Class project of computational photography

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

Prerequisites

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

Folder list

• ERRNet-ref folder contains the debugged original code of ERRNet
• ERRNet-mydata folder contains the code for training using our own synthetic data (Code for data generating is in sync_model folder）
• ERRNet-TR folder contains our newly proposed network: ERRNet-TR and the relating training and test code
• ERRNet-TR-lite folder contains our newly proposed network: ERRNet-TR-lite and the relating training and test code
• sync_model folder contains code for data generating and a data sample

Quick Start

Testing

• Running the following command after entering the corresponding folder python test_errnet.py --name errnet -r --icnn_path checkpoints/errnet/ourmodel.pt --hyper --gpu_ids -1 (gpu_ids=-1 means using CPU to test)

Training

• Running the following command after entering the corresponding folder python train_errnet.py --name errnet --hyper （Need to prepare the dataset in the folder reflection_data before training）
• Viewing options/errnet/train_options.py for more training options