The code still needs to be cleaned up (removing print statements, checking some bbox logic works for non-LATISS detectors, etc), but the preliminary results are shown in the attached images, which have matched color bars and scales, and show (clockwise from top left): tl: Current DM code with MEDIAN_PER_ROW; tr: MEDIAN_PER_ROW applied in both serial and parallel directions, no BIAS applied; br: MEDIAN fitting applied in both serial and parallel directions, standard BIAS applied; bl: MEDIAN_PER_ROW applied in both serial and parallel directions, standard BIAS applied.
My initial look at this shows that if we adopt this as the standard processing, we'll need to regenerate the BIAS/DARK/FLAT to match this, otherwise we'll be double-correcting the horizontal structure (as both the parallel overscan and the BIAS attempt to model this trends). The zoom suggests that the amplifier boundaries (which have had a "wobble" on the +x edges) are somewhat better corrected by the parallel overscan, so there should be improvement there. Bad column defects definitely need to be masked better, and the inversion of some of these features suggests we may need to extend those masks into the overscan region (or fully mask the columns).
Improvement
RFC-862 Drop piece-wise "bias jump" overscan correction code.
Paul Price
added a comment - Reviewed on GitHub PR.
The code still needs to be cleaned up (removing print statements, checking some bbox logic works for non-LATISS detectors, etc), but the preliminary results are shown in the attached images, which have matched color bars and scales, and show (clockwise from top left): tl: Current DM code with MEDIAN_PER_ROW; tr: MEDIAN_PER_ROW applied in both serial and parallel directions, no BIAS applied; br: MEDIAN fitting applied in both serial and parallel directions, standard BIAS applied; bl: MEDIAN_PER_ROW applied in both serial and parallel directions, standard BIAS applied.
My initial look at this shows that if we adopt this as the standard processing, we'll need to regenerate the BIAS/DARK/FLAT to match this, otherwise we'll be double-correcting the horizontal structure (as both the parallel overscan and the BIAS attempt to model this trends). The zoom suggests that the amplifier boundaries (which have had a "wobble" on the +x edges) are somewhat better corrected by the parallel overscan, so there should be improvement there. Bad column defects definitely need to be masked better, and the inversion of some of these features suggests we may need to extend those masks into the overscan region (or fully mask the columns).