Advsal

A general python framework for training and testing SODGAN, based on PyTorch.

This release also includes many new features, including:

Multi-GPU training（You need to retrain the model for first training step）
PyTorch v1.3 support

Highlights

SODGAN network

Training Framework: LTR

LTR is a general framework for training SODGAN network.

trainers

The toolkit contains the implementation of the following methods.

Publication and citation

SODGAN is proposed in our paper accepted by JVCIR 2020. Detailed explanation of our method can be found in the paper:

@article{wuadvsal2020,
  author={Yong Wu and Zhi Liu and Xiaofei Zhou}, 
  title = {Saliency detection using adversarial learning networks},
  journal = J. Vis. Commun. Image Represent,
  volume = {67},
  pages={102761},
  year = {2020},
  month={Feb.},
}

The paper can be downloaded here.

SODGAN

Official implementation of the SODGAN network. SODGAN is two-stage training architecture, which can accelerate training speed. And it takes only 35 minutes to train on a Titan Xp GPU. Our model can detect salient objects better.

Model Zoo

The models trained using PyTorch. Your can download trained well models model zoo.

Installation

Clone the GIT repository.

git clone https://github.com/yongwuSHU/Advsal.git

Install dependencies

PyTorch >=0.4.1 (we have tested PyTorch v1.3 with Python 3.7)
Python 3
Ubuntu 16.04 (we don't recommend OS 18.04)
You need to install pydensecrf
You can download the resnet101 model

Let's start !

Training

Activate the right environment and run it.

python training.py sodgan

Testing

Activate the right environment and run it

python testing.py sodgan --dataset pascal(ecssd,hkuis,dutste,dutomron,....)

Contributors

Yong Wu

Any question, please email: yong_wu1@163.com, yong_wu@shu.edu.cn

yongwuML / Advsal