Closed super233 closed 3 years ago
Hi,
Our method does not rely on adversarial learning.
It is only based on self-training, and a very simple variant where the pseudolabels are simply assigned the class of the highest probability, for every pixel, and from uniformly sampling images as usual.
Thanks for your reply, it's a really an awesome work! 😄
Does there exist wamup stage of adversarial training in your method? And is your method only based on self-training? I haven't found the corresponding content in paper so I'm going to ask you.