I didn't get much done today after a nearly-all-nighter helping with the Fergus paper on high-contrast imaging and various papers by Bovy and Bolton. However, I did pair-code a bit of The Thresher with Foreman-Mackey; we showed that we can take a small number of PanSTARRS cutouts with different seeing and make a coadded
version that is higher in signal-to-noise and resolution than any of the original images from which it is built. (Phil Marshall made the image below, and also noted that we are "over-deconvolving" which is a problem I have to agree.) That could be useful to the survey team. Late in the day, Maria Kapala and I beat our heads against the HST data reduction pipelines used in the HST Archive and the PHAT HLSP Archive. There seem to be inconsistencies in the image pixel histograms that are very difficult to explain by any reasonable model of what these pipelines do.
The background in the "Threshed" exposure looks worrying. Have you characterized the amount of correlated noise in your output images?
ReplyDelete@doug: 1. stretch is harder on that image, and 2. of course you get correlated noise! You can't have high resolution *and* uncorrelated noise. If you want both, don't combine your images!
ReplyDeleteSure you introduce some correlated noise, but it's still legitimate to try to minimize it, or at least cause it to have tractable statistical properties. Most tools for photometry or morphology can't handle correlated noise. [And now some needless provocation:] If you can't run those tools, then what is the purpose for the thresher?
ReplyDelete@doug: Great question. Two answers (a) if you are photometering stacked images, you are making a mistake, and (b) when we are combining more than four images (yes, that is a combination of only four images) the correlations decay as something like 1/sqrt(N). More soon on this!
ReplyDelete