The implementation of "Delving Deeper Into Pixel Prior for Box-Supervised Semantic Segmentation", IEEE TIP.

• python 3.7 / pytorch 1.2.0
• pydensecrf
• opencv

• Pascal VOC 2012 Dataset
  • extract 'VOCtrainval_11-May-2012.tar' to 'VOCdevkit/'
• Pascal VOC 2012 Augment Dataset
  • extract labels to 'VOCdevkit/VOC2012/SegmentationClassAug'
• WSSL pseudo labels
  • extract pseudo labels to 'VOCdevkit/VOC2012/SegmentationClassBboxCRF'
• SDI pseudo labels
  • extract pseudo labels to 'VOCdevkit/VOC2012/VOC12_M&G+'

Finally, it should like this

VOCdevkit
└── VOC2012
    ├── ImageSets
    ├── JPEGImages
    ├── SegmentationClass
    ├── SegmentationClassAug
    ├── SegmentationClassBboxCRF
    └── VOC12_M&G+

python main.py --type=train

python main.py --type=test --use_crf

python evalate.py

1. trained model.
2. For DeepLabv2-ResNet-101 backbone, we refer to deeplab-pytorch. We also provide the ImageNet pretrained model.

1. DeepLab-V1-PyTorch

@ARTICLE{9684236,
  author={Ma, Tianqi and Wang, Qilong and Zhang, Hongzhi and Zuo, Wangmeng},
  journal={IEEE Transactions on Image Processing}, 
  title={Delving Deeper Into Pixel Prior for Box-Supervised Semantic Segmentation}, 
  year={2022},
  volume={31},
  number={},
  pages={1406-1417},
  doi={10.1109/TIP.2022.3141878}}