Deep Learning for Seeing Through Window With Raindrops

by Yuhui Quan, Shijie Deng, Yixin Chen, Hui Ji

Shijie Deng is in charge of this code repo. If you have any questions, please contact shijie.deng.cs@foxmail.com

See more details on http://csyhquan.github.io/

Environment

• Python3.5
• Tensorflow 1.9 with NVIDIA GPU

Testing

The testing checkpoints can be downloaded at ：https://pan.baidu.com/s/1Ocp-xM83s2Irts1ssd5UVQ access code：y70s

We use the dataset with rainy and clean images created by https://github.com/rui1996/DeRaindrop. Edges maps are added for attention, all pictures are under the directory /testing_real/ and the outputs are under /testing_result/. Run the code below to get the result of our model.

python run_model.py --inputdata_path ./testing_real/ --output_path ./testing_result --phase test --restore_step 258000

Quantitative results of PSNR and SSIM will be printed, you can check /testing_result/ for a qualitative evaluation.

Training

Put your training rainy images at "train_img/data/" and corresponding gt clean images at "train_img/gt/"(images should be *.jpg or *.png format).

Then run the train_list.sh to generate datalist raindrop.txt to train.

raindrop.txt should be like this, one gt image path with one rainy image path after it :

train_img/gt/0_clean.png train_img/rain/0_rain.png

Run the code below to train the model.

python run_model.py  --phase train --restore_step 0