search

jiupinjia / SkyAR

Official Pytorch implementation of the preprint paper "Castle in the Sky: Dynamic Sky Replacement and Harmonization in Videos", in arXiv:2010.11800.
https://jiupinjia.github.io/skyar/
2.01k stars 247 forks source link

Question regarding CoordConv #20

Closed yggs1401 closed 3 years ago

yggs1401 commented 3 years ago

Hello,

First of all, nice job. I was wondering why did you only encode the y coordinate for the CoordConv instead of both x, y or only x. Can you explain the reason behind this?

Thank you!

jiupinjia commented 3 years ago

Hi @yggs1401, thanks for your question. The reason why I was only using y instead of using (x, y) is that sky pixels usually appear in the upper region of an image. There is no clear clue as to which part of the sky tends to appear in the x-direction.

yggs1401 commented 3 years ago

Thank you for your answer! From upper region I need to understand the y-axis related to the height of the image and thus we locate the sky region?

jiupinjia commented 3 years ago

Yes, exactly. The pixels at the top of the image are more likely to be sky pixels than the pixels at the bottom.

yggs1401 commented 3 years ago

I understand now, thank you! I will close the issue now