This is a project of csc-2547. Based on the pytorch implementation of "Single Image Reflection Removal Exploiting Misaligned Training Data and Network Enhancements".

• Python >=3.5, PyTorch >= 0.4.1
• Requirements: opencv-python, tensorboardX, visdom

you can use environment.yaml to create your environment

• 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

• 100 synthetic testing images from CEILNet dataset (testdata_reflection_synthetic_table2)
• 20 real testing images from Berkeley real dataset.

• Train the model by python train_errnet.py --name errnet --hyper --pixel_loss [mse+grad|ms_ssim_l1+grad|ms_ssim_l1|highpass] --model errnet_alw_model , by changing the value of pixel_loss, you can train the model with different kind of loss functions.
• Check options/errnet/train_options.py to see more training options.

Evaluate the model performance by python test_errnet.py --name errnet -r --icnn_path [model-path] --hyper --testcase [1|2], set testcase to 1 for synthetic data or 2 for read data