Gaussian splashing enables direct volumetric rendering underwater.

Area of Science:

• Computer Vision
• Robotics
• Photogrammetry

Background:

• Underwater imaging is challenging due to light scattering and occlusion.
• Existing 3D reconstruction methods like NeRFs and 3DGS perform poorly in underwater environments.
• Previous underwater adaptations of NeRFs offer high quality but are impractically slow.

Purpose of the Study:

• To develop a fast and high-quality 3D reconstruction method for underwater scenes.
• To adapt 3D Gaussian Splatting (3DGS) for effective underwater scene representation.
• To address limitations in rendering speed and reconstruction time for underwater 3D modeling.

Main Methods:

• Introduced Gaussian Splashing, unifying 3DGS speed with a scattering-aware image formation model.
• Innovated rendering and depth estimation procedures tailored for underwater conditions.
• Developed a new loss function for 3DGS to improve underwater scene reconstruction.

Main Results:

• Achieved reconstruction times in minutes, a significant improvement over previous methods.
• Enabled novel underwater scene rendering at 140 FPS, drastically increasing speed.
• Demonstrated competitive reconstruction quality across datasets, particularly for distant details affected by backscatter.

Conclusions:

• Gaussian Splashing offers a practical and efficient solution for underwater 3D reconstruction.
• The method effectively handles scattering and occlusion, improving detail preservation.
• This advancement enables faster and more detailed exploration and analysis of underwater environments.