Detections with Celeste

[andrevitorelli]

Shear Reading 2021-04-08 10:00

Bastien: Celeste

• Not detection, but a full generative model.
• Completely different from sextractor & tractor
• build a model from the parameter disribution space
• more like a model fitting than a detection procedure
• choose 1 reference band and colors to find other filters
• testing on SDSS-like generated images
• some parameters are fixed in advance
• each pixel may have contributions from several objects
• brightness: gamma function distributed
• color distribution: multivariate gaussian
• mixture component is a categorical distribution
• galaxy models: bulge + disk
• mu_s position/detection
• priors from the position, from other surveys
• proceed with fitting
• when no positions available: convolve wihth matched filters to increase snr. find pixels value exceeds local + upper bound of noise.
• Test: Stripe82 using Photo (Lupton) as prior.
• positions are better by ~1-% (N=654)

Axel:

• are they detecting galaxies? (Table 2 of celeste paper)
• why are they detecting more galaxies? Blends?

Andre:

• The input catalogue is just a prior.

Bastien:

• Then the fit is optimized.

Andre:

• Starnet partitions images to control positions/parameter space

Bastien:

• part of LSST DESC, same people
• generative models in starnet represent objects better.

Andre:

• do they quote size/storage demands?

Bastien:

• celeste takes 5 min in a 4 megapixel image with hundreds of obj.
• in starnet tiling, in celeste, are there limits of objects contributing to a pixel?

Axel:

• at least one problem: fit the number of obj at some point: you have to fix the # of obj at some point.
• not see the point, you'll have to fix # at some point.

Andre:

• you can integrate on all distributions (but it ridiculous).
• could this give us a mitigation for the starlink problem?

Bastien:

• Celeste was kind of a first step, starnet came after.

Axel:

• do they assume models for galaxies:

Bastien:

• yes, bulge plus disk

Axel:

• is this used for the detection

Bastien:

• as you fit everything at once, yes, you can't separate it

Axel:

• if you have too much variables, things explode, its computationally difficult, and gains of it are low.

Bastien:

• in the end they do fix a lot of things.

Axel:

• interesting for metadetection. there are artifacts on images that are difficult to treat.
• for blends is a really good way

Bastien:

• in non-blends maybe this wont get better positions, but in blends, maybe yes.
• starnet uses the same tests on coadds

Andre:

• first & second pass

Axel:

• do they need a prior on the psf?

Bastien:

• they use a mix of Gaussian fitted to known stars.