Closed jmbalasalle closed 1 year ago
Hello,
Thank you for your kind message and for bringing up this point. My intuition is the same as yours.
For new AOIs with no GT data I would go with some public elevation model to get reasonable values for min_alt
, max_alt
. E.g. task SRTM (check srtm4 !) at the central [lat, lon] point and subtract/add a reasonable margin that gives you an altitude range more or less suited to your target area.
Concerning the second question. Depending on how big the range is after defining min_alt
and max_alt
, yes, it may be wise to increase the number of sampling points (--n_samples
). More samples will take more time though. In case it's of any use, in our AOIs we never exceeded 100m between min_alt
and max_alt
.
Good luck with your experiments !
Hi, this is fantastic work, great job.
I have a detailed question. It appears that you are using ground truth data to do the actual training. Specifically the values of
min_alt
andmax_alt
. These are read from the json file here. And these .json file are created by thecreate_satellite_dataset.py
code here.It therefore appears that when you are creating all the rays, you are using the
min_alt
andmax_alt
values (for the near and far planes). If we were to run this on a new or different AOI, that data would not be available. My question is this: how sensitive to these values is the code overall? Could we approximate these values using something like GTOPO30?In the case of approximating
min_alt
andmax_alt
would we need to increase the number of sampling points on the rays?I'm trying to run this on different areas, so I will probably try the approximation technique and see how it goes.
Thanks again for such a promising project!