Huanglin Yu, Ke Chen*, Kaiqi Wang, Yanlin Qian, Zhaoxiang Zhang, Kui Jia     AAAI 2020

This implementation uses Pytorch.

Please install Anaconda firstly.

git clone https://github.com/yhlscut/C4.git
cd C4
## Create python env with relevant packages
conda create --name C4 python=3.6
source activate C4
pip install -U pip
pip install -r requirements.txt
conda install pytorch torchvision cudatoolkit=9.0 -c pytorch  # cudatoolkit=10.0 for cuda10

Tested on pytorch >= 1.0 and python3.

Shi's Re-processing of Gehler's Raw Dataset:

• Download the 4 zip files from the website and unzip them
• Extract images in the /cs/chroma/data/canon_dataset/586_dataset/png directory into ./data/images/, without creating subfolders.
• Masking MCC chats:

  bash ./data/run.sh

• Pretrained models can be downloaded here. To reproduce the results reported in the paper, the pretrained models(*.pth) should be placed in ./trained_models/, and then test model directly

Open the visdom service

python -m visdom.server -p 8008

• Please train the three-fold models (modify foldnum=0 to be foldnum=1 or foldnum=2 in line 6 of ./scripts/train_sq_1stage.sh and ./scripts/train_sq_3stage.sh accordingly)
• Train the C4_sq_1stage first:

bash ./scripts/train_sq_1stage.sh

• Train the C4_sq_3stage (Before that, please move the directory ./log/C4_sq_1stage to ./trained_model):

bash ./scripts/train_sq_3stage.sh

• After training, move the trained models directory in ./log/C4_sq_3stage to ./trained_model/, and run:

bash ./scripts/test_sq_3stage.sh

• To reproduce the results reported in the paper, move the pretrained models(*.pth) downloaded from here to ./trained_models/, and then test model directly.