DID-MDN
Density-aware Single Image De-raining using a Multi-stream Dense Network
[Paper Link] (CVPR'18)
We present a novel density-aware multi-stream densely connected convolutional neural network-based algorithm, called DID-MDN, for joint rain density estimation and de-raining. The proposed method enables the network itself to automatically determine the rain-density information and then efficiently remove the corresponding rain-streaks guided by the estimated rain-density label. To better characterize rain-streaks with dif- ferent scales and shapes, a multi-stream densely connected de-raining network is proposed which efficiently leverages features from different scales. Furthermore, a new dataset containing images with rain-density labels is created and used to train the proposed density-aware network.
@inproceedings{derain_zhang_2018,
title={Density-aware Single Image De-raining using a Multi-stream Dense Network},
author={Zhang, He and Patel, Vishal M},
booktitle={CVPR},
year={2018}
}
Prerequisites:
- Linux
- Python 2 or 3
- CPU or NVIDIA GPU + CUDA CuDNN (CUDA 8.0)
Installation:
-
Install PyTorch and dependencies from http://pytorch.org (Ubuntu+Python2.7) (conda install pytorch torchvision -c pytorch)
-
Install Torch vision from the source. (git clone https://github.com/pytorch/vision cd vision python setup.py install)
-
Install python package: numpy, scipy, PIL, pdb
Demo using pre-trained model
python test.py --dataroot ./facades/github --valDataroot ./facades/github --netG ./pre_trained/netG_epoch_9.pth Â
Pre-trained model can be downloaded at (put it in the folder 'pre_trained'): https://drive.google.com/drive/folders/1VRUkemynOwWH70bX9FXL4KMWa4s_PSg2?usp=sharing
Pre-trained density-aware model can be downloaded at (Put it in the folder 'classification'): https://drive.google.com/drive/folders/1-G86JTvv7o1iTyfB2YZAQTEHDtSlEUKk?usp=sharing
Pre-trained residule-aware model can be downloaded at (Put it in the folder 'residual_heavy'): https://drive.google.com/drive/folders/1bomrCJ66QVnh-WduLuGQhBC-aSWJxPmI?usp=sharing
Training (Density-aware Deraining network using GT label)
python derain_train_2018.py --dataroot ./facades/DID-MDN-training/Rain_Medium/train2018new --valDataroot ./facades/github --exp ./check --netG ./pre_trained/netG_epoch_9.pth.
Make sure you download the training sample and put in the right folder
Density-estimation Training (rain-density classifier)
python train_rain_class.py --dataroot ./facades/DID-MDN-training/Rain_Medium/train2018new --exp ./check_class
Testing
python demo.py --dataroot ./your_dataroot --valDataroot ./your_dataroot --netG ./pre_trained/netG_epoch_9.pth
Reproduce
To reproduce the quantitative results shown in the paper, please save both generated and target using python demo.py into the .png format and then test using offline tool such as the PNSR and SSIM measurement in Python or Matlab. In addition, please use netG.train() for testing since the batch for training is 1.
Dataset
Training (heavy, medium, light) and testing (TestA and Test B) data can be downloaded at the following link: https://drive.google.com/file/d/1cMXWICiblTsRl1zjN8FizF5hXOpVOJz4/view?usp=sharing
License
Code is under MIT license.
Acknowledgments
Great thanks for the insight discussion with Vishwanath Sindagi and help from Hang Zhang