Aeye-coder/Supervised-Fog-Removal-Dataset

Supervised Fog Removal Dataset Overview

This dataset contains 80,000 paired images designed for supervised image dehazing / fog removal tasks. Each sample consists of:

a clean image (ground truth)

a synthetically fogged version of that image

The fog is generated using a physics-inspired atmospheric scattering model combined with depth estimation, allowing the fog to behave realistically with respect to scene geometry.

Unlike simple uniform haze overlays, this dataset simulates depth-aware fog, where distant objects receive stronger fog while nearby regions remain clearer.

The dataset is intended for training and evaluating deep learning models for fog removal and visibility restoration.

Dataset Structure

The dataset is split into two subsets:

train/ validation/

Each row contains two images:

{ "clean": original clear image "foggy": synthetically generated fog version }

Total dataset size:

Train samples: ~70,000 Pairs Validation samples: ~10,000 Pairs

Both images share the same resolution and alignment, making them directly usable for supervised learning.

Fog Generation Method

Fog was generated using a depth-aware atmospheric scattering model.

The process follows three main stages:

1. 1.Depth Estimation

Depth maps are estimated using MiDaS (DPT-Large), a state-of-the-art monocular depth estimation model.

The predicted depth allows the algorithm to determine which parts of the image should contain stronger fog.

1. 1.Atmospheric Light Estimation

To simulate realistic atmospheric conditions, the Dark Channel Prior (DCP) method is used to estimate the airlight component.

This helps determine how light scatters through fog.

1. 1.Fog Synthesis

Fog is applied using a modified atmospheric scattering model:

I_fog(x) = I(x) t(x) + A (1 - t(x))

Where:

I(x) → original pixel intensity

A → atmospheric light

t(x) → transmission map dependent on depth

The transmission is computed as:

t(x) = exp(-β * depth)

Multiple fog densities are simulated using randomly sampled β values.

Additional Realism Enhancements

Several modifications were added to produce more natural fog:

Non-linear depth scaling to exaggerate distance fog

Horizon brightening to simulate atmospheric scattering

Foreground preservation, keeping near objects clearer

Random noise injection for texture realism

Random fog density jitter to increase variation

These steps help avoid the overly uniform haze commonly seen in synthetic datasets.

Fog Density Parameters

Fog intensity is controlled using the following β values:

β ∈ {0.04, 0.05, 0.065, 0.075}

Random jitter of ±10% is applied per image to further diversify fog density.

Example Usage

Load the dataset directly from Hugging Face:

from datasets import load_dataset

ds = load_dataset("Aeye-coder/Supervised-Fog-Removal-Dataset")

sample = ds["train"][0]

cleanimage = sample["clean"] foggyimage = sample["foggy"] Intended Use

This dataset can be used for:

training image dehazing networks

benchmarking fog removal models

research in visibility restoration

supervised image-to-image translation tasks

Limitations

Although the fog is physically inspired, it is still synthetic. Real-world fog can exhibit additional effects such as:

dynamic scattering

weather variations

sensor noise

illumination changes

Therefore models trained on this dataset may still require domain adaptation for real-world environments.

Citations Please cite the following foundational works alongside this repository:

@inproceedings{yu2020bdd100k, title={Bdd100k: A diverse driving dataset for heterogeneous multitask learning}, author={Yu, Fisher and others}, booktitle={CVPR}, year={2020} }

@misc{mdfahimbinamin_2023, title={100K Vehicle Dashcam Image Dataset}, url={https://kaggle.com}, journal={Kaggle}, author={Md. Fahim Bin Amin}, year={2023} } License This dataset is a derivative work built using images from an Apache 2.0-licensed Kaggle dataset, which itself is based on BDD100K-derived dashcam imagery. All original source rights remain with their respective owners.

The NoFog-80K dataset includes additional processing, fog synthesis, pairing, and curation performed by the author. It is shared for research purposes under the Apache 2.0 license with appropriate attribution to all upstream data sources.