Outline of the paper

[b-biswas]

Opening this issue to exchange on what we want to study/plot in the paper for Debvader.

As of now we discussed going back to the previous setting (as in Bastien's paper) and trying to fix the issue of high residuals. The idea will be to start from simpler light profiles and proceed step by step to complex simulations.

Paper plan (preliminary):

• [ ] #83
  • [ ] Parametric models / (real galaxies)
  • [ ] BTK / (CosmoDC2)
  • [ ] Make LSST-like fields in multi-band
• [ ] Method
  • [ ] Check Bastien's paper and use that as a starting point
  • [ ] Extension with the iterative procedure
• [ ] Metric (here we only describe how the metric is built)
  • [ ] Do we need a blending metric? Or we try to have "realistic" blended scene? (I think this has already been discuss in Bastien's paper)
  • [ ] How do we asses the quality of deblending? #80
    • [ ] Check literatture (Scarlet, BLISS?, MOF in DES/metadetect)
    • [ ] Pixel error? (not the greatest..)
    • [ ] Bias in flux/color -> impact on redshift (compare measured flux by SExtractor with and without deblending)
    • [ ] Bias in size -> impact on morphology (same)
    • [ ] Bias in shear recovery (see project with Thomas)
  • [ ] Comparison with Scarlet / SExtractor / (other ?) (I am adding SExtractor because it more or less comes for free..)
    • [ ] Previously mentioned metrics
    • [ ] #81
      • [ ] How do we reliably measure the speed? (this is very important if we want people to believe the results)
• [ ] Results (here we show the results of the metric)
  • [ ] Pixel error
  • [ ] Flux error
  • [ ] size error
  • [ ] (shear error)
  • [ ] SNR dependence
• [ ] Comparison (might be merge with previous section)
• [ ] Discussion
  • [ ] Improvement on science
  • [ ] What is the best solution for deblending in terms of speed AND accuracy?

[BastienArcelin]

Two comments on that:

• for the simulation part: not sure that coming back to parametric models of GalSim is useful as we already show in our previous paper that the network was working fine on these galaxies. People will be much more interested on the results on DC2 like simulations (or even directly DC2)
• for the method part: do not forget to add the probabilistic output: this is the only deblender that outputs a probability distribution per pixel

[aguinot]

• for the simulation part: not sure that coming back to parametric models of GalSim is useful as we already show in our previous paper that the network was working fine on these galaxies. People will be much more interested on the results on DC2 like simulations (or even directly DC2)

DC2 is also using parametric models if I am not mistaken, right? Right now I put those just to fill the section but will see as we do the tests what we use at the end for the paper.

• for the method part: do not forget to add the probabilistic output: this is the only deblender that outputs a probability distribution per pixel

I don't remember exactly but I thought the probabilistic output was already included in your paper?