SonarSplat: Novel View Synthesis of Imaging Sonar via Gaussian Splatting.

IEEE Robotics and Automation Letters 2025

Advaith V. Sethuraman¹, Max Rucker¹, Onur Bagoren¹, Pou-Chun Kung¹, Nibarkavi N.B. Amutha¹, and Katherine A. Skinner¹

¹Department of Robotics, University of Michigan, Ann Arbor, MI, USA.

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 lake environment. Compared to the state-of-the-art, SonarSplat offers improved image synthesis capabilities (+3.2 dB PSNR) and more accurate 3D reconstruction (77% lower Chamfer Distance). We also demonstrate that SonarSplat can be leveraged for azimuth streak removal.

Installation

Create a conda environment with Python 3.10:

conda create -n sonarsplat python=3.10 -y
conda activate sonarsplat

Dependency: Please install Pytorch first.

Install the required dependencies:

git submodule update --init --recursive
pip install -e . 

pip install -r requirements.txt

Data

Download the released dataset from:

dataset/
├── basin_horizontal_empty1/
├── basin_horizontal_infra_1/
└── monohansett_3D/
    ├── Data/
    ├── sonar_images/
    ├── bounds.txt
    ├── Config.json
    └── gt.ply (only in 3D datasets)

Data contains .pkl files with sensor pose (data['PoseSensor']) and sonar images (data['ImagingSonar']). sonar_images/ is used primarily for debugging/choosing subsequences. bounds.txt has the 3D bounds of the scene.

Experiments

SonarSplat allows for novel view synthesis and 3D reconstruction. We have provided some training commands with parameters to reproduce some of our experiments from the paper.

Novel View Synthesis

Training

Train the model for novel view synthesis on the infra_360_1 dataset:

bash scripts/run_nvs_infra_360_1.sh --data_dir <data_dir> --results_dir <results_dir>

Evaluation

We evaluate SonarSplat and baselines using PSNR, SSIM, LPIPS. Please organize your rendered images like so:

root_dir/
├── baseline1/
├── baseline2/
└── sonarsplat/
    ├── scene1/
    └── scene2/
        └── sonar_renders/
            ├── gt_sonar_images/
            │   ├── 0000.png
            │   ├── 0001.png
            │   └── ...
            └── sonar_images/
                ├── 0000.png
                ├── 0001.png
                └── ...

Run:

python examples/evaluate_imgs.py --root_folder <root_folder>

The script will save a .csv file conveniently listing all metrics compared to baselines.

In the event that you want to validate all methods use the same GT images, run:

python examples/evaluate_imgs.py --validate_only

3D Reconstruction

Training

Train the model for 3D reconstruction on the Monohansett Shipwreck dataset:

bash scripts/run_3D_monohansett.sh --data_dir <data_dir> --results_dir <results_dir>

Evaluation

First, we will need to convert the splat into a mesh:

python mesh_gaussian.py --ply_path <results_dir>/renders/output_step<iter>.ply 

Note: you must perform ICP/manual alignment to get the GT and predicted mesh in the same frame first. Then, organize your meshes in this structure. Please copy over the GT meshes from concrete_piling_3D and monohansett_3D:

root_dir/
├── gt/
│   ├── <scene1>/
│   │   └── gt.ply
│   ├── <scene2>/
│   │   └── gt.ply
│   └── ...
├── preds/
│   ├── baseline1/
│   │   ├── <scene1>/
│   │   │   └── pred.ply
│   │   ├── <scene2>/
│   │   │   └── pred.ply
│   │   └── ...
│   ├── baseline2/
│   │   ├── <scene1>/
│   │   │   └── pred.ply
│   │   ├── <scene2>/
│   │   │   └── pred.ply
│   │   └── ...
│   └── sonarsplat/
│       ├── <scene1>/
│       │   └── pred.ply
│       ├── <scene2>/
│       │   └── pred.ply
│       └── ...

You can compute metrics with the following command:

python scripts/compute_pcd_metrics_ply.py --gt_root <root_dir>/gt --pred_root <root_dir>/preds

Citation

@ARTICLE{11223217,
  author={Sethuraman, Advaith V. and Rucker, Max and Bagoren, Onur and Kung, Pou-Chun and Amutha, Nibarkavi N.B. and Skinner, Katherine A.},
  journal={IEEE Robotics and Automation Letters}, 
  title={SonarSplat: Novel View Synthesis of Imaging Sonar via Gaussian Splatting}, 
  year={2025},
  volume={10},
  number={12},
  pages={13312-13319},
  keywords={Sonar;Three-dimensional displays;Azimuth;Imaging;Acoustics;Rendering (computer graphics);Reflectivity;Neural radiance field;Robots;Covariance matrices;Mapping;deep learning for visual perception;marine robotics},
  doi={10.1109/LRA.2025.3627089}}

License

SonarSplat is licensed under CC BY-NC-SA 4.0