This is the implementation of the method described in paper: toward practical weakly supervised semantic segmentation via point-level supervision.

• Python3.7+
• Pytorch1.0+
• Numpy, OpenCV
• Pydensecrf

1. Prepare the dataset, e.g., VOC, Cityscapes, and ADE20k. We only use the image and point-level labels as supervision, where the used point labels can be downloaded here.

2. Run the scripts to reproduce the results. Change the dataset pathes if necessary, which are typically determined by train-image-root, test-image-root, train-label-file, and test-gt-root. The results with different datasets can be reproduced by simply running scripts train_xxx.py, e.g.,

python train_voc_all.py --gpus 0,1,2,3

The scripts will automatically conduct evaluation and print logs during running.