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
LTR is a general framework for training SODGAN network.
The toolkit contains the implementation of the following methods.
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.
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.
The models trained using PyTorch. Your can download trained well models model zoo.
git clone https://github.com/yongwuSHU/Advsal.git
- 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
Activate the right environment and run it.
python training.py sodgan
Activate the right environment and run it
python testing.py sodgan --dataset pascal(ecssd,hkuis,dutste,dutomron,....)
Any question, please email: yong_wu1@163.com, yong_wu@shu.edu.cn