WeaSeL (Weakly-supervised Segmentation Learning), is a method for few-shot semantic segmentation from sparse labels. This repository included the implementation of the WeaSeL method and the ProtoSeg method.
We also included our implementation of the Guided Networks.
The paper where this method is presented was published in IEEE Transactions on Multimedia
This is a Pytorch implementation.
Visualization of WeaSel. In the inner loop, the parameters are optimized using sparse labels from support sets. The optimal weights would be obtained if dense labels were presented in meta-training, similar to the ones in the query set that are used to compute the task outer loss. The model learns to intrinsically minimize this difference Delta between parameters, and thus fastly adapt to the few-shot task.
The following requirements are essential:
- python >= 3.5
- pytorch >= 1.5 && torchvision >= 0.6
- torchmeta >= 1.5.3
- pillow >= 7.0
- scikit-image >= 0.17
- numpy >= 1.18
- scikit-learn >= 0.23
To install, simply clone this repository with git clone https://github.com/pedrohtg/weasel.git.
Adjust the configurations in the config.ini file, and eventually the class ListDataset in the list_dataset.py file, if using different datasets.
Due to licensing we cannot provide direct acess to the medical datasets used, but mostly are publicly available upon request to owners.
The agricultural datasets, Brazilian Coffee and Orange Orchards, can be downloaded in this link.
With your datasets downloaded, and correctly setup (more info on data), set the default root path in the list_dataset.py.
Then check the configurations in config.ini file, lastly run the train script with:
python train_weasel.py
train_protoseg.py
train_guided.py
If you use this code, please cite:
@article{gama2022weakly,
title={Weakly Supervised Few-Shot Segmentation Via Meta-Learning},
author={Gama, Pedro Henrique Targino and Oliveira, Hugo Neves and Marcato, Jose and Dos Santos, Jefersson},
journal={IEEE Transactions on Multimedia},
year={2022},
publisher={IEEE}
}
If you find any issues or have any problem let us know. You can open a new issue, or contact us via email:
- Pedro Henrique Targino Gama (@pedrohtg) : phtg [at] dcc.ufmg.br
- Hugo Oliveira (@hugo-oliveira) oliveirahugo [at] vision.ime.usp.br