About Eq.5

[swift1988]

To obtain the Eq.5, you supposed the ls = lt + n, and n is a constant tensor. I am confused with this presumption. Actually, I think this presumption can not be held for each pixel, and Eq.5 is only a very special case.

I wonder if you can help me to better understand your analysis.

[TinyTigerPan]

Sorry for the late reply, because I was busy with other work.
Yes, you are right, this is just an extreme case to illustrate our motivation. It is not common to encounter this situation in a real-world example. We just want to explain the inconsistency between softmax and sigmoid cross-task protocol.
Thanks for your attention to our work!