figures & tables

Fig. 1 : The proposed controlled cross-regime study, exemplified with EIVA Survey’s renderings and point clouds. The point clouds correspond to the Gaussians’ centres.

	Per-Gaussian state	Medium realisation	Extra supervision / inputs
M0 vanilla 3DGS [ 1 ]	pos., SH-3 colour, opacity, scale, rot.	—	—
M1 UIE $\rightarrow$ 3DGS [ 1 ]	as M0	removed in 2D pre-pass: gray world + CLAHE	restored images as input
M2 WaterSplatting [ 2 ]	as M0 (gsplat param.)	separate field: dir.-conditioned MLP $\rightarrow$ medium colour, attenuation and backscatter densities	medium regularisers
M3 SeaSplat [ 3 ]	pos., SH-0 colour , opacity, scale, rot.	analytic Eq. ( 1 ) in image space; CNNs infer $\beta^{D}\!,B^{\infty}$ from rendered depth	depth-smooth, depth-weighted recon., dark-channel, grey-world
M4 UW-GS [ 4 ]	as M0 + clean/degraded colour pair via view–distance MLP	per-Gaussian distance-dependent transform	depth-alignment loss; mono-depth maps required

TABLE I : System anatomy. All systems initialise from the shared per-scene SfM cloud of Sec. III-C . SH = spherical harmonics. Entries in bold are undocumented in the original paper.

NTU	Ctrl $\uparrow$	System	PSNR $\uparrow$	Err $\downarrow$ [mm]	Floater $\downarrow$	#G (k)
0.0	24.1	M0 3DGS	31.97	99	0.071	117
		M1 UIE $\rightarrow$ 3DGS	26.85 ∗	68	0.058	296
		M2 WaterSplatting	32.87	289	0.001	154
		M3 SeaSplat	28.17	438	0.047	269
		+ COLMAP poses	28.30	—	—	129
6.0	31.5	M0 3DGS	35.54	845	0.076	22

TABLE II : E2 turbidity sweep and E3 pose source on SOTRUE. Ctrl is the geometry-free control of Sec. III-D . The indented

+

COLMAP poses rows replace the encoder poses with a free COLMAP reconstruction (E3). † SeaSplat’s default densification exceeds 24 GB at 7 NTU; that cell alone uses a doubled threshold.

Fig. 2 : Turbidity dose–response on SOTRUE (E2), for the three systems that run there: M0, M2 and M3, the last without a 6 NTU cell. PSNR (top) rises with turbidity while surface depth error against the stereo reference (bottom, log scale) grows by up to

8.5\times

Scene	System	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Chamfer $\downarrow$	Floater mass $\downarrow$	Train (min)	#G (M)
S1 Curaçao	M0 3DGS	29.3	0.877	0.199	n/a	0.21	24	0.58
	M1 UIE $\rightarrow$ 3DGS	25.0 ∗	0.807 ∗	0.251 ∗	n/a	0.191	24	1.34
	M2 WaterSplatting	29.5	0.919	0.141	n/a	0.077	40	1.06
	M3 SeaSplat	30.2	0.895	0.173	n/a	0.37	82	3.35
	M4 UW-GS	30.6	0.898	0.179	n/a	0.551	188	0.93
S4 EIVA	M0 3DGS	25.9	0.780	0.509	58	0.024	28	0.24

TABLE III : E1 cross-regime comparison, with the S3 deep-water probe (E5) below. Chamfer is against each scene’s photogrammetric reference; S1 has none, so geometry there is floater mass. ∗ M1 is scored against its own enhanced targets and is not comparable in-medium.

Fig. 3 : Renders of one held-out view per regime (rows) for ground truth and all five systems (columns). In the clear tank (S2) the AprilTags act as a sharpness scale and the visual ordering follows the measured surface error of Table II : M1 and M0 hold the tags, M2 softens them, M3 washes out the right half of the frame. Under co-moving light (S3) M2’s render is the blurred residue of a 27 k-Gaussian model after its medium field absorbed the scene; on the survey (S4) M3 is visibly hazier. ∗ M1 renders its own enhanced images, so its column differs from ground truth by construction; M4 completes only on S1.

Config	Med.	Dep.	Bks.	PSNR $\uparrow$	SSIM $\uparrow$	Floater $\downarrow$ $\tau{=}.05/.1/.2$	#G (M)
A0 full	✓	✓	✓	30.22	0.895	.41 / .37 / .33	3.35
A1 $-$ backsc.	✓	✓	–	19.84	0.825	.31 / .26 / .21	1.30
A2 $-$ depth	✓	–	✓	27.46	0.881	.30 / .25 / .21	0.62
A3 $-$ medium	–	✓	–	23.98	0.865	.29 / .24 / .21	1.58

TABLE IV : E4: component ablation inside SeaSplat (M3) on S1. A0

\rightarrow

A3 is the total underwater contribution; A0

\rightarrow

A2 isolates the depth prior; A0

\rightarrow

A1 isolates the backscatter term. S1 carries no metric geometry reference, so the geometric axis is the GT-free floater mass of Sec. III-D .

Explore similar work

Underwater360: Reconstructing Underwater Scenes from Panoramic Images with Omnidirectional Gaussian Splatting
May 26, 2026Jiangbei Hu, Weichao Song, Shibo Yu +83D Gaussian Splatting Underwater Imaging
School of Software, Dalian University of Technology · College of Computing and Data Science, Nanyang Technological University
WaterClear-GS: Optical-Aware Gaussian Splatting for Underwater Reconstruction and Restoration
Jan 27, 2026Xinrui Zhang, Yufeng Wang, Zesheng Wang +33D Gaussian Splatting Underwater Image Enhancement
Beihang University
Swimm3R: Splatting with Medium-aware SfM for Underwater 3D Reconstruction
Aug 2, 2026Minseong Kweon, Junaed Sattar3D Gaussian Splatting Underwater Imaging
Department of Computer Science & Engineering and the Minnesota Robotics Institute (MnRI), University of Minnesota–Twin Cities, Minneapolis, MN 55455 USA

May 26, 2026J/K move · Enter open · S save

Underwater360: Reconstructing Underwater Scenes from Panoramic Images with Omnidirectional Gaussian Splatting

Jiangbei Hu, Weichao Song, Shibo Yu, Mohan Wang+7

School of Software, Dalian University of Technology · College of Computing and Data Science, Nanyang Technological University

3D Gaussian Splatting Underwater Imaging 3D Reconstruction Novel View Synthesis Underwater 3D Reconstruction Gaussian Splatting

Open paper AskPDF arXiv

Underwater scene reconstruction is essential for immersive exploration of aquatic environments, yet remains challenging due to complex participating-media effects such as absorption and scattering, as well as the limited field of view (FoV) of conventional cameras. Although combining panoramic imaging with 3D Gaussian Splatting (3DGS) offers a promising direction for photorealistic underwater rendering, traditional 3DGS struggles with both spherical projection distortion and underwater medium degradation. In this paper, we propose \textbf{Underwater360}, a physics-informed omnidirectional 3DGS framework for underwater panoramic scene reconstruction. First, we introduce an Omnidirectional Gaussian Splatting module that performs ray casting directly in spherical camera space instead of relying on 2D projection approximations, thereby reducing geometric distortions under 360

^\circ

FoV. Second, we design a physics-based appearance-medium modeling architecture with pose-conditioned appearance embeddings to explicitly decouple intrinsic scene radiance from depth-dependent backscatter and attenuation, enabling physically grounded scene appearance restoration. Finally, we establish a new panoramic underwater benchmark dataset containing both synthetic and real-world scenes. Extensive experiments demonstrate that Underwater360 achieves superior performance in underwater novel view synthesis and scene appearance restoration, delivering improved rendering quality and cross-view consistency in complex underwater environments. The code and datasets are released at https://github.com/SwcK423/Underwater360

Abstract

Explore similar work

Underwater360: Reconstructing Underwater Scenes from Panoramic Images with Omnidirectional Gaussian Splatting

WaterClear-GS: Optical-Aware Gaussian Splatting for Underwater Reconstruction and Restoration

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