search

barnabytprowe / great3-public

Public repository for the Third Gravitational Lensing Accuracy Testing Challenge
BSD 3-Clause "New" or "Revised" License
9 stars 11 forks source link

Negative pixel values #12

Open adonnarum opened 10 years ago

adonnarum commented 10 years ago

Hi everyone, it looks like in a few cases the background subtraction led to regions with negative pixel values - see the example in the attached snapshot (field 28 in the real_galaxy-ground-constant branch). bk_ex

barnabytprowe commented 10 years ago

Aha good spot! Yes, that does seem to have happened here. I would think this must be a background subtraction defect that is present in the HST COSMOS postage stamp images used to make the real galaxy branch simulations. In fact it looks a lot like a near neighbour masking-related oversubtraction to me, but that's just a hunch.

No background subtraction can ever be perfect, and the COSMOS treatment used to make these training set images (which I think used a SExtractor-estimated background map) is no different. We did do eyeball checks for these sorts of things but missed this! How frequently do you see this occurring?

rmandelb commented 10 years ago

I think your interpretation of what’s going on in this image is correct. We worked hard to reduce the incidence of weird background subtraction and masking defects to <~0.5%, but we know that our efforts in that regard were not perfect. Can you (roughly) estimate the incidence of this kind of defect?

If I could bother you for a bit more info, can you send me either the (X, Y) position or the galaxy ID from the catalog? Given the branch and subfield which you already posted and the information about the galaxy location in the field, I can figure out exactly which COSMOS galaxy this is, and try to understand if there are better sets of cuts or better way to handle the image processing that could eliminate these problems. Obviously, this does not help GREAT3, so this is more for “life beyond GREAT3”. :)

adonnarum commented 10 years ago

Hello,

How frequently do you see this occurring?

so far we noticed just some (< 10) cases in a few fields; if you want, we can let you know if we find other examples.

If I could bother you for a bit more info, can you send me either the (X, Y) position or the galaxy ID from the catalog?

Sure, no bother at all: the nominal image coordinates for that galaxy are 1224,1608 (one-indexed). Thanks for your replies!

rmandelb commented 10 years ago

If you see <10 cases in a single field, that's <10/10^4 or <0.1%. So, I would love it if they weren't there at all :) but at least this is not a higher-incidence problem!

I'll check into the galaxy for which you gave me coordinates, will post if I find something interesting...

rmandelb commented 10 years ago

I can confirm, looking at the original COSMOS image for this object, that it is one of the nastiest cases we have to deal with:

(1) It is very near two objects that are bright enough to make sky determination difficult. (2) The light profile actually overlaps slightly with those objects, so deblending is a problem.

We tried to make some masking cuts that we could use to automatically weed out such problem galaxies, but for whatever reason those cuts didn't work here.