In preparation for writing our first paper on the Tractor, Lang and I discussed the relationship between the Tractor and the things that have come before, like SDSS Photo (Robert Lupton's baby) and SExtractor (Emmanuel Bertin's). We need to get right the ways in which Photo is and isn't an approximation to a likelihood maximization. For example, in fitting galaxy shapes, it is very close to doing maximum likelihood. However, it doesn't set the centroids of the objects or do aperture photometry that way. Of course, the Tractor can do sampling—that is, it can go way beyond optimization—but we don't want to be unfair to the precedents. We also want to slip in lots of the good ideas we gathered from Photo, including profile truncation and sampling considerations.
In the morning, Fergus and I spent some quality time with Andrew Flockhart (NYU), who is doing cosmic-ray classification in single-epoch imaging using nearest-neighbor (and maybe soon more insane methods) given a good training set derived from HST data. We decided to fully understand the system by making lots of visualizations of all the things it is doing; this is all related to my ideas about functional testing: You don't find bugs by looking at your code; you find them by producing results and comparing with expectations.
No comments:
Post a Comment