BMNet

CVPR 2022 oral (Official implementation of Bijective Mapping Network for Shadow Removal.

Yurui Zhu†, Jie Huang†, Xueyang Fu∗, Feng Zhao, Qibin Sun, Zheng-Jun Zha

†Equal Contributions *Corresponding Author

University of Science and Technology of China (USTC)

Introduction

This repository is the official implementation of the paper, "Bijective Mapping Network for Shadow Removal", where more implementation details are presented.

0. Hyper-Parameters setting

Overall, most parameters can be set in options/train/train_Enhance_ISTD.yml or options/train/train_Enhance_SRD.yml

1. Dataset Preparation

1. Download datasets and set the following structure

|-- ISTD_Dataset
    |-- train
        |-- train_A # shadow image
        |-- train_B # shadow mask
        |-- train_C # shadow-free GT
    |-- test
        |-- test_A # shadow image
        |-- test_B # shadow mask
        |-- test_C # shadow-free GT

2. Create xx_train.txt and xx_test.txt files and put them in BMNet/MainNet/data/.

python create_txt.py

2. Training

Firstly, we need to step into ColorTrans folder and train the subnetwork for color map restoration:

python train.py --opt options/train/train_Enhance.yml

Then, save the .pth file and put the file path to the ConditionNet in the Enhance_arch.py in MainNet folder.

Next, you can train the shadow removal network as:

python train.py --opt options/train/train_Enhance_ISTD.yml or train_Enhance_SRD.yml or train_Enhance_ISTD.yml

3. Inference

You should modify the path of pre-training weights and run:

python eval.py --opt options/test/test_Enhance_ISTD.yml or test_Enhance_SRD.yml or test_Enhance_AISTD.yml

Dataset

ISTD dataset/SRD dataset/AISTD dataset

Please refer to previous project of shadow removal (see https://github.com/jinyeying/DC-ShadowNet-Hard-and-Soft-Shadow-Removal)

Our results

Results on ISTD dataset (I have uploaded to https://drive.google.com/file/d/1cKRS26fgSOyIDqriD2fQFIcvyi2V8PIC/view?usp=sharing)

Results on SRD dataset (I have uploaded to https://drive.google.com/file/d/1Evi9-MWigJHuEwUov0w4v-gQqmZF1NPV/view?usp=sharing)

Results on AISTD dataset (I have uploaded to https://drive.google.com/file/d/1rg_hjihxIw4ypeQsiUavTWQ3dXD01qGu/view?usp=sharing)

Pre-trained Weights

The pre-trained weights (ISTD, AISTD and SRD) have been uploaded in BMNet/MainNet/pretrain/.

Contact

If you have any problem with the released code, please do not hesitate to contact me by email ([email protected] or [email protected]).

Cite

@InProceedings{Zhu_2022_CVPR,
    author    = {Zhu, Yurui and Huang, Jie and Fu, Xueyang and Zhao, Feng and Sun, Qibin and Zha, Zheng-Jun},
    title     = {Bijective Mapping Network for Shadow Removal},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {5627-5636}