This repository contains the official PyTorch implementation of the paper "Unlocking the Potential of Ordinary Classifier: Class-specific Adversarial Erasing Framework for Weakly Supervised Semantic Segmentation" paper (ICCV 2021) by Hyeokjun Kweon and Sung-Hoon Yoon.

We have developed a framework that extract the potential of the ordinary classifier with class-specific adversarial erasing framework for weakly supervised semantic segmentation. With image-level supervision only, we achieved new state-of-the-arts both on PASCAL VOC 2012 and MS-COCO.

If our code be useful for you, please consider citing our ICCV paper using the following BibTeX entry.

@inproceedings{kweon2021unlocking,
  title={Unlocking the potential of ordinary classifier: Class-specific adversarial erasing framework for weakly supervised semantic segmentation},
  author={Kweon, Hyeokjun and Yoon, Sung-Hoon and Kim, Hyeonseong and Park, Daehee and Yoon, Kuk-Jin},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={6994--7003},
  year={2021}
}

• Tested on Ubuntu 16.04, with Python 3.6, PyTorch 1.5.1, CUDA 10.1, both on both single and multi gpu.
• You can create conda environment with the provided yaml file.

conda env create -f od_cse.yaml

• The PASCAL VOC 2012 development kit: You need to specify place VOC2012 under ./data folder.
• ImageNet-pretrained weights for resnet38d are from [resnet_38d.params]. You need to place the weights as ./pretrained/resnet_38d.params.
• PASCAL-pretrained weights for resnet38d are from [od_cam.pth]. You need to place the weights as ./pretrained/od_cam.pth.

• Please specify the name of your experiment.
• Training results are saved at ./experiment/[exp_name]

python train.py --name [exp_name] --model model_cse

python infer.py --name [exp_name] --model model_cse --load_epo [epoch_to_load] --vis --dict --crf --alphas 6 10 24

python evaluation.py --name [exp_name] --task cam --dict_dir dict

python evaluation.py --name [exp_name] --task crf --dict_dir crf/06

we heavily borrow the work from AffinityNet repository. Thanks for the excellent codes!