jsws

Environment

clone repo:

git clone https://github.com/zengxianyu/jsws.git
git submodule init 
git submodule update

prepare environment:

conda env create --file=environments.yaml

Prepare Data

DUTS-train (Onedrive)

ECSSD (Onedrive)

VOC2012 (Onedrive)

SegmentationClassAug (Onedrive)

Train stage 1

train using image-level class labels and saliency ground-truth:

weak_seg_full_sal_train.py

Open the file output/logs/train1.html in browser for visualization

It should be easy to achieve MIOU>54 but you may need to try multiple times to get the score MIOU 57.1 or more than that in Table. 5 of the paper.

Train stage 2

train a more complex model using the prediction of the model trained in the stage 1.

syn_stage1.py
self_seg_full_sal_train.py

Saliency results

download saliency maps (Google Drive); One Drive

Citation

@inproceedings{zeng2019joint,
  title={Joint learning of saliency detection and weakly supervised semantic segmentation},
  author={Zeng, Yu and Zhuge, Yunzhi and Lu, Huchuan and Zhang, Lihe},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  year={2019}
}