Two question about upsample process

[Ikhwansong]

Hi. nice to meet you !

Your logic for solving different scale problem was very reasonable !. It was interesting to me!

I have a one question about the process of upsample in the paper.
That is why you didn't use transpose convolutional method for upsampling.

Usually, for better performance, Transpose convolutional be used for upsampling (i.g, Grid-Net...)
or within similar architecture for image reconstruction.

But just only bilinear-upsampling be used.

Was it better?

---

Additional question is this.

According to Figure 4. the last size of output upper down scale convolutional path has 1 by 1.
but it meet a Conv[256, 3x3, 1x1, 256] directly.
How can the output which has 1 by 1 size of [w,h], feed into a Conv[256, 3x3, 1x1, 256] ??
Is it possible?

Please, reply my question. thanks you!