Question about the input data

[liuliuliu1111]

Hello writers,

Thank you so much for your reply to my previous question! Now I have some new questions.

I know that you use two-stage network to combine both spatial and temporal information to get better results. I have learned your code, paper and discussion #152, it seems like that the final loss fuction only focus on only one frame. Although you put three frames into three frame analysis U-Nets, how do you make sure that they work on the way you want? In another way, my question is that, how do you use the emitter id to make temporal information useful? I know it is used somewhere, but I didn't find it in the code.

Thanks!

[Haydnspass]

@liuliuliu1111 thanks for the question.
The network's target is only emitters on the "middle" frame. The network learns to automatically and implicitly utilise the information from the neighbouring frames to better determine the emitters on the middle frame. This will reduce the loss compared to when it had only the middle frame; that's why it learns to do so.
Hope that helps :)

[liuliuliu1111]

So do you put any secial weight on the same id bubble or the area around them? Or you just let the network dig out the relationship between them by it self?

[liuliuliu1111]

Not secial, but special, I'm sorry I type it wrong.

[Haydnspass]

The network does it implicitly; it's only incentivised by the loss to fit the target emitters as well as possible.

[liuliuliu1111]

OK, I get it. Thanks a lot!