Question about paper

[lovelyzzkei]

Hi, sir.
Thanks for great work and opened source for sharing.

I'm also looking into monocular depth estimation and have a question about tile-based depth estimation approach.

In section 3.1.(iii), there's about inconsistency issue of tile-based approach.

(iii) Fusion and Guided Fusion Network: To resolve the inconsistency issues with patch-wise prediction, our goal in this step is to transfer the global information from the coarse depth obtained in (i) to the patch-wise predictions without compromising their higher details. While the transfer could be implemented by simply learning a pix2pix U-Net [27, 34], our key idea is to exploit the multi-scale features ...

I know you proposed a G2L module for this and it's the novel point of this paper, but I was also interested in the pix2pix approach. How can I resolve this issue with pix2pix U-Net? May I know approach to this?

Thanks.