Synthetic Fog Simulation

Updated 15 April 2026

Synthetic fog simulation is the algorithmic generation of fog effects in visual data using physics-based models like Koschmieder’s law and volumetric radiative transfer.
Implementation methods include 2D image compositing, 3D scene reconstruction, active sensor simulation, and neural rendering with adversarial losses.
Applications span robust training for perception, defogging evaluation, and improved detection metrics, with gains observed in IoU, PSNR, and mAP.

Synthetic fog simulation refers to the algorithmic generation of fog effects in sensor data or rendered images, enabling controlled adverse-weather augmentation for vision, graphics, and autonomous systems research. Modern synthetic fog simulation leverages physically based rendering, radiative transfer, sensor modeling, deep learning, and high-fidelity numerical solvers to create photo-realistic or task-oriented data for training and evaluation. Approaches span 2D image-based compositing, 3D scene reconstruction with volumetric integration, neural rendering, and physically-validated simulation targeted at both optical and active sensors.

1. Physical and Mathematical Models for Fog Synthesis

Rigorous fog simulation is rooted in the physics of atmospheric light scattering and radiative transfer. The dominant model for homogeneous terrestrial fog is Koschmieder’s law, which describes pixel-wise foggy radiance as a convex combination of the clear scene intensity and a global atmospheric light term, weighted by a distance-dependent transmittance: $I(x) = R(x)\,t(x) + L\,[1-t(x)]$ where $t(x) = \exp(-\beta\,\ell(x))$ , with $\beta > 0$ the scattering coefficient and $\ell(x)$ the scene depth at pixel $x$ (Sakaridis et al., 2018). Meteorological optical range is given by $\text{MOR} = 2.996 / \beta$ , with fog traditionally defined at $\beta \geq 2.996 \times 10^{-3} \,\mathrm{m}^{-1}$ .

For more physically comprehensive modeling, volumetric radiative transfer equations are employed, accounting for multiple scattering, in-scattering (airlight), wavelength dependence, and phase functions, leading to computationally intensive solutions such as path tracing and Monte Carlo integration. State-of-the-art pipelines (e.g., SynFog (Xie et al., 2024)) use the full radiative transfer equation (RTE) with Henyey–Greenstein phase functions and explicit global illumination: $E(d, \lambda) = E_0(\lambda) e^{-\beta(\lambda)d} + L_h(\infty, \lambda)[1-e^{-\beta(\lambda)d}]$ and even integrate secondary effects due to artificial lighting, sensor optics, noise, and image signal processors.

Active sensor fog simulation, as for LiDAR, employs the Beer–Lambert law, simulating both direct signal attenuation and distributed backscatter from fog droplets: $I(d) = I_0\,\exp(-\beta_\text{ext}\,d)$ where $\beta_\text{ext} = N\,\sigma_\text{ext}$ combines droplet density and size (Hahner et al., 2021). The received power includes both “hard” returns (attenuated surface echo) and “soft” fog returns (volumetrically integrated backscatter).

2. Algorithmic Pipelines and Implementation Frameworks

Fog synthesis pipelines differ in engineering detail by sensor modality, dataset format, and photo-realism targets:

2D Image-Based Synthesis: Given a clear image, depth (or disparity), and estimated atmospheric light, pixels are composited under the homogeneous-fog model. Transmission maps are refined via cross-bilateral or guided filtering with color or semantic cues (Sakaridis et al., 2017, Sakaridis et al., 2018). Dual-reference cross-bilateral filters utilize both color (in CIELAB) and semantic label maps, strictly blocking smoothing across object boundaries and suppressing texture-transfer artifacts in the fog mask.
3D Scene and Volumetric Methods: Recent frameworks such as 3D Gaussian Splatting (3DGS) (Sang et al., 26 May 2025, Fiebelman et al., 7 Apr 2025) reconstruct the scene as a set of spatially varying Gaussian primitives. Fog is applied in screen space through an exponential depth-based blending post-process, or more physically, as a dynamic particle system using the Material Point Method (MPM) for fog dynamics and volume rendering for radiative transfer.
Physically-Based Synthetic Datasets: Pipelines such as SynFog (Xie et al., 2024) and Foggy Cityscapes (Sakaridis et al., 2017) employ procedural scene construction, physically based path tracing with fog media, and detailed sensor-noise and ISP modeling. Annotated data includes depth, segmentation, and raw sensor output.
Active Sensing (LiDAR): Each point in a clear-air scan is reweighted according to physical attenuation models, and synthetic backscatter is injected based on particle cross-sections and expected fog visibility (MOR) levels (Hahner et al., 2021).
Neural and Adversarial Image Synthesis: Image-to-image translation architectures (e.g., AnalogicalGAN (Gong et al., 2020)) transfer the “gist” of fog—pixel-wise attenuation maps and residuals—learned on synthetic paired data to the real domain via adversarial and cycle-consistency losses, enabling zero-shot fog generation on unpaired clear images. Physically inspired constraints can be injected as auxiliary depth or perceptual losses.

The following table organizes core pipelines by methodological family:

Methodology	Input/Output	Physical Model	Refinement/Controls
Analytical 2D Compositing	RGB, depth/D, semantics → RGB	Koschmieder / ASM	Color/semantic bilateral filter; density β
Volumetric Path Tracing	3D scene, lights → RGB	Full RTE, phase fn, volumetric	Scene geometry, light, photometric sensor/ISP simulation
3DGS/Particle	Multi-view, 3DGS → RGB	Exponential screen-space or MPM+RTE	Gaussian densities, particle physics parameters
Active Sensor (LiDAR)	Point cloud → foggy cloud	Beer–Lambert law, backscatter	α (attenuation), calibrated fog visibility
GAN/Image Translation	Clear RGB (real/synth), synth fog → real fog	Learned pixel map (physically encouraged)	Adversarial/cycle losses, supervised or perceived depth

3. Fog Density Estimation, Control, and Evaluation

Realistic fog simulation requires accurate control and estimation of fog density (scattering coefficient, β) to match target conditions:

Density Control: Image pipelines sample β in the range $t(x) = \exp(-\beta\,\ell(x))$ 0 (MOR $t(x) = \exp(-\beta\,\ell(x))$ 1 m) to span from light to dense fog (Sakaridis et al., 2018, Sakaridis et al., 2017, Xie et al., 2024). For volumetric approaches, σ_s is chosen accordingly and can be adjusted at inference for continuous density variation.
Density Estimation: CNN-based fog-density estimators (e.g., an AlexNet regressor) are trained on synthetically fogged datasets, enabling annotation and curriculum adaptation of real foggy datasets (Sakaridis et al., 2018).
Qualitative Validation: Mechanical Turk fog-density ranking studies yield ~89% agreement with human judgments, attesting to simulation realism (Sakaridis et al., 2018).
Downstream Metrics: Impact is quantified via semantic segmentation mean IoU, object detection AP/mAP, and defogging metrics such as PSNR, SSIM, and no-reference DHQI (Xie et al., 2024). Synthetic fog data consistently improves performance on real foggy images across models and benchmarks.

4. Comparative Analysis: Classical vs. Modern and Domain-Adaptive Approaches

Classical synthetic fog simulators are limited by the assumptions of homogeneous fog, global atmospheric light, and sensor-agnostic imaging:

Limitations: No spatially varying fog density, no multiple scattering, global $t(x) = \exp(-\beta\,\ell(x))$ 2, and dependence on error-prone disparity depth. This can lead to blocky or over-smoothed fog at depth boundaries and artifacts in scenes with missing or noisy depth (Sakaridis et al., 2017, Sakaridis et al., 2018).
Augmentation by Semantics: Cross-bilateral filtering with semantics strictly respects instance boundaries, yielding higher-fidelity object contours (Sakaridis et al., 2018).
Full Imaging Simulation: Modern fog simulators such as SynFog incorporate end-to-end physical image simulation, including optics, sensor, and ISP, with volumetric path tracing and phase functions, validated against fog-chamber real data (Xie et al., 2024).
Neural Rendering and Inverse Modeling: ScatterNeRF introduces a disentangled volumetric NeRF-style architecture with fog and clear scene MLPs, enabling bidirectional rendering (dehazing as $t(x) = \exp(-\beta\,\ell(x))$ 3), control of fog density, and learning under physics-inspired entropy and photometric supervision (Ramazzina et al., 2023).
Domain Adaptation: Cycle-consistent adversarial and analogical architectures enable fog “style” transfer from synthetic/paired domains to real/unpaired domains. AnalogicalGAN achieves state-of-the-art mIoU on downstream semantic segmentation in real fog, surpassing physics-only and standard GAN methods (Gong et al., 2020).

5. Applications, Datasets, and Empirical Results

Synthetic fog simulation underpins progress in robust perception for autonomous driving, adverse-weather computer vision, and graphics:

Training and Benchmarking: Simulated data enables training semantic segmentation, defogging, and 3D object detection models for conditions where real foggy annotations are scarce (Sakaridis et al., 2018, Sakaridis et al., 2017, Hahner et al., 2021).
Public Datasets:
- Foggy Cityscapes-DBF: Cityscapes images rendered with dual-cross-bilateral filtered fog, β ∈ {0.005, 0.01, 0.02} (Sakaridis et al., 2018).
- SynFog: 4,000 images across three fog densities, dual lighting modes, full volumetric/sensor/ISP simulation, with multi-modal annotations (Xie et al., 2024).
- Foggy Zurich: 3,808 real foggy images, pixel-level semantic annotations for 16–40 dense fog images, used for cross-domain evaluation (Sakaridis et al., 2018, 1901.01415).
- LiDAR Foggy Pointclouds: Fogifying clear-weather point clouds for 3D detection, with visibility sampled down to 50 m MOR (Hahner et al., 2021).
Performance Gains: Curriculum Model Adaptation (CMAda) stages models through increasing fog densities, yielding a 5–6 pp absolute gain in mean IoU on dense fog test sets compared to baseline (Sakaridis et al., 2018). In LiDAR, domain-adaptive fog simulation improves 3D object detection mAP by 1–3 pp and helps maintain high recall under dense fog (Hahner et al., 2021). SynFog-trained defogging and detection networks yield +1–2 dB PSNR and higher mAP than competing synthetic datasets on real foggy imagery (Xie et al., 2024).

6. Limitations and Prospects

Key limitations include the assumption of homogeneous fog, the computational cost of volumetric path tracing, the reliance on depth estimation quality, and the challenge of scaling highly realistic simulation pipelines to city-scale data volumes (Sakaridis et al., 2017, Xie et al., 2024). Screen-space and single-scattering approximations may break down for extremely dense fog and fail to reproduce volumetric shadowing, colored multiple scattering, or inhomogeneous haze (Sang et al., 26 May 2025, Ramazzina et al., 2023).

Research is addressing these gaps via:

Volumetric particle methods (MPM (Fiebelman et al., 7 Apr 2025)) for dynamic and physically-plausible volumetric effects.
End-to-end neural architectures incorporating physics-informed losses and domain adaptation (e.g., ScatterNeRF (Ramazzina et al., 2023), AnalogicalGAN (Gong et al., 2020)).
Comprehensive sensor and ISP modeling for pipeline realism, as in SynFog (Xie et al., 2024).
Combined multi-modal simulation (joint camera+LiDAR fogifying) for sensor-fusion studies (Hahner et al., 2021).

Advances in computational efficiency, depth estimation fidelity, radiative transfer solutions, and machine learning for physical parameter estimation are likely to further improve the scope, realism, and utility of synthetic fog simulation for both simulation and learning-based perception under adverse weather conditions.