3D-UIR: 3D Gaussian for Underwater 3D Scene Reconstruction via Physics-Based Appearance-Medium Decoupling

Area of Science:

• Computer Vision
• Computer Graphics
• Underwater Imaging

Background:

• Underwater scenes present complex light-media interactions, including scattering and absorption, challenging conventional rendering.
• Existing methods like 3D Gaussian Splatting (3DGS) produce artifacts in inhomogeneous underwater environments due to scattering media.

Purpose of the Study:

• To develop a physics-based framework for high-quality novel view synthesis and physically accurate scene restoration in underwater environments.
• To address the limitations of current methods in handling inhomogeneous underwater media and improve geometric fidelity.

Main Methods:

• A physics-based framework using tailored Gaussian modeling to disentangle object appearance from water medium effects.
• Introduction of appearance embeddings for explicit medium representations (backscatter, attenuation) to enhance scene consistency.
• A depth-guided optimization strategy using pseudo-depth maps with regularization and scale penalty terms for improved geometric fidelity.

Main Results:

• Achieved high-quality novel view synthesis and physically accurate scene restoration by integrating appearance, medium modeling, and an underwater imaging model.
• Demonstrated significant improvements in rendering quality and restoration accuracy compared to existing methods.
• Successfully addressed artifacts and inconsistencies caused by scattering media in underwater environments.

Conclusions:

• The proposed physics-based framework effectively handles challenging underwater light-media interactions for improved 3D scene reconstruction.
• The integration of appearance embeddings and depth-guided optimization leads to enhanced scene consistency and geometric accuracy.
• This work advances the state-of-the-art in underwater computer vision and graphics, offering robust solutions for complex aquatic environments.