Energy-based Domain Adaptation without Intermediate Dataset for Foggy Scene Segmentation

This work was accepted at IEEE Transactions on Image Processing (TIP). Paper

Requirements

This repository is implemented on

• Ubuntu 16.04
• Conda 4.12.0
• CUDA 11.6
• Python 3.7.15
• Pytorch 1.12.0

To install required environment for the training:

conda env create -n [your env name] -f environment.yaml
conda activate [your env name]

You need to install densetorch package. To install densetorch:

git clone https://github.com/drsleep/densetorch
cd densetorch
pip install -e .

Dataset

You can download the dataset here:

• Cityscapes: Download "leftImg8bit_trainvaltest.zip" and "gtFine_trainvaltest.zip" from the Cityscapes Dataset, and put it in the './data/Cityscapes' directory.

• Foggy Zurich: Download "Foggy Zurich.zip" from the Foggy Zurich Dataset, and put it in the './data/Foggy_Zurich' directory.

• Foggy Driving: Downlaod "Foggy Driving.zip" from the Foggy Driving Dataset, and put it in the './data/Foggy_Driving' directory.

• ACDC: Download "rgb_anon_trainvaltest.zip" and "gt_trainval.zip" from the ACDC Dataset, and put it in the './data/ACDC' directory.

Pre-trained Models

You can download pretrained model here:

• Cityscapes pre-trained model: ./Cityscapes_pretrained_model.pth

• Pre-trained DAEN: ./DAEN_1.pth

Evaluation

To evaluate DAEN on real foggy datasets:

CUDA_VISIBLE_DEVICES=[gpu_id] python evaluate.py --file-name 'DAEN_result' --restore-from './DAEN_1.pth'

Training

To train DAEN:

CUDA_VISIBLE_DEVICES=[gpu_id] python main.py --file-name 'DAEN' --restore-from './Cityscapes_pretrained_model.pth'

Acknowledgements

This code is built on FIFO. We thank the authors for sharing their codes.