ReefMapGS: Enabling Large-Scale Underwater Reconstruction by Closing the Loop Between Multimodal SLAM and Gaussian Splatting

Daniel Yang; Jungseok Hong; John J. Leonard; Yogesh Girdhar

cs.ROApr 13, 2026

ReefMapGS: Enabling Large-Scale Underwater Reconstruction by Closing the Loop Between Multimodal SLAM and Gaussian Splatting

Authors: Daniel Yang, Jungseok Hong, John J. Leonard, Yogesh Girdhar

Organizations: Massachusetts Institute of Technology · Woods Hole Oceanographic Institution

Abstract

3D Gaussian Splatting is a powerful visual representation, providing high-quality and efficient 3D scene reconstruction, but it is crucially dependent on accurate camera poses typically obtained from computationally intensive processes like structure-from-motion that are unsuitable for field robot applications. However, in these domains, multimodal sensor data from acoustic, inertial, pressure, and visual sensors are available and suitable for pose-graph optimization-based SLAM methods that can estimate the vehicle's trajectory and thus our needed camera poses while providing uncertainty. We propose a 3DGS-based incremental reconstruction framework, ReefMapGS, that builds an initial model from a high certainty region and progressively expands to incorporate the whole scene. We reconstruct the scene incrementally by interleaving local tracking of new image observations with optimization of the underlying 3DGS scene. These refined poses are integrated back into the pose-graph to globally optimize the whole trajectory. We show COLMAP-free 3D reconstruction of two underwater reef sites with complex geometry as well as more accurate global pose estimation of our AUV over survey trajectories spanning up to 700 m.

Figures & tables

Fig. 1 : System overview ReefMapGStakes as input RGB images and inertial, acoustic, and pressure sensor data from an AUV. Sensor data is fused into odometry while the known, central landmark is detected from RGB frames. Together, this information is integrated into a factor graph to estimate the whole AUV trajectory, shown in the center. We incrementally build a 3DGS model of the scene, starting from the region of highest certainty near the center. The solid blue line shows observations incorporated into the current map. From this current map, we expand and incorporate new observations at the frontier, the solid red line. The observations are aligned with the current map starting from the pose estimate from the factor graph. Together with depth information from a metric depth estimation model, we initialize new Gaussians and then optimize the model, yielding an expanded map. We repeat this process to incorporate the remaining observations, shown in the dashed blue line.

Fig. 2 : Incremental, radially-expanding 3DGS optimization framework. Local camera pose refinement occurs as we radially and incrementally expand our 3DGS-based map representation, incorporating new cameras and their visual data into the scene. Given a current 3DGS map, target image, and initial camera pose estimate, we can refine the camera pose with gradient descent (as shown from the light to dark frustums). Sample target, initial pose render, and final pose render images are shown on the left for the blue frustum. As camera poses are refined, we add these refined camera pose estimates as factors into our pose-graph to globally update the whole robot trajectory, as shown between the light (before) and dark (after) overlayed trajectories. Local pose refinement aligns new observations before expanding the scene, while global trajectory optimization propagates these changes to poses far away.

	Tektite		Yawzi
Method	Length (m)	RMSE (m)	Length (m)	RMSE (m)
Metashape (ref)	347.305	0.000	695.504	0.000
GTSAM	347.198	0.328	701.151	0.283
ORB-SLAM3	17.714	0.202	6.931	0.222
DROID-SLAM	18.343	4.305	142.180	4.532
VGGT-SLAM	11.938	5.029	83.476	6.085

TABLE I : Trajectory estimation metrics. ATE RMSE is calculated in meters. Some methods are unable to track all frames, so we report RMSE over those frames that are tracked as well as total path length. Lower RMSE is better but the path length should also be closer to the reference. Best results using non-oracle data are highlighted as first , second , and third . f denotes complete tracking failure. For RMSE, lower is better

\downarrow.

Fig. 3 : Qualitative trajectory evaluation Top-down view of robot trajectories estimated by ReefMapGSand other baselines methods are plotted against the reference trajectory estimated by Metashape, dashed green, at both reef sites, Tektite (above) and Yawzi (below).

		Tektite				Yawzi
	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	RMSE (m) $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	RMSE (m) $\downarrow$
Oracle	Metashape+sfm+batch	20.447	0.499	0.886	0.000	19.977	0.492	0.918	0.000
	Metashape+sfm+inc	22.395	0.577	0.673	0.070	22.522	0.594	0.680	0.064
	Metashape+md+batch	20.273	0.501	0.880	0.000	20.060	0.496	0.917	0.000
	Metashape+md+inc	21.975	0.564	0.702	0.066	22.158	0.583	0.701	0.059
In the wild	GTSAM+md+batch	17.846	0.461	0.945	0.328	17.683	0.467	0.958	0.283

TABLE II : Reconstruction and tracking metrics across different methods. We ablate across different configurations of input data and 3DGS optimization and with WildGS-SLAM. Initial poses can come from GTSAM or Metashape , Gaussians are initialized either from the SfM point cloud, sfm , or depth estimator outputs, md , and the 3DGS can be optimized either in batch, batch , or incrementally following our method, inc . Input data configurations that use oracle data from SfM are grayed . Best results using non-oracle data are highlighted as first and second , while best results among all are bolded .

	Runtime (hh:mm)
Method	Tektite	Yawzi
COLMAP	10:40	17:14
Metashape	8:04	6:57
ORB-SLAM3	0:01	0:02
DROID-SLAM	0:03	0:05
VGGT-SLAM	0:08	0:06

TABLE III : Processing time. ReefMapGScuts down processing time significantly compared to open source (COLMAP) and proprietary (Metashape) SfM software, requiring roughly 3 hours for both scenes. While learning-based methods are fast none yield coherent results on our scenes.

Underwater 3D Reconstruction

Underwater 3D reconstruction recovers scene geometry and appearance from camera, sonar, and robot-sensor data despite refraction, scattering, and motion.

See all 38 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
MoonSplat: Monocular Online Gaussian Splatting with Sim(3) Global Optimization
Jun 16, 2026Guo Pu, Yixuan Han, Haofeng Li +33D Gaussian Splatting Language-Embedded 3DGS
Wangxuan Institute of Computer Technology, Peking University, China · Beijing Hydrogen Intelligent Tech. Co., Ltd., China
LightSplat: Real-Time High-Fidelity 3D Gaussian SLAM with Loop Closure
Sep 7, 2026Junze Bao, Ye Gao, Yiming Huang +5Simultaneous Localization and Mapping 3D Gaussian Splatting
School of Automation Science and Electrical Engineering, Beihang University, Beijing 100191, P. R. China, and also with Zhejiang Key Laboratory of Industrial Big Data and Robot Intelligent Systems, Hangzhou Innovation Institute, Beihang University, Hangzhou 310051, P. R. China. · School of Mechanical Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, P. R. China.

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

Underwater 3D Reconstruction

Explore similar work

AquaFlow: A Monocular Gaussian Splatting SLAM for Underwater Streaming Reconstruction

MoonSplat: Monocular Online Gaussian Splatting with Sim(3) Global Optimization

LightSplat: Real-Time High-Fidelity 3D Gaussian SLAM with Loop Closure