DehazePhys-NeRF: Neural Radiance Fields for Dehazing With Physical Priors

Area of Science:

• Computer Vision
• Computer Graphics
• Photorealistic Rendering

Background:

• Neural Radiance Fields (NeRFs) excel at 3D scene reconstruction but struggle with adverse conditions like haze due to scattering and attenuation.
• Existing dehazing NeRF methods often suffer from imprecise disentanglement of clear content and artifacts from unreliable priors.

Purpose of the Study:

• To develop a novel framework, DehazePhys-NeRF, that addresses the limitations of current NeRFs in hazy environments by leveraging physical priors.
• To achieve high-fidelity 3D scene reconstruction and novel view synthesis in challenging atmospheric conditions.

Main Methods:

• Introduced Particle Existence Probability (PEP) to model spatial distributions of surfaces and scattering media within the NeRF volume.
• Derived a physics-based rendering equation for hazy scenes and a formalized expression for the clear scene.
• Developed the Physical Consistency-Guided Depth Compensation Network (PCDC-Net) and Semantic Dark Channel Prior (SDCP) for improved parameter estimation and depth consistency.

Main Results:

• Demonstrated significant enhancement in the fidelity and realism of clear-scene rendering compared to existing methods.
• Experimental results on multiple hazy scenes validated the superior performance of DehazePhys-NeRF.
• Achieved more accurate disentanglement of clear scene content and reduced artifacts.

Conclusions:

• DehazePhys-NeRF provides a robust, physically grounded approach for 3D scene reconstruction in hazy conditions.
• The framework successfully integrates physical priors for improved dehazing and scene representation, expanding NeRF applicability.
• The proposed method offers a significant advancement for realistic rendering in adverse weather scenarios.