Prompt-based test-time real image dehazing: a novel pipeline

(ECCV 2024)

Zixuan Chen, Zewei He, Ziqian Lu, Xuecheng Sun, Zheming Lu

Zhejiang University

📣 News

• 2024.10.27: 🔥🔥🔥 Implementations of AECRNet-PTTD is available now.
• 2024.7.2: 🎉🎉🎉 Accepeted by ECCV 2024
• 2023.09.29: Arxiv version of the paper are available now.

Getting Started

Environment

1. Clone this repo:

git clone https://github.com/cecret3350/PTTD-Dehazing.git
cd PTTD-Dehazing/

2. Create a new conda environment and install dependencies:

conda create -n pytorch_2_1_1 python=3.8
conda activate pytorch_2_1_1
conda install pytorch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 pytorch-cuda=12.1 -c pytorch -c nvidia

3. Build DCNv2 for AECRNet

cd model/DCNv2
rm -rf build
python setup.py build develop

Inference

Since our proposed PTTD is a test-time dehazing method, we do not need any additional training procedures. Your can run the following command to perform PTTD based on pre-trained AECRNet.

python pttd.py --source RIDCP --model AECRNet --pretrained_path ./pretrained_checkpoints --ys ./ys/0543.jpg --input input_samples/ --output results

You can use --save_all to save the dehazed results generated by the original pre-trained model and other intermediate results generated during PTTD-Dehazing.

You can also try other $y_s$ from the GTs of the source dataset.

Evaluation

We also provide scripts for performance evaluation on NTIRE datasets (resized O-HAZE dataset and I-HAZE dataset can be downloaded from Zero-Restore).

python pttd.py --source RIDCP --model AECRNet --pretrained_path ./pretrained_checkpoints --ys ./ys/0543.jpg --input <PATH_TO_OHAZE_DATASET_INPUT> --output results_OHAZE
python val.py --result_dir results_OHAZE --gt_dir <PATH_TO_OHAZE_DATASET_GT>

🤟 Citation

If you find our paper and repo are helpful for your research, please consider citing:

@article{chen2023promptbased,
      title={Prompt-based test-time real image dehazing: a novel pipeline}, 
      author={Zixuan Chen and Zewei He and Ziqian Lu and Zhe-Ming Lu},
      booktitle = {European Conference on Computer Vision (ECCV)},
      year = {2024}
}

Contact

If you have any questions or suggestions about our paper and repo, please feel free to contact us via zxchen@zju.edu.cn or zeweihe@zju.edu.cn.