Table of Content

1. Introduction
2. Datasets
3. Trained model
4. Getting Started
   • Requirements
   • Usage Example
5. Generated images

Introduction

This is the PyTorch implementation of our paper entitiled "Single Image Dehazing via Semi-Supervised Domain Translation and Architecture Search "

Datasets

Training sets：

We adopted the same training set as DA_dehazing. All the images come from RESIDE dataset. Including 3000 synthetic hazy images in the Outdoor Training Set (OTS), 3000 synthetic hazy images in the Indoor Training Set (ITS), and 1000 real-world hazy images in the Unannotated Real Hazy Images set (URHI)

• DA_DATASETS.

Testing sets：

We use four testing sets to evaluate our method:

• Benchmarking Single Image Dehazing and Beyond SOTS-OD
• O-HAZE:A Dehazing Benchmark with Real Hazy and Haze-free Outdoor Images O-HAZE
• Densehaze:A Benchmark for Image Dehazing with Dense-haze and Haze-free Images DENSE-HAZE
• NH-HAZE: An Image Dehazing Benchmark with Non-homogeneous Hazy and Haze-free Images NH-HAZE

Trained model

• The trained model of fusion dehazing network is at Google drive: Checkpoint

Getting Started

Requirements

1. Python 3.7
2. PyTorch 1.8.0
3. CUDA 9.1
4. Ubuntu 20.04

Usage

Clone the repo

git clone https://github.com/jklp2/SID_Semi-Supervised_Domain_Translation.git
cd SID_Semi-Supervised_Domain_Translation

Download the trained checkpoints Put your hazy images in the input folder, and run:

python test.py --model cra_unrolled_final --resume --ckpt_path $CKPT_PATH

where $CKPT_PATH denotes the path of the checkpoints. The results will be saved in the output folder.

Examples of dehazed images and pseudo-real hazy images

SOTS-OD

Download link

O-HAZE

Download link

DENSE-HAZE

Download link

NH-HAZE

Download link

Serveral S2R Results: Examples of pseudo-real hazy images generated in the S2R task.