LG-ShadowNet

Shadow Removal by a Lightness-Guided Network with Training on Unpaired Data.

@article{liu2021shadow,
  title={Shadow Removal by a Lightness-Guided Network with Training on Unpaired Data},
  author={Liu, Zhihao and Yin, Hui and Mi, Yang and Pu, Mengyang and Wang, Song},
  journal={IEEE Transactions on Image Processing},
  volume={30},
  pages={1853--1865},
  year={2021},
  publisher={IEEE}
}

Dependencies

This code uses the following libraries

python 3.7+
pytorch 1.1+ & tochvision
scikit-image

Train and test on the adjusted ISTD dataset

Train

Set the path of the dataset in train_aistd_module1.py
Run train_aistd_module1.py
Set the paths of the saved module1 models (netG_A2B.pth,netG_B2A.pth) and the dataset in train_aistd.py
Run train_aistd.py

Test

Set the paths of the dataset and saved LG-ShadowNet models (netG_A2B.pth) in test_aistd.py
Run test_aistd.py

Evaluate

Set the paths of the shadow removal result and the dataset in evaluate.m
Run evaluate.m

Acknowledgments

Code is implemented based on Mask-ShadowGAN.

Results of LG-ShadowNet

GoogleDrive: AISTD/ISTD/USR

BaiduNetdisk: AISTD/ISTD/USR (Access code: 1111)

All codes will be released to public soon.

AISTD Results (size: 480x640)

Method	Shadow	Non-shadow	All
Mask-ShadowGAN(our run)	11.5*	4.5*	5.5
LG-ShadowNet	10.6*	4.0*	5.0

AISTD Results (size: 256x256)

Method	Shadow	Non-shadow	All
Mask-ShadowGAN(our run)	10.8*	3.8*	4.8
LG-ShadowNet	9.9*	3.4*	4.4

*Since the ~~RMSE~~ (MAE) in shadow and non-shadow regions are computed on each image first and then compute the average of all images, the results may be different from yours.

hhqweasd/LG-ShadowNet