SonarSplat: Novel View Synthesis of Imaging Sonar via Gaussian Splatting
Advaith V. Sethuraman 1 Max Rucker 1 Onur Bagoren 1 Pou-Chun Kung 1 Nibarkavi N.B. Amutha 1 Katherine A. Skinner 1 · 2025
Read SonarSplat: Novel View Synthesis of Imaging Sonar via Gaussian Splatting by Advaith V. Sethuraman 1 Max Rucker 1 Onur Bagoren 1 Pou-Chun Kung 1 Nibarkavi N.B. Amutha 1 Katherine A. Skinner 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-00159.txt.
Introduction
Advaith Sethuraman and co-authors render fresh views from imaging sonar with a splatting method in a 2025 paper. Underwater robots map with sonar, whose grainy fan images defy photo-style tools. Their SonarSplat scatters the scene into soft 3D blobs tuned for sound, not light.
Each blob carries echo strength and saturation traits, and a custom renderer draws true range-and-bearing views. A new trick models the sideways streaking that smears sonar images. Tank and river data from a robot showed crisper new views and far finer 3D shapes than prior art. The same model can also strip streaks from raw sonar.
Catalog record
| CREATOR | Advaith V. Sethuraman 1 Max Rucker 1 Onur Bagoren 1 Pou-Chun Kung 1 Nibarkavi N.B. Amutha 1 Katherine A. Skinner 1 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-00159
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
SonarSplat: Novel View Synthesis of Imaging Sonar via Gaussian Splatting Advaith V. Sethuraman 1 Max Rucker 1 Onur Bagoren 1 Pou-Chun Kung 1 Nibarkavi N.B. Amutha 1 Katherine A. Skinner 1 Abstract In this paper, we present SonarSplat, a novel Gaussian splatting framework for imaging sonar that demonstrates realistic novel view synthesis and models acoustic streaking phenomena. Our method represents the scene as a set of 3D Gaussians with acoustic reflectance and saturation properties. We develop a novel method to efficiently rasterize Gaussians to produce a range/azimuth image that is faithful to the acoustic image formation model of imaging sonar. In particular, we develop a novel approach to model azimuth streaking in a Gaussian splatting framework. We evaluate SonarSplat using real-world datasets of sonar images collected from an underwater robotic platform in a controlled test tank and in a real-world river environment. Compared to the state-of-the-art, SonarSplat offers improved image synthesis capabilities (+3.2 dB PSNR) and more accurate 3D reconstruction (52% lower Chamfer Distance). We also demonstrate that SonarSplat can be leveraged for azimuth streak removal. I Introduction Acoustic sensors, such as imaging sonar, are commonly used for infrastructure inspection, large-area mapping, and target detection in underwater environments [ 12 , 39 , 24 ] . Unlike optical sensors, which are severely range-limited due to water column effects on light propagation, acoustic sensors can capture data at long ranges to provide critical information about subsea environments. Although sonar exhibits many desirable qualities such as longer range, invariance to lighting conditions, and the ability to discern material properties, there exist acoustic phenomena, including elevation ambiguity, azimuth streaking, and multi-path reflections, which make sonar interpretation difficult for operators and computer vision algorithms alike. Fig. 1: We present SonarSplat, a novel 3D Gaussian splatting framework that enables novel view synthesis of realistic sonar images, image de-streaking, and 3D reconstruction. We model the acoustic reflectivity and azimuth streak probabilities for all Gaussians, yielding a 3D scene representation that can be queried for a variety of downstream tasks. Radiance field methods offer a promising approach to addressing these challenges. Neural radiance fields (NeRFs) have demonstrated the potential for high-fidelity data synthesis, denoising, and 3D reconstruction using optical imagery [ 26 , 45 , 30 , 5 ] . The benefits of neural rendering have been recognized by the underwater perception community, with prior work exploring neural rendering for 3D object reconstruction using underwater cameras [ 18 , 41 ] and sonar data [ 33 , 34 , 50 , 6 , 49 ] . While these innovations are promising, training and processing NeRFs is costly and time-consuming, making deploying them in real-time or on resource-constrained devices difficult. More recently, 3D Gaussian splatting (3DGS) [ 14 ] was developed as a faster alternative to NeRF. Gaussian splatting has been leveraged for underwater applications as well [ 51 , 19 , 27 ] . Most relevant to our proposed work, ZSplat proposes a Gaussian splatting framework for RGB-sonar fusion [ 35 ] . However, ZSplat leverages the fusion of sonar data to improve the rendering of RGB images and does not focus on enabling high-fidelity data synthesis for sonar imagery or provide evaluation for the quality of rendered sonar images or 3D reconstruction. Thus, there is a clear gap in the literature for a framework capable of efficient and effective sonar image synthesis and 3D reconstruction using Gaussian splatting. In this paper, we propose SonarSplat ( Fig. 1 ), a novel Gaussian splatting framework for imaging sonar that enables efficient and high quality novel view synthesis, sonar image de-streaking, and 3D reconstruction for underwater robotic applications. To the best of our knowledge, SonarSplat is the first sonar-only method that leverages Gaussian splatting to perform these tasks. Our main contributions are summarized as follows: • We develop a novel 3D Gaussian splatting framework for rasterization of range/azimuth sonar images from known and novel viewpoints. • We develop a novel method for learning the probability of azimuth streaking in a differentiable manner, which allows for removal of unwanted azimuth streaking artifacts from rendered images. • We introduce a novel densification strategy called the Elevation Sampling Densification Strategy (ESDS) that places Gaussians on elevation arcs and yields better novel view synthesis and 3D reconstruction. • We perform quantitative and qualitative evaluation of our method on real imaging sonar data collected from a robotic platform with a variety of sonar ranges and environments. Code and data will be made publicly available on a project website after review. II Related Work Fig. 2: Overview of SonarSplat. Our method takes as input a sensor pose and an initial set of 3D Gaussians representing the scene. Then, we transform the 3D Gaussians into the sensor’s image space. We splat the reflectivity parameter \nu_{k} to get the unsaturated image I_{u} . Additionally, we optimize and splat per-Gaussian azimuth streaking probabilities p_{k} . All of the probabilities in a range bin r_{i} are considered in our novel Adaptive Gain Module, which adjusts the receiver gain applied to I_{u} . Finally, we produce \hat{I} by multiplying the gain A by I_{u} . All parameters are optimized using gradient descent by taking losses with respect to the sonar image pixel values. All sonar images shown are polar (range/azimuth) coordinates. II-A Radiance Field Methods NeRFs represent scenes implicitly using neural networks and have been explored for a variety of robotic applications, including novel view synthesis, data generation, Simultaneous Localization and Mapping (SLAM), manipulation, image dehazing, and trajectory optimization [ 26 , 2 , 11 , 37 , 43 , 5 , 7 ] . Despite their success as a scene representation, a key challenge for the practical application of NeRFs is the long training time and the slow rendering speeds, primarily due to the computationally expensive volumetric rendering process [ 26 ] . A recent development in radiance field methods is Gaussian splatting, which uses more efficient rasterization instead of volumetric rendering to optimize and produce a scene representation, enabling usage in a variety of real-time applications [ 14 , 23 , 22 , 25 ] . Despite their impressive performance and rapid adoption, radiance field methods were developed with optical sensors in mind, limiting their direct application to non-visual sensors. Recent developments have integrated the physics-based image formation models of various non-visual sensors, including radar, LIDAR, sonar, thermal cameras, and event cameras into NeRF and Gaussian splatting frameworks for various vision tasks [ 3 , 4 , 9 , 33 , 20 ] . We are inspired by the faster rendering speeds and high fidelity of Gaussian splatting methods and propose a novel Gaussian splatting framework for imaging sonar. II-B Perception Challenges for Imaging Sonar Imaging sonar is a popular underwater sensor for perception tasks like object detection, segmentation, reconstruction, and inspection [ 33 , 8 , 12 , 44 , 28 ] . However, the sensor is susceptible to multi-path effects, nonlinear noise patterns, and phenomena like azimuth streaking [ 31 ] . Azimuth streaking, in particular, can make sonar image understanding difficult by suppressing returns from certain azimuth angles, causing black bars in the sonar image [ 31 ] . In addition, the sensor inherently loses information from 3D into 2D when capturing the acoustic returns,