SURGE: Sonar-fUsed Reconstruction and localization via image-gated Graph Estimation

Mohammed Ibrahim M; Vallabh Deogaonkar; Trung Dong; Jane Shin; Abhilash Somayajula; Xiaomin Lin

cs.ROOct 5, 2026

SURGE: Sonar-fUsed Reconstruction and localization via image-gated Graph Estimation

Authors: Mohammed Ibrahim M, Vallabh Deogaonkar, Trung Dong, Jane Shin, Abhilash Somayajula, Xiaomin Lin

Organizations: Department of Electrical and Computer Engineering University of South Florida Tampa, FL, USA · Department of Ocean Engineering Indian Institute of Technology Madras Chennai, India · Department of Naval Architecture and Ocean Engineering Seoul National University Seoul, South Korea

Abstract

Remotely operated vehicles (ROVs) are widely used to explore and inspect underwater environments such as caves, shipwrecks, and submerged infrastructure. These missions require accurate 3D understanding of the surrounding environment, which depends on both reliable vehicle localization and metric scene reconstruction. However, external positioning is often unavailable underwater, requiring small ROVs to rely primarily on onboard perception. Optic vision provides rich visual and geometric information but suffers from scale ambi- guity and trajectory drift, whereas 2D imaging sonar provides metric range but incomplete 3D geometry. Existing underwater reconstruction approaches typically address these limitations separately or assume known sensor poses, leaving localization and reconstruction disconnected. We present SURGE, a camera sonar framework that jointly estimates the ROV trajectory and target location by integrating visual and acoustic observations within a factor graph, then uses the recovered metric poses for sonar Gaussian splatting. Experiments on real underwater RGB sonar observations show that SURGE substantially improves localization consistency over conventional vision based pose estimation and produces a more compact, natively metric reconstruction than RGB Gaussian splatting baselines.

Figures & tables

Fig. 1 : A representative scenario of an ROV scanning a cinder-block target, with the imaging sonar’s and camera’s fields of view projected onto the object for localization and reconstruction.

System	Modalities	Poses	Metric	Joint Loc.+Recon.	Setting
AONeuS [ 8 ]	RGB + Sonar	GT	Yes	No	Tank
Z Splat [ 9 ]	RGB + Sonar	GT	Yes	No	Tank
SonarSplat [ 10 ]	Sonar	Est.	Yes	Yes	Tank + River
Pose free 3D [ 12 ]	Sonar	Est.	Yes	Yes	Tank
SeaSplat [ 13 ]	RGB	COLMAP	No	No	Real
WaterSplatting [ 14 ]	RGB	COLMAP	No	No	Real

TABLE I : Comparison of representative underwater 3D reconstruction methods. “Real” denotes data collected outside controlled tank environments.

Fig. 2 : Pipeline overview. RGB frames and sonar scans pass through visual and acoustic feature extraction into object segmentation, visual pose estimation, and sonar peak estimation. Factor-graph modelling fuses a multi-view consistent peak, the segmentation centroid, a prior visual pose, and an object-position anchor; factor optimization recovers metric ROV poses and the target landmark. Cone-arch seed collection and 2D-sonar Gaussian splatting then produce the reconstruction reprojected onto RGB.

Fig. 3 : The data collection platform. A BlueROV2 carries an RGB camera and a 2D forward looking sonar in a rigid, forward facing configuration.

Fig. 4 : Sonar fan with candidate peaks. The RGB camera cone constrains admissible peaks to the target’s angular extent; peaks outside the cone are discarded. Among in-cone candidates, the multi-view vote elects the frame’s sonar observation.

Factor	Weight	Residual
$F_{\mathrm{sonar}}$	$4.0$	Projected $b$ vs. orange; zero cost inside a $12$ cm polar disk.
$F_{\mathrm{bbox}}$	$1.8$	Projected AABB of $b\pm h$ vs. SAM box.
$F_{\mathrm{uv}}$	$1.5$	Projected $b$ vs. SAM mask centroid.
$F_{\mathrm{colmap}}$	$0.25$	Soft absolute anchor to COLMAP ( $\sigma_{c}{=}0.8$ m); down-weighted near teleports.
$F_{\mathrm{smooth}}$	$2.5$	Small $\Delta c$ ; matches $\Delta c^{\mathrm{C}}$ when $\\|\Delta c^{\mathrm{C}}\\|\leq 0.4$ m.
$F_{\mathrm{yaw}}$	$0.7$	Small $\delta\psi$ and $\Delta\delta\psi$ .

TABLE II : Factors in the reported configuration.

Fig. 5 : Correspondence between RGB and sonar observations for a representative frame. Left: RGB with SAM silhouette (yellow), mask AABB (blue), and mask centroid (green). Right: matching FLIS fan with RGB-mask cone (cyan), factor-graph cone (magenta), projected brick (white), locked orange return, and projected centre $b$ (red).

Config	Inl.	$\|\Delta r\|$	$\|\Delta\mathrm{az}\|$	bbox
		(cm)	( ∘ )	(px)
COLMAP frozen	$312$	$189$	$12.4$	$210$
Pose opt., $b$ fix	$872$	$12.1$	$3.60$	$108$
Centre $+$ disk	$\mathbf{874}$	$\mathbf{10.6}$	$\mathbf{3.33}$	$\mathbf{93}$

TABLE III : Localization on the

880

-frame sequence.

|\Delta r|

and

|\Delta\mathrm{az}|

are medians of projected

b

against locked orange; bbox is mean projected-box vs. SAM error.

Fig. 6 : Recovered ROV trajectory (red) and target landmark (green) in the world frame, overlaid on the COLMAP prior (blue). The factor graph metricises the COLMAP path against the consensus orange and the frozen landmark, absorbing large SfM teleports without copying them.

Method	med $\to b$ (m)	$\leq\!0.5$ m	In mask	Metric
Vanilla 3DGS	$1.41$	$6.8\%$	$20\%$	no
SeaSplat	$2.20$	$1.6\%$	$8\%$	no
WaterSplatting	$1.43$	$7.1\%$	$17\%$	no
AquaSplat	$1.43$	$6.4\%$	$17\%$	no
Ours	$\mathbf{0.65}$	$\mathbf{30\%}$	$\mathbf{36\%}$	yes

TABLE IV : Reconstruction on the

880

-frame sequence (opaque

\alpha\!\geq\!0.05

). RGB methods use COLMAP poses and are Sim(3)-aligned to FG cameras before scoring; ours uses FG poses and is natively metric. med

\to b

: median distance to landmark

b

;

\leq\!0.5

m: fraction within

0.5

m of

b

; In mask: mean projected mass inside the SAM mask.

Fig. 7 : Reconstruction projected into representative frames. In each pair, raw RGB is left and the same view with projected Gaussian centres is right. The cloud is compact and correctly located near the target. Vertical spread follows the unmeasured $\pm 25^{\circ}$ FLIS aperture under a mostly horizontal orbit.

Explore similar work

ISOPoT: Imaging Sonar Odometry by Point Tracking
Jun 22, 2026Jaša Samec, Vid Rijavec, Marko Peljhan +4Forward-Looking Sonar Inertial Navigation Systems
Faculty of Computer and Information Science, University of Ljubljana, Večna pot 113, 1000 Ljubljana, Slovenia · OZON Research Group, TIMTEC, Goriska cesta 6c, Vipava, 5271, Slovenia · MAT Systemics, University of California, Santa Barbara, 3309 Phelps Hall, Santa Barbara, 93106-6065, California, United States of America +1
Odometry-Aided Real-Time Mapping for Underwater Robots Using Forward-Looking Sonar
Sep 21, 2026Siyuan Du, Kanzhong Yao, Youdong Wang +5Forward-Looking Sonar Underwater Robots
Institute of Artificial Intelligence (TeleAI), China Telecom, China. · Tongji University, Shanghai, China.
DIVO: Continuous-time DVL-Inertial-Visual Odometry for Unmanned Underwater Vehicles
Jul 6, 2026Kyungmin Jung, Angad Bajwa, Junha Yoo +2Autonomous Underwater Vehicles Underwater Object Detection
Department of Mechanical Engineering, McGill University, Montreal, QC, Canada

Jun 22, 2026, cs.ROJ/K move · Enter open · S save

ISOPoT: Imaging Sonar Odometry by Point Tracking

Jaša Samec, Vid Rijavec, Marko Peljhan, Aleksander Grm+3

Faculty of Computer and Information Science, University of Ljubljana, Večna pot 113, 1000 Ljubljana, Slovenia · OZON Research Group, TIMTEC, Goriska cesta 6c, Vipava, 5271, Slovenia · MAT Systemics, University of California, Santa Barbara, 3309 Phelps Hall, Santa Barbara, 93106-6065, California, United States of America+1

Forward-Looking Sonar Inertial Navigation Systems Keypoint Detection

Open paper AskPDF arXiv

Reliable navigation in underwater environments remains a key challenge in marine robotics. In such scenarios, forward-looking sonars are a natural choice for long-range perception, offering wide coverage even in turbid, low-visibility conditions. However, sonar images are inherently noisy, contain artifacts, and lack rich semantic structure, causing standard computer vision methods for keypoint detection and matching to perform poorly. In this paper, we introduce ISOPoT, an imaging sonar odometry method based on modern point tracking techniques. We propose a sonar odometry pipeline that uses multi-frame point tracks as its primary correspondence representation, augmented with lightweight optimizations to improve robustness. We evaluated the proposed method on the Aracati 2017 dataset, as well as on an internal sonar dataset collected in real-world underwater environments. Our results show that ISOPoT outperforms previous state-of-the-art methods consistently in both sonar-only scenarios and in multi-sensor settings.

Abstract

Explore similar work

ISOPoT: Imaging Sonar Odometry by Point Tracking

Odometry-Aided Real-Time Mapping for Underwater Robots Using Forward-Looking Sonar

DIVO: Continuous-time DVL-Inertial-Visual Odometry for Unmanned Underwater Vehicles