U-FRRN

Fast User-Guided Single Image Reflection Removal via Edge-aware Cascaded Networks

Representitive Results

Different levels of user guidance. SG means sparse guidance and DG means dense guidance.

Qualitative Comparison

Environment Preparing

python2.7
torch==0.4.1
torchvision==0.2.0
numpy
cv2

You can install the supports with pip install -r requirement.txt. Here we only provide the GPU version code!

Dataset preparing

Prepare the background synthetic dataset SUN2012, reflection synthetic dataset VOC2012, and test dataset SIRR, then change the root in train.py.

We have already prepared several demo validation images in ./demo/, we use pictures in input as original input, and pictures in edge_R and edge_B as user-guide hints.

Training process

Download VGG pretrained model from vgg16 , and then change the pretrained_vgg16 in train.py

Pre-generate the mask which we used for synthesis training data in function main

then run the following command

python train.py

Testing process

We embed the test function in train.py, you can directly use it.

Demo

We will provide a demo soon!

If you find this work useful for you, please cite

@article{zhang2019fast,
  title={Fast User-Guided Single Image Reflection Removal via Edge-aware Cascaded Networks},
  author={Zhang, Huaidong and Xu, Xuemiao and He, Hai and He, Shengfeng and Han, Guoqiang and Qin, Jing and Wu, Dapeng},
  journal={IEEE Transactions on Multimedia},
  year={2019},
  publisher={IEEE}
}