ERRNet

The implementation of CVPR 2019 paper "Single Image Reflection Removal Exploiting Misaligned Training Data and Network Enhancements"

News (19/09/2019): Fix the broken link; our pretrained model and collected unaligned dataset are now available at OneDrive

Highlights

• Our network can extract the background image layer devoid of reflection artifacts, as in the example:

• We captured a new dataset containing 450 unaligned image pairs that are considerably easier to collect. Image samples from our unaligned dataset are shown below:

• We introduce a simple but powerful alignment-invariant loss function to facilitate exploiting misaligned real-world training data. Finetuning on unaligned image pairs with our loss leads to sharp and reflection-free results, in contrast to the blurry ones when using a conventional pixel-wise loss (L1, L2, e.t.c.). The resulting images finetuned by different losses are shown below: (Left: Pixel-wise loss; Right: Ours)

Prerequisites

• Python >=3.5, PyTorch >= 0.4.1
• Requirements: opencv-python, tensorboardX, visdom
• Platforms: Ubuntu 16.04, cuda-8.0

Quick Start

1. Preparing your training/testing datasets

Training dataset

• 7,643 cropped images with size 224 × 224 from Pascal VOC dataset (image ids are provided in VOC2012_224_train_png.txt, you should crop the center region with size 224 x 224 to reproduce our result).

• 90 real-world training images from Berkeley real dataset

Testing dataset

• 100 synthetic testing images from CEILNet dataset (testdata_reflection_synthetic_table2)
• 20 real testing images from Berkeley real dataset.
• Three sub-datasets, namely ‘Objects’, ‘Postcard’, ‘Wild’ from SIR^2 dataset

Once the data are downloaded, you must organize the dataset according to our code implementation (see the source code of datasets.CEILDataset, e.t.c.)

2. Playing with aligned data

Testing

• Download our pretrained model from OneDrive and move errnet_060_00463920.pt to checkpoints/errnet/.
• Evaluate the model performance by python test_errnet.py --name errnet -r --icnn_path checkpoints/errnet/errnet_060_00463920.pt --hyper

Training

• Reproduce our results by python train_errnet.py --name errnet --hyper
• Check options/errnet/train_options.py to see more training options.

3. Playing with unaligned data

• Reproduce our finetuned model by python train_errnet_unaligned.py --name errnet_unaligned_ft --hyper -r --icnn_path checkpoints/errnet/errnet_060_00463920.pt --unaligned_loss vgg

Citation

If you find our code helpful in your research or work please cite our paper.

 @inproceedings{wei2019single,
   title={Single Image Reflection Removal Exploiting Misaligned Training Data and Network Enhancements},
   author={Wei, Kaixuan and Yang, Jiaolong and Fu, Ying and David, Wipf and Huang, Hua},
   booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
   year={2019},
 }

Contact

If you find any problem, please feel free to contact me (kxwei at princeton.edu kaixuan_wei at bit.edu.cn). A brief self-introduction is required, if you would like to get an in-depth help from me.

Acknowledgments

• Our code architecture is inspired by CycleGAN and EDSR.

• Special thanks to @fqnchina and @ceciliavision for some discussions of this work.