We propose a new method for effective shadow removal by regarding it as an exposure fusion problem. Please refer to the paper for details: https://openaccess.thecvf.com/content/CVPR2021/papers/Fu_Auto-Exposure_Fusion_for_Single-Image_Shadow_Removal_CVPR_2021_paper.pdf.
- For data folder path (ISTD), train_A: shadow images, train_B: shadow masks, train_C: shadow free images, organize them as following:
--ISTD+
--train
--train_A
--1-1.png
--train_B
--1-1.png
--train_C_fixed_official
--1-1.png
--train_params_fixed # generate later
--1-1.png.txt
--test
--test_A
--1-1.png
--test_B
--1-1.png
--test_C
--1-1.png
--mask_threshold # generate later
--1-1.png
- Run the code
./data_processing/compute_params.ipynb
for exposure parameters generation. The result will be put in./ISTD/train/train_params_fixed
. Here, namestrain_C_fixed_official
andtrain_params_fixed
are for ISTD+ dataset, which are consitent withself.dir_C
andself.dir_param
in./data/expo_param_dataset.py
. - For testing masks, please run the code
./data_processing/test_mask_generation.py
. The result will be put in./ISTD/mask_threshold
.
We release our pretrained model (ISTD+, SRD) at models
pretrained model (ISTD) at models
Modify the parameter model
in file OE_eval.sh
to Refine
and set ks=3, n=5, rks=3
to load the model.
Modify the corresponding path in file OE_train.sh
and run the following script
sh OE_train.sh
- For the parameters:
DATA_PATH=./Datasets/ISTD or your datapath
n=5, ks=3 for FusionNet,
n=5, ks=3, rks=3 for RefineNet.
model=Fusion for FusionNet training,
model=Refine for RefineNet training.
The trained models are saved in ${REPO_PATH}/log/${Name}
, Name
are customized for parameters setting.
In order to test the performance of a trained model, you need to make sure that the hyper parameters in file OE_eval.sh
match the ones in OE_train.sh
and run the following script:
sh OE_eval.sh
- The pretrained models are located in
${REPO_PATH}/log/${Name}
.
The results reported in the paper are calculated by the matlab
script used in other SOTA, please see evaluation for details. Our evaluation code will print the metrics calculated by python
code and save the shadow removed result images which will be used by the matlab
script.
- Comparsion with SOTA, see paper for details.
- Penumbra comparsion between ours and SP+M Net
- Testing result
The testing results on dataset ISTD+, ISTD, SRD are:results
More details are coming soon
@inproceedings{fu2021auto,
title={Auto-exposure Fusion for Single-image Shadow Removal},
author={Lan Fu and Changqing Zhou and Qing Guo and Felix Juefei-Xu and Hongkai Yu and Wei Feng and Yang Liu and Song Wang},
year={2021},
booktitle={accepted to CVPR}
}