This repository includes inplementation of the paper "Uncertainty-aware semi-supervised few shot segmentation" published in Pattern Recognition.
1. Download Pascal VOC dataset set the path for the directory in config.py file
2. Download pretrained model ResNet50 Resnet101 put them under FewShotSeg-dataset/cache/ folder.
- Train 5-shot model
/bin/bash script/train_logp.sh 0 5
python aggr_pth.py --model_dir PPNet_res101_Logp --fname 24000
- Fine-tune uncertainty estimation module
/bin/bash script/train_logp_additional.sh 0 5
python aggr_pth.py --model_dir PPNet_res101_Logp_additional
- Test 5-shot semi-supervised prediction using 6 unlabeled images
/bin/bash script/train_semi_sigma_sep.sh 0 5 6 0
- summarize test results
python aggr_json.py --model_dir PPNet_res101_Logp_additional_semi_w_sigma_qp_un6
If this code is helpful for your study, please cite:
@article{kim2023uncertainty,
title={Uncertainty-aware semi-supervised few shot segmentation},
author={Kim, Soopil and Chikontwe, Philip and An, Sion and Park, Sang Hyun},
journal={Pattern Recognition},
volume={137},
pages={109292},
year={2023},
publisher={Elsevier}
}