Research map / 3D vision / Depth and stereo
SynFog: A Photo-realistic Synthetic Fog Dataset based on End-to-end Imaging Simulation for Advancing Real-World Defogging in Autonomous Driving
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds an end-to-end physically based foggy image simulator (volumetric path tracing plus a physical camera model) and releases SynFog, a synthetic driving fog dataset that yields defogging models that generalize better to real fog than ASM-based or game-engine datasets.
Problem
Synthetic fog datasets built with the atmospheric scattering model or game engines do not match real imaging, so defogging models trained on them transfer poorly to real foggy driving images.
Approach
Simulate the full foggy imaging chain: volumetric path tracing of scene radiance with a Henyey-Greenstein phase function and global illumination, followed by a physical optics model, a calibrated cellphone sensor noise model (shot noise), and an ISP, producing photo-realistic paired clear and foggy images with sky-light and active-light conditions at three visibility levels.
Key results
- On the SynFog test set, DehazeFormer attains PSNR 27.96 and SSIM 0.923, while GridDehazeNet gets the best LPIPS of 0.163.
- AECR-Net trained on SynFog reaches PSNR 15.43 and SSIM 0.6116 on O-Haze, versus 14.46 for Foggy Cityscapes and 13.90 for Virtual KITTI, despite SynFog having fewer training pairs (1350 vs 8925 and 3452).
- Detection mAP on Foggy Zurich rises to 71.5% with SynFog-trained AECR-Net, compared with 69.7% (Foggy Cityscapes) and 68.9% (Virtual KITTI); removing sensor noise drops it to 69.5%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train AECR-Net on SynFog instead of Foggy Cityscapes | AECR-Net · Foggy Zurich | DHQI: 43.40 -> 44.46 | Helped |
| train AECR-Net on SynFog instead of Virtual KITTI | AECR-Net · Foggy Zurich | DHQI: 42.80 -> 44.46 | Helped |
| train AECR-Net on SynFog instead of Foggy Cityscapes | AECR-Net · O-Haze | PSNR: 14.46 -> 15.43 | Helped |
| train AECR-Net on SynFog instead of Foggy Cityscapes | AECR-Net · Foggy Zurich | mAP (%): 69.7 -> 71.5 | Helped |
| train AECR-Net on SynFog instead of Virtual KITTI | AECR-Net · Seeing Through Fog | mAP (%): 53.3 -> 55.5 | Helped |
| add sensor noise simulation (ablation) | SynFog training data for AECR-Net · O-Haze | PSNR: 15.16 -> 15.43 | Helped |
| add sensor noise simulation (ablation) | SynFog training data for AECR-Net · O-Haze | SSIM: 0.5795 -> 0.6116 | Helped |
| add sensor noise simulation (ablation) | SynFog training data for AECR-Net · Foggy Zurich | mAP (%): 69.5 -> 71.5 | Helped |
| replace end-to-end simulated fog with ASM-generated fog (ablation) | DehazeFormer training data · Foggy Zurich | mAP (%): 69.7 -> 59.7 | Hurt |
Surprising findings
- Models trained on the ASM-based SynFog-beta variant can underperform on Foggy Zurich detection (59.7% mAP with DehazeFormer), lower than models trained on other datasets.
- Fog-chamber color analysis shows ASM-simulated images do not follow the real per-channel color trend as fog density increases, while the proposed pipeline does.
Limitations
- Volumetric path tracing is computationally expensive, so the dataset is smaller (500 scenes, 1350 pairs) than other synthetic fog datasets.
- Only fog is simulated; other scattering media are left for future work.
- Evaluation of real-world transfer relies on qualitative comparisons and mAP on a few real datasets; no quantitative real-world defogging ground truth.
Experimental setup
Datasets: SynFog Foggy Zurich Seeing Through Fog Foggy Cityscapes Virtual KITTI 2 RESIDE O-Haze BeDDE Foggy Driving
Benchmarks: SynFog test set RESIDE O-Haze Foggy Zurich Seeing Through Fog Foggy Driving BeDDE
Metrics: PSNR SSIM LPIPS DHQI mAP
Baselines: DCP CAP GridDehazeNet MSBDN FFA-Net AECR-Net DehazeFormer Foggy Cityscapes Virtual KITTI 2 SynFog-beta (ASM-based)
Builds on: Atmospheric Scattering Model ISETCam YOLOv8
Closest papers by meaning
- SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving (CVPR 2025 · similarity 0.84)
- PRISM: Rethinking Atmospheric Scattering Reconstruction as a Unified Understanding and Restoration Model for Real-world Dehazing (NeurIPS 2026 · similarity 0.84)
- Learning Generalized Segmentation for Foggy-Scenes by Bi-directional Wavelet Guidance (AAAI 2024 · similarity 0.83)
- PEIE: Physics Embedded Illumination Estimation for Adaptive Dehazing (AAAI 2025 · similarity 0.83)
- RainyGS: Efficient Rain Synthesis with Physically-Based Gaussian Splatting (CVPR 2025 · similarity 0.82)
- Depth-Centric Dehazing and Depth-Estimation from Real-World Hazy Driving Video (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.