WaterSplat-SLAM: Photorealistic Monocular SLAM in Underwater Environment

Kangxu Wang; Shaofeng Zou; Chenxing Jiang; Yixiang Dai; Siang Chen; Shaojie Shen; Guijin Wang

cs.ROApr 6, 2026

WaterSplat-SLAM: Photorealistic Monocular SLAM in Underwater Environment

Authors: Kangxu Wang, Shaofeng Zou, Chenxing Jiang, Yixiang Dai, Siang Chen, Shaojie Shen, Guijin Wang

Organizations: Department of Electronic Engineering, Tsinghua University, Beijing 100084, China · Key Laboratory of Marine Robotics, Shenyang 110169, China University of Chinese Academy of Sciences, Beijing 101408, China · Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Hong Kong, China

Abstract

Underwater monocular SLAM is a challenging problem with applications from autonomous underwater vehicles to marine archaeology. However, existing underwater SLAM methods struggle to produce maps with high-fidelity rendering. In this paper, we propose WaterSplat-SLAM, a novel monocular underwater SLAM system that achieves robust pose estimation and photorealistic dense mapping. Specifically, we couple semantic medium filtering into two-view 3D reconstruction prior to enable underwater-adapted camera tracking and depth estimation. Furthermore, we present a semantic-guided rendering and adaptive map management strategy with an online medium-aware Gaussian map, modeling underwater environment in a photorealistic and compact manner. Experiments on multiple underwater datasets demonstrate that WaterSplat-SLAM achieves robust camera tracking and high-fidelity rendering in underwater environments.

Figures & tables

Fig. 1: System overview of WaterSplat-SLAM: The system takes an RGB sequence as input and generates an online medium-aware Gaussian map. The RGB sequence is input into the segmentation model to generate semantic segmentation of images. The object regions are then fed into the Camera Tracking and Pointmap Estimation module to generate keyframe poses. Then, Gaussian primitives are initialized through keyframe poses and object parts. For each encoded ray vector, the Medium Network predicts the medium parameters. The Gaussian map is optimized using a semantic-guided photometric loss. Upon loop closure, we also perform adaptive adjustment and merging of Gaussian primitives.

Fig. 2: Illustration of medium-aware Gaussian mapping: Encoded ray vectors are passed through the Medium Network to predict three medium attributes: $\sigma^{\text{attn}}$ (attenuation density), $\sigma^{\text{bs}}$ (backscatter density), and $c^{\text{med}}$ (medium color). During rendering, the contributions of the medium and objects are explicitly separated.

Fig. 3: Gaussian primitives merging pipeline: When consecutive keyframes

k,k+1

detect loop closure with history frame

j

, we extract their anchored Gaussian primitives and establish a 3D voxel grid. All primitives falling within the same voxel are then merged into a single Gaussian primitive.

Fig. 4: Detailed comparison of reconstruction results for Curacao, JapRedSea, and Panama sequences on SeaThru-NeRF dataset. All three sequences exhibit the common challenge of turbid water and significant obscureness. We select a non-keyframe which is not included in training views across all methods in three sequences. Specific regions are zoomed in to highlight reconstruction details. WaterSplat-SLAM shows high-fidelity reconstruction and details, demonstrating its strong capability in clearly rendering both foreground and background objects despite the challenging underwater conditions.

Method	Type	Curasao			RedSea			JapRedSea			Panama			Avg
Method	Type	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
GO-SLAM [ 33 ]	NeRF	18.70	0.353	0.795	11.89	0.131	0.791	15.72	0.294	0.841	16.63	0.289	0.898	15.74	0.267	0.831
GLORIE-SLAM [ 22 ]	NeRF	21.79	0.567	0.558	16.02	0.415	0.575	16.11	0.446	0.494	18.79	0.452	0.523	18.18	0.470	0.538
MonoGS [ 9 ]	3DGS	20.83	0.613	0.643	Fail	Fail	Fail	14.76	0.465	0.553	18.54	0.597	0.557	18.04	0.558	0.584
HI-SLAM2 [ 12 ]		15.24	0.247	0.913	14.28	0.503	0.688	12.04	0.277	0.782	12.10	0.152	0.936	13.42	0.295	0.830
OpenGS-SLAM [ 24 ]		Fail	Fail	Fail	Fail	Fail	Fail	17.11	0.573	0.581	19.59	0.652	0.710	18.35	0.613	0.646

TABLE I: Rendering quality evaluations on SeaThru-NeRF dataset.

Fig. 5: Detailed reconstruction comparisons for Big_gate and Pipe_local and Pool_up2 sequences on our dataset. We select a non-keyframe which is not included in training views for all methods across three sequences. Specific regions are zoomed in to highlight reconstruction details.

Method	Type	Big_gate			Pipe_local			Pool_up			Pool_up2			Pool_loop			Avg
Method	Type	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
GO-SLAM [ 33 ]	NeRF	17.76	0.348	0.616	14.20	0.269	0.743	10.97	0.144	0.824	19.83	0.510	0.609	10.02	0.192	0.779	14.56	0.293	0.714
GLORIE-SLAM [ 22 ]	NeRF	19.51	0.643	0.556	21.54	0.682	0.597	Fail	Fail	Fail	20.48	0.649	0.519	Fail	Fail	Fail	20.51	0.658	0.557
MonoGS [ 9 ]	3DGS	25.97	0.765	0.672	23.71	0.860	0.649	23.21	0.864	0.629	26.44	0.897	0.572	17.25	0.747	0.719	23.32	0.827	0.648
HI-SLAM2 [ 12 ]		24.01	0.650	0.575	25.61	0.820	0.451	21.39	0.762	0.591	26.09	0.834	0.460	22.41	0.746	0.618	23.90	0.762	0.539
OpenGS-SLAM [ 24 ]		Fail	Fail	Fail	Fail	Fail	Fail	23.01	0.752	0.585	24.76	0.880	0.570	Fail	Fail	Fail	23.89	0.816	0.578

TABLE II: Rendering quality evaluations on WaterSplat-SLAM dataset.

Method	Big_gate	Pipe_local	Pool_up	Pool_up2	Pool_loop	Avg
WaterSplatting [ 28 ]	27.49	28.37	30.62	31.74	26.16	28.88
WaterSplat-SLAM	29.55	30.20	30.29	33.22	27.69	30.19

TABLE III: PSNR comparison between WaterSplatting and WaterSplat-SLAM on WaterSplat-SLAM dataset.

Method	Big_gate	Pipe_local	Pool_up	Pool_up2	Pool_loop	Avg
ORB-SLAM3 [ 35 ]	0.091	2.768	Fail	Fail	Fail	$\times$
GO-SLAM [ 33 ]	0.041	0.292	2.532	0.205	3.231	1.260
GLORIE-SLAM [ 22 ]	0.036	0.236	Fail	0.086	Fail	$\times$
MonoGS [ 9 ]	1.729	3.472	2.899	2.578	3.240	2.784
HI-SLAM2 [ 12 ]	0.034	0.135	2.533	1.213	2.343	1.252
OpenGS-SLAM [ 24 ]	Fail	Fail	2.063	2.249	Fail	$\times$

TABLE IV: Tracking accuracy (ATE(m)

\downarrow

) on WaterSplat-SLAM dataset.

Dataset	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Panama	w/o Water Mask	24.85	0.701	0.357
	w/o Primitives Adjustment	25.63	0.705	0.326
	WaterSplat-SLAM	25.95	0.723	0.270
Pool_loop	w/o Water Mask	27.21	0.797	0.353
	w/o Primitives Adjustment	26.56	0.790	0.376
	WaterSplat-SLAM	27.69	0.808	0.326

TABLE V: Ablation study of different modules in Panama and Pool_loop on rendering performance

	Big_gate	Pipe_local	Pool_up	Pool_up2	Pool_loop	Avg
w/o Water mask	0.077	0.160	0.323	0.274	0.375	0.242
WaterSplat-SLAM	0.066	0.164	0.289	0.224	0.353	0.219

TABLE VI: Water Mask module in WaterSplat-SLAM dataset on ATE(m)

\downarrow

Fig. 6: NVS result w/o Water Mask module.

Method	Primitives Size	Reconstruction Quality
Method	Primitives Size	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
w/o Primitives Merge	137657	27.59	0.807	0.303
WaterSplat-SLAM	104264	27.69	0.808	0.326

TABLE VII: Ablation study on primitives merge strategy in Pool_loop

Method	Big_gate	Pipe_local	Pool_up	Pool_up2	Pool_loop
MonoGS [ 9 ]	0.88 $\pm$ 0.24	0.84 $\pm$ 0.20	1.02 $\pm$ 0.20	1.12 $\pm$ 0.37	0.67 $\pm$ 0.41
HI-SLAM2 [ 12 ]	0.24 $\pm$ 2.31	0.17 $\pm$ 1.57	0.24 $\pm$ 2.02	0.39 $\pm$ 3.33	0.39 $\pm$ 1.74
WaterSplat-SLAM	0.21 $\pm$ 0.29	0.22 $\pm$ 0.25	0.31 $\pm$ 0.37	0.30 $\pm$ 0.43	0.47 $\pm$ 1.50

TABLE VIII: Per-frame runtime analysis (second) on WaterSplat-SLAM dataset.

Visual SLAM

Visual simultaneous localization and mapping (SLAM) estimates camera motion while building maps from images for navigation and scene understanding.

See all 108 papers

AquaFlow: A Monocular Gaussian Splatting SLAM for Underwater Streaming Reconstruction
Aug 24, 2026Yingxiang Xu, Kerui Ren, Wenqi Guo +4Underwater 3D Reconstruction 3DGS SLAM
Zhejiang University · Shanghai AI Laboratory · Shanghai Jiao Tong University +3
SurfSLAM: Sim-to-Real Underwater Stereo Reconstruction For Real-Time SLAM
Jan 15, 2026Onur Bagoren, Seth Isaacson, Sacchin Sundar +6Simultaneous Localization and Mapping Metric Depth Estimation
Department of Robotics, University of Michigan, Ann Arbor, MI 48109, USA
APVI-SLAM: Real-Time Acoustic-Pressure-Visual-Inertial Localization and Photorealistic Mapping System in Complex Underwater Environment
Jul 7, 2026Hanwen Zhang, Yipeng Zhu, Xiaopeng Guo +2Underwater Robotics Visual-Inertial SLAM
Division of Integrative Systems and Design, Hong Kong University of Science and Technology · School of Interdisciplinary Science, Beijing Institute of Technology · Division of Integrative Systems and Design, the Department of Computer Science and Engineering, and the Department of Ocean Science, Hong Kong University of Science and Technology

Aug 24, 2026J/K move · Enter open · S save

AquaFlow: A Monocular Gaussian Splatting SLAM for Underwater Streaming Reconstruction

Yingxiang Xu, Kerui Ren, Wenqi Guo, Changjian Jiang+3

Zhejiang University · Shanghai AI Laboratory · Shanghai Jiao Tong University+3

Underwater 3D Reconstruction 3DGS SLAM Streaming 3D Reconstruction Underwater Imaging

Open paper AskPDF arXiv

Recent monocular 3D Gaussian Splatting (3DGS) streaming reconstruction methods have achieved impressive performance by balancing reconstruction quality and efficiency. However, extending these frameworks to underwater scenes remains challenging due to severe visual degradation, such as light attenuation and scattering, which degrades camera pose tracking and distorts scene geometry. To address these challenges, we propose AquaFlow, a monocular Gaussian Splatting streaming reconstruction framework for efficient and high-fidelity underwater reconstruction. Specifically, AquaFlow fine-tunes a 3D vision foundation model on large-scale underwater data for robust pose and pointmap estimation, and introduces a medium-guided incremental Gaussian initialization strategy for streaming mapping. Furthermore, we develop a streaming-compatible hybrid scene representation that integrates structured, distance-conditioned neural Gaussians with a physics-inspired optical model to compensate for underwater image formation effects, enabling accurate scene reconstruction. We evaluate AquaFlow on a comprehensive dataset of 62 diverse underwater trajectories, collected from both public benchmarks and in-the-wild web videos across various scales. Extensive experiments demonstrate that AquaFlow achieves state-of-the-art tracking and rendering performance, reducing average localization error by 13.2% and improving PSNR by 4.74 dB compared to WaterSplat-SLAM.

Abstract

Visual SLAM

Explore similar work

AquaFlow: A Monocular Gaussian Splatting SLAM for Underwater Streaming Reconstruction

SurfSLAM: Sim-to-Real Underwater Stereo Reconstruction For Real-Time SLAM

APVI-SLAM: Real-Time Acoustic-Pressure-Visual-Inertial Localization and Photorealistic Mapping System in Complex Underwater Environment