ryo-abiko / canet

Repo of "Channel Attention GAN Trained with Enhanced Images for Single Image Shadow Removal"

Channel Attention GAN Trained with Enhanced Dataset for Single-Image Shadow Removal (CANet)

Our Paper is published in IEEE Access. You can check the paper here.

Datasets

• ISTD : ST-CGAN
• SRD : Deshadow-Net
• ISTD+, SRD+
  Datasets are not available but you can create the dataset by applying the MATLAB code provided by the authors.

Our results

• ISTD,ISTD+,SRD,SRD+ [Google Drive]
• Trained weights [Google Drive]

Create environment with Anaconda

conda create -n CANet python=3.8 -y
conda activate CANet
conda init
conda install -c pytorch torchvision=0.10.0 -y
conda install -c conda-forge scikit-image -y

Train

1. Prepare the training dataset. You need to prepare the edge mask by yourself. Run makemask.m with MATLAB on the mask images to obtain the edge mask dataset.
2. Execute the following command.

(This is an example.)

python CANet_train.py \
    --train_name train1 \
    --shadow_dir_path datasets/ISTD/train/train_A \
    --gt_dir_path datasets/ISTD/train/train_C \
    --mask_dir_path datasets/ISTD/train/train_B \
    --mask_edge_dir_path datasets/ISTD/train/train_B_edge \
    --pretrained_d detect_ISTD.pth(optional) \
    --pretrained_r remove_ISTD.pth(optional)

Test

1. Prepare the trained weights.
   Our pre-trained weights can be downloaded from Google Drive.
2. Prepare the test images and put them in a single directory.
3. Execute the following command.

python CANet_test.py \
    --detection_pth_path <path of .pth> \
    --removal_pth_path <path of .pth>  \
    --input_img_path <path of directory>

4. Directory output_CANet is automatically created and the output images are saved here.

Evaluation

1. Put your results under results/<dataset name>/<method name>.
2. Make sure that the test images are located under datasets/<dataset name>.
3. Run the following command.

python evaldata.py \
    --method_name <method name> \
    --dataset_name ISTD \
    --resized 1

Directory structure

CANet
├── evaldata.py
├── getDatasetPath.py
├── README.md
├── datasets/
│   ├── ISTD/
│       ├── test/
│           ├── test_A/
│               ├── 101.png
│           ├── test_B/
│           ├── test_C/
│       ├── train/
├── results/
│   ├── ISTD/
│       ├── CANet/
│           ├── 101.png

Results

Comparison with the other methods. Please see our paper for the details.

Citation

@ARTICLE{ryo2022,
  author={Ryo Abiko and Masaaki Ikehara},
  journal={IEEE Access}, 
  title={Channel Attention GAN Trained With Enhanced Dataset for Single-Image Shadow Removal}, 
  year={2022},
  volume={10},
  number={},
  pages={12322-12333},
  doi={10.1109/ACCESS.2022.3147063}}

canet's Issues

doubts about the test restoration results in the paper

Hi! thanks for you paper, the paper was cool.

but I have some doubts about the test restoration result.
like in table3 and table4, The AEF[17] results in ISTD+ and ISTD shadow part were 8.77 and 10.05, respectively. But the original paper utilize the RMSE evaluation metric and color space same in LAB were 6.5 and 7.77, there is a relatively obvious difference between the two result,So it is used the evaluation metric in SID[14] that causes the relatively obvious difference in the results?

closed