Detections with Tractor

[andrevitorelli]

tractor

benjamin:

• fwd modelling approach
• big generative modelling fitted to the observed image
• statistical approach
• detection: simple threshold - each band individually x stacked (grz in DESI)

axel:

• i think it's like sextractor (but six sigma cut is too high)

benjamin:

• model is the same for each obj is the same for all the bands
• objects detected on the stack but not single image, the model is forced on the single band with forced photometry. Increases the # of objects detected.

alexandre:

• when they do that, they build a color model of the object

benjamin:

• a kind of deblending with this method

axel:

• the deblending is based on detection in another band

alexandre:

• which is a selection bias

axel:

• i wouldn't call it a deblending

alexandre:

• what do they set as image pixels
• object is all information (? - something more is missing here)

benjamin:

• still in development, many empty sections

axel:

• are all objects fitted simultaneously, or iteratively?

axel:

• do they try models individually, or is it automated in some way?

axel:

• preparation of the sources for DESI

andre:

• so is it a prior?

axel:

• they wanted to, but it was nor working with lensing

Bastien:

• Maxime , CNN for detections. https://www.youtube.com/watch?v=IoNu5L3l8xQ
  (^^^this will be a new full issue/notes soon)

Thomas:

• Maxime's thesis has a review

Axel:

• no other way to do it from scratch not sx
• others will use sx as prior for the foreseeable future

Bastien:

• we might reach computational power to use Bayesian catalogs

extra:

• psf is constructed from PSFEx.