Official code for ICME2021(oral) paper: Contrastive Feature Decomposition for Image Reflection Removal.

​ The overall framework of our proposed contrastive feature decomposition method for image reflection removal.

The model is built in PyTorch 1.2.0 and tested on Ubuntu 16.04 environment (Python3.7, CUDA9.0, cuDNN7.5).

You can installation the environment via the following:

pip install -r requirements.txt

• Training dataset: synthetic and real dataset from PLNet.

• Test dataset: test sets of $\rm SIR^2$ (divided into three datasets: Solid, Wild, Postcard) and Real20.

Please have yours real and synthetic data paths set up correctly according to option.py.

CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 train.py

Please have yours test data paths set up correctly according to option.py.

Download the pre-trained model here.

python test.py

If you find this work useful for your research, please cite:

@inproceedings{Feng2021CFDNet,
    title={Contrastive Feature Decomposition for Image Reflection Removal},
    author={Xin Feng, Haobo Ji, Bo Jiang, Wenjie Pei, Fanglin Chen, and Guangming Lu},
    booktitle={IEEE International Conference on Multimedia and Expo (ICME)},
    year={2021}
}