Hazy/Dusty Image Synthesis for Driving Scenarios

Paper: Toward Improving Robustness of Object Detectors against Domain Shift (IEEE GECOST 2024)

Authors: Le-Anh Tran, Chung Nguyen Tran, Dong-Chul Park, Jordi Carrabina, David Castells-Rufas

Dependencies

This repo is based on the following project/packages:

• Monodepth2
• Pytorch
• OpenCV

Setup

• Step 1: Create virtual environment:

conda create -n hazesynt python=3.6
conda activate hazesynt

• Step 2: Install required packages as in Monodepth2 or just run this command:

pip install -r requirements.txt

• Step 3: Download pre-trained model from Monodepth2 and place it in 'models/{model_name}', e.g., 'models/mono+stereo_640x192'.

Image Synthesis

Run the following command to generate synthetic image:

python main.py --image_path ./inputs --output_image_path ./outputs --model_name mono+stereo_640x192 --beta 2.0 --airlight 150

The values of beta and airlight can be changed (recommended: beta = [1.0,3.0], airlight = [150,255]).

Citation

If you feel this repo is helpful for your study, please cite our work:

@inproceedings{tran2024toward,
  title={Toward improving robustness of object detectors against domain shift},
  author={Tran, Le-Anh and Tran, Chung Nguyen and Park, Dong-Chul and Carrabina, Jordi and Castells-Rufas, David},
  booktitle={2024 International Conference on Green Energy, Computing and Sustainable Technology (GECOST)},
  pages={01--05},
  year={2024},
  organization={IEEE}
}

Have fun!

LA Tran