README.md

RAGNet

• The implementation of "Two-Stage Single Image Reflection Removal with Reflection-Aware Guidance".

Prerequisites

• The code has been test on a PC with following environment
  • Ubuntu 18.04
  • Python 3.7.5
  • PyTorch 1.2.0
  • cudatoolkit 10.0
  • NVIDIA RTX 2080Ti

Datasets

Training datasets

• Synthetic: 7643 images from Pascal VOC dataset, image choices and the data synthesis protocol are same as ERRNet.
• Real: 90 real-world images from Berkeley real dataset.

Testing datasets

• Real20: 20 real testing images from Berkeley real dataset.

• Real45: 45 real testing images from CEILNet dataset.

• SIR dataset: three sub-datasets (Solid, Postcard, Wild) from SIR dataset.

  We provide Real20 and Real45 in ./testsets folder, the SIR dataset is not provided due to their policy, download here and put it under ./testsets folder. Please organize the SIR dataset according to our code implementation.

Test

• Download our pre-trained model and put the pretrain.pth into ./checkpoint folder

Test with the pre-trained model

$ cd RAGNet
$ python test.py

Train

• Download the vgg19-pretrained model and put it into ./checkpoint folder.
• Organize the training dataset according to our code implementation, i.e.,

$ cd synthetic
$ mkdir transmission_layer
$ mkdir blended

Start training

$ cd RAGNet
$ python train.py