More data splits of VOC about pixelssl HOT 2 CLOSED

Thanks for your attention!

I have uploaded the PascalVOC data splits used in our paper to the master branch. You can try them now. Please create new scripts under the folder pixelssl/task/sseg/script and set new sublabeled_path in the scripts to retrain models (4 GPUs are required for each training).

The script pixelssl/task/sseg/dataset/PascalVOC/tool/random_sublabeled_samples.py is used to create all PascalVOC data splits used in our paper. Actually, you can randomly generate new data splits for experiments. Note that different data splits may cause a large performance gap, especially when the amount of labeled data is small. For example, the mIOU gap of SupOnly could be 1.5% under different 1/16 labeled data. Therefore, for each setup, the average over multiple runs (use different data splits) is reported in our paper.

If you have any problems in the experiment, please feel free to contact me.

BTW, It's my pleasure if you could contribute your implementation of the experiments on Cityscape to PixelSSL after you finish your work. The contributing document and the API document are coming soon. :)

from pixelssl.

Thank you!

from pixelssl.