WaterSplatting: Fast Underwater 3D Scene Reconstruction Using Gaussian Splatting

Huapeng Li, Wenxuan Song, Tianao Xu, Alexandre Elsig, Jonas Kulhanek

WaterSplatting: Fast Underwater 3D Scene Reconstruction Using Gaussian Splatting

Underwater 3D reconstruction is notoriously difficult due to the optical properties of water, which acts as a participating medium that scatters and absorbs light. Traditional methods like Neural Radiance Fields (NeRFs) struggle with slow training and rendering times, while newer techniques like 3D Gaussian Splatting (3DGS) fail to account for the volumetric nature of water. WaterSplatting addresses these challenges by combining the efficiency of 3DGS with a dedicated volumetric field to model the scattering medium, achieving real-time performance and state-of-the-art quality.

The Challenge of Underwater Imaging

Standard computer vision assumes light travels through a vacuum. In underwater environments, light interacts with particles, leading to attenuation (color shifting and fading) and backscatter (a foggy appearance). Specialized NeRFs, such as SeaThru-NeRF, can model these effects but are computationally expensive, often requiring hours to train and seconds to render a single frame. Conversely, standard 3DGS is fast but "explicit," attempting to represent water as solid particles, which results in "muddy" artifacts and inaccurate geometry.

Methodology: A Dual Representation Approach

WaterSplatting employs a hybrid architecture that decouples scene geometry from the surrounding medium.

The WaterSplatting pipeline combines 3D Gaussians for geometry with an MLP-based volumetric field for the medium.

The framework consists of four main components:

  1. 3DGS for Geometry: Explicit surfaces like rocks and coral are represented using standard 3D Gaussians.
  2. Volumetric Field for the Medium: A separate MLP-based field captures the scattering properties (attenuation and backscatter) of the water.
  3. Integrated Rendering: The renderer interleaves alpha compositing of the Gaussians with the integration of the medium along each ray. This produces a final image composed of a Direct Component (attenuated object color) and a Backscatter Component (the water's glow).
  4. Perception-Aware Loss: Optimization uses a regularized $L_1$ and D-SSIM loss function that prioritizes dark regions, aligning the reconstruction with human visual perception in low-light underwater scenes.

Experimental Results

The method was evaluated on the real-world SeaThru-NeRF dataset and a simulated foggy dataset based on Mip-NeRF 360.

Comparison of WaterSplatting (Ours) against 3DGS and SeaThru-NeRF. Our method achieves superior PSNR and real-time frame rates with significantly shorter training times.

Key findings include:

  • Efficiency: WaterSplatting achieves real-time rendering (~42-56 FPS) on an RTX 4080, a massive leap from SeaThru-NeRF's 0.07 FPS. Training is completed in 7–10 minutes, compared to the 10 hours required by previous state-of-the-art methods.
  • Reconstruction Quality: It consistently outperforms baselines in PSNR and SSIM. In the "Curaçao" scene, it reached a 32.67 PSNR, surpassing SeaThru-NeRF's 29.30.
  • Scene Restoration: By setting the medium density to zero during inference, WaterSplatting can "de-water" scenes, producing clear, color-corrected images with sharper details than competing restoration techniques.

Conclusion and Future Work

WaterSplatting demonstrates that decoupling the medium from the geometry is essential for efficient underwater reconstruction. While the system excels in most scenarios, it still faces challenges in extremely low-visibility conditions where distant objects become indistinguishable from the medium. Future work will explore scaling to larger environments and handling hybrid atmospheric-underwater scenes.

Summary updated

BibTeX
@inproceedings{li2025watersplatting,
  author = {Huapeng Li and Wenxuan Song and Tianao Xu and Alexandre Elsig and Jonas Kulhanek},
  title = {{WaterSplatting: Fast Underwater 3D Scene Reconstruction Using Gaussian Splatting}},
  booktitle = {3DV},
  year = {2025},
  url = {https://openreview.net/forum?id=Z9yn9YgNIz},
}
Figure