Awesome

Robust Shadow Detection by Exploring Effective Shadow Contexts

This repository contains Pytorch code for the paper titled Robust Shadow Detection by Exploring Effective Shadow Contexts at ACM Multimedia 2021.

In this article, we propose a novel end-to-end shadow detection method based on the encoder-decoder structure. Our key contribution is the integration of the regular deep feature with discrimative multi-scale features to fulfil robust shadow augmentation. This integration is obtained by a new module called Effective-Context Augmentation (ECA).

Architecture

Attached below is the architecture diagram as given in the paper. Built on the encoder-decoder structure, it takes ECA as the main building block in the encoder while also adopts the features from those ECAs to decode robust shadow classification results.

Requirements

• Pytorch
• Python3.X
• numpy
• cv2
• PIL

Usage

• You can search and download the datasets from the Internet.
• ResNext101 has been adopted，and you can download the ResNet101's settings from here，you can put it in the ./ directory.

Training

python train.py

Testing

python test.py

Results

(Left to right: Input, ground truth, detection result)

More results can be downloaded here.

Some pretrained models can be downloaded here for quick program setups.

Citation

Please cite the following paper if you think this project is useful for your work. Thanks.

@inproceedings{

FangHWS2021Shadow,

author = { Fang, Xianyong and He, Xiaohao and Wang, Linbo and Shen, Jianbing},

booktitle = {Proceedings of the 29th ACM International Conference on Multimedia - ACM Multimedia 2021},

title = {{Robust Shadow Detection by Exploring Effective Shadow Contexts}},

address = {Chengdu, China},

year = {2021}

}