Abstract
Neural Radiance Fields (NeRFs) have achieved significant advancements in 3D scene reconstruction and novel view synthesis, with applications increasingly spanning both academic research and industrial practice. However, while NeRF excels under ideal conditions, challenges such as airlight scattering and attenuation complicate reconstruction in adverse weather, such as haze, thereby limiting its broader applicability. Although existing dehazing NeRF methods have made progress by incorporating prior information, they remain hindered by the imprecise disentanglement of clear scene content and artifacts induced by unreliable priors. To address these challenges, we propose DehazePhys-NeRF, a framework grounded in physical priors. We first introduce the Particle Existence Probability (PEP) to explicitly model the spatial distributions of object surfaces and scattering media within the NeRF volume. Based on this formulation, we derive a rendering equation specifically for hazy scenes and establish a formalized expression for the underlying clear scene. To ensure accurate disentanglement, we systematically analyze the physical relationship between hazy and clear observations. Furthermore, we introduce the Physical Consistency-Guided Depth Compensation Network (PCDC-Net) to improve depth-dependent physical consistency modeling and propose a Semantic Dark Channel Prior (SDCP) to refine parameter estimation. By integrating these components with depth priors and structural regularization, our method significantly enhances the fidelity and realism of clear-scene rendering. Experimental results across multiple hazy scenes validate the significant performance improvements of the proposed DehazePhys-NeRF.