Auto-exposure fusion for single-image shadow removal
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.
Dataset
- 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.ipynbfor exposure parameters generation. The result will be put in./ISTD/train/train_params_fixed. Here, namestrain_C_fixed_officialandtrain_params_fixedare for ISTD+ dataset, which are consitent withself.dir_Candself.dir_paramin./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.
Pretrained models
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.
Train
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.
Test
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}.
Evaluation
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.
Results
- 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
Bibtex
@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}
}
![@fl82hope Hi, thanks for your prompt reply. To be more clear, did you mean that the variables addgt and mulgt are the same as input['param'] which will not be used during testing mode?](https://avatars.githubusercontent.com/u/60936436?v=4)
67 Dec 28, 2022
13 Nov 21, 2022
95 Dec 18, 2022
68 Jan 3, 2023
39 Nov 23, 2022
2 Jan 5, 2022
117 Dec 20, 2022
24 Oct 4, 2022
105 Dec 30, 2022
2 Aug 19, 2022
5 Jan 5, 2023
163 Dec 22, 2022
64 Nov 2, 2022
464 Dec 27, 2022
2 Sep 15, 2022
13 Apr 26, 2021
27 Oct 26, 2022
89 Jan 4, 2023
162 Jan 5, 2023