PSF models

In the morning, Fergus and I discussed the exoplanet imaging project we are doing, and funding opportunities related to it. The project can be seen as an attempt to make an extremely precise model of a point-spread function in a very non-trivial optical system. We also discussed relationships between computer vision and astronomy, in the context of things I learned at AISTATS.

In the afternoon, Federica Bianco (LCOGT) showed me some lucky imaging data where she finds that the Hirsch et al. online blind deconvolution method does not give good results, whereas old-school lucky imaging works okay. In a deconvolution problem, or an image modeling problem, you can either put the structure you see in the image into the scene or into the PSF. Where the structure goes is explicitly degenerate, although some choices might have much higher marginalized likelihood under sensible priors, or much lower cost under properly regularized cost functions. The challenge is setting those priors or cost functions. (And sampling or optimizing them!) Anyway, Bianco challenged Foreman-Mackey and me to beat her lucky imaging stack with a deconvolution run. We accepted.