The Jupyter notebooks https://github.com/lsst/meas_deblender/blob/u/fred3m/deblender/examples/testNmfLeeSeung.ipynb and https://github.com/lsst/meas_deblender/blob/u/fred3m/deblender/examples/testNmfHSC.ipynb illustrate the current status of the NMF deblender when running on simulated and HSC data respectively. The simulated data was created by Bob Armstrong as part of
Until we have good PSF homogenized data to run through the NMF deblender it is difficult to say how the deblender performs on real data, however it's performance on the simulated data is encouraging. Compared to the current deblender, which the NMF deblender uses as initial conditions, the NMF algorithm shows significant improvements, although more thorough analysis needs to be performed to verify that this holds for the general case. It does, however, perform poorly on the current calibrated HSC images due to color gradients created by their differing PSF's.
There are still many details that need to be worked out, including fine tuning of the constraints and a reasonable stopping criteria for each blend, but the biggest point of failure is the inability to deblend sources with nearly the same SED (as we expected). One of the first tasks in the next sprint will be to develop a constraint that enforces monotonicity (
DM-9143), which ideally will solve the problem of degenerate SED's. This will require the creation of a linear monotonicity operator to act on the intensity matrix H that will penalize pixels that are not monotonically decreasing from its center.