Generation of Synthetic Non-Homogeneous Fog by Discretized Radiative Transfer Equation.

Area of Science:

• Computer Vision
• Image Synthesis
• Scientific Computing

Background:

• Realistic fog synthesis is crucial for applications like autonomous navigation, augmented reality, and visual effects.
• Existing methods often assume uniform fog and simplified scattering, limiting physical accuracy.
• Traditional techniques include Koschmieder's law and Generative Adversarial Network (GAN)-based image translation.

Purpose of the Study:

• To develop a physics-driven approach for synthesizing spatially inhomogeneous fog with anisotropic multi-scattering.
• To enhance the structural consistency and perceptual plausibility of synthetic fog effects.
• To improve the physical realism of image augmentation under challenging visibility conditions.

Main Methods:

• Discretization of the Radiative Transfer Equation (RTE) to model complex fog phenomena.
• Incorporation of spatially varying fog distributions.
• Modeling of anisotropic multi-scattering effects for enhanced realism.

Main Results:

• The proposed RTE-based method demonstrated superior performance compared to Koschmieder-based and GAN-based baselines.
• Achieved a 10% lower Fréchet Inception Distance (FID) compared to Koschmieder's law.
• Achieved a 42% lower FID compared to CycleGAN, with improved Pearson correlation.
• Generated structurally consistent and perceptually plausible fog effects.

Conclusions:

• The RTE-based fog synthesis method offers significant improvements in physical realism and visual fidelity.
• It enables more accurate image augmentation for applications requiring challenging visibility conditions.
• High memory requirements for tensor computations pose a challenge for real-time or large-scale deployment.