The code and trained models of:

Weakly-Supervised Semantic Segmentation via Sub-category Exploration, Yu-Ting Chang, Qiaosong Wang, Wei-Chih Hung, Robinson Piramuthu, Yi-Hsuan Tsai and Ming-Hsuan Yang, CVPR 2020 [Paper]

We develope a framework to generate semantic segmentation labels of images given their image-level class labels only. We propose a simple yet effective approach to improve the class activation maps by introducing a self-supervised task to discover sub-categories in an unsupervised manner. Our algorithm produces better activation maps, thereby improving the final semantic segmentation performance.

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

@InProceedings{Chang_2020_CVPR,
author = {Chang, Yu-Ting and Wang, Qiaosong and Hung, Wei-Chih and Piramuthu, Robinson and Tsai, Yi-Hsuan and Yang, Ming-Hsuan},
title = {Weakly-Supervised Semantic Segmentation via Sub-category Exploration},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2020}
}

• Tested on Ubuntu 14.04, with Python 3.5, PyTorch 1.0.1, Torchvision 0.2.2, CUDA 9.0, and 1x NVIDIA TITAN X GPU.
• The PASCAL VOC 2012 development kit: You need to specify the path ('voc12_root') of your downloaded dev kit.
• To extract the feature for the first round training, you need to download the pretrained weight of the AffinityNet [res38_cls.pth]
• To train the classification model, you need to download the pretrained weight of the Mxnet and ResNet-38 model [ilsvrc-cls_rna-a1_cls1000_ep-0001.params]
• Please create a folder named weights and save pretrained models in the folder.

• We iteratively train the model with three rounds to achieve the best performance. You can directly run script.sh to complete the three-round training.
• [Note] To run the whole pipeline, you need to specify the path to the saved model for each round. Please see the command in script.sh.

bash script.sh

• The iteratively training process includes the step of feature extraction, sub-category label generation, and classification model training. You can use following scripts to run each step.
• [Note] Please specify the argument in the command. You can also check script.sh to see more details.

bash python extract_feature.py --weights [your_weights_file] --infer_list ./voc12/train_aug.txt --voc12_root [your_voc12_root_folder] --save_folder [your_save_folder] --from_round_nb [round_number]

bash python create_pseudo_label.py --save_folder [your_save_folder] --k_cluster 10 --for_round_nb [round_number]

bash python train_cls.py --lr 0.01 --voc12_root [your_voc12_root_folder] --weights [your_weights_file] --round_nb [round_number] --save_folder [your_save_folder]

• With the classification model, you can infer the classifier and generate the response map

bash python infer_cls.py --infer_list voc12/train.txt --voc12_root [your_voc12_root_folder] --weights [your_weights_file] --save_path [your_save_cam_folder] --save_out_cam 1 --round_nb [round_number]

• Refinement: We adopt the random walk method via affinity to refine the map as pixel-wise pseudo ground truths for semantic segmentation. Please refer to the repo of AffinityNet [1].
• Segmentation network: We utilize the Deeplab-v2 framework [2] with the ResNet-101 architecture [3] as the backbone model to train the segmentation network. Please refer to the repo.

1. Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. CVPR, 2018.
2. Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2017.
3. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016.