question about paper about torchsemiseg

[charlesCXK]

Hi, we use all the pixels as the target for the other network during training, without any threshold selection. This makes our approach simple and clean.

In the early stage of training, there must exist some pixels with wrong predictions in both networks. Let's consider these situations (two columns in each row means whether the prediction of each network is correct):

(1) Correct, Correct
(2) Correct, Wrong
(3) Wrong, Correct
(4) Wrong, Wrong

In (1)(3), the target is correct, so the prediction is supervised with the correct signal.
In (4), both the prediction and the target are wrong, but the prediction will be punished. So it is hard to say whether it is harmful or not.
In (2), the supervision signal is wrong, which may be harmful.

I think our CPS works well because the advantages outweigh the disadvantages that exist in these four situations. One may design an algorithm to suppress the wrong supervision signal in the early training, which may further improve CPS.

[czy341181]

Thanks for your patient reply! The method is very simple and effective.

By the way, If it is simply set the threshold of segmentation score of networks output, above a certain threshold as a pseudo label, like GCT. Does it work? Have you done any experiments like this?

[charlesCXK]

We haven't conducted experiments like this. Since the code and data are released, one could try different strategies themselves~