davidwhogg/ballpeenhammer

calibrating zeros, darks, and flats by science data alone

Ball-Peen Hammer

A few love taps with the Ball Peen Hammer and everything's flat!

A project to self-calibrate all astronomical imaging: Inferring zeros, darks, flats, and point-spread functions by use of the science data alone.

authors

• Ross Fadely (NYU)
• David W. Hogg (NYU)

license

Copyright 2013, 2014 the authors. All rights reserved. This project is visible on the web, but not yet licensed for re-use or redistribution.

ballpeenhammer's Issues

switch to taking not top 1 percent variance patches, but rather top N patches per pixel

This way we ensure we have patches for every pixel.

Unfortunately, if we do this we have to exclude same-pixel neighbors, which sux.

check KNN algorithm description in flatfield.tex

Either adjust or comment; there are some shift-scale things there I think we should be doing.

switch to initialization with truth

come up with title for proposal

check in new code

where is the plotting code?

write sensible things about the PSF and what we assume about it

get discussion points into outline form in document

KDE is expensive

~30s for 20k samples in 2D.

Need to look into FFT tricks, perhaps. Are there non-gaussian ways of doing such things?

get in touch with an STScI calibrator in re proposal

What's the status of this?

Thank Weiner and Dalcanton in papers

make appt with Sonntag to discuss believe prop and etc.

patch-based calibration method needs a general name

so that I can refer to something in the paper when I contrast it with other methods.

We could call it "KNN" but that should be the SUB-type of the general method name, which should be something like "patch-based caibration" except that I hate that name. And self-calibration is too general (it covers Holmes's method too).

Work with a subgrid of pixels

Update inside a small patch, allow the flat to vary but only update in a small patch.

compute an entropy

figure out a heuristic entropy to calculate

calculate it

show it decreases as the fit proceeds.

draft budget with CAI

make code to plot a pair of stretches consistently to visualize a flat

show that final noise in inferred flat is (relatively) uncorrelated with flat

for this test, initialize at truth

get decisions made about PSF inference likelihood into code and document

show that flat error is properly related to coverage

the flat rms should scale like coverage to the 0.5 power.

make publication-quality plots that show this

make code to build a webpage summarizing a run.