UberSeg end to end on #933's tile detection (SEG_VIGNET from the tile SExtractor run)

astra.yaml

@@ -1421,22 +1421,25 @@ analyses:
       defect_fill:
         label: Image content of defect pixels before metacal
         rationale: >-
-          A defect pixel (nonzero instrument flag, zero exposure weight or
-          invalid background RMS) gets weight 0 in prepare_ngmix_weights. Its
+          A defect pixel (nonzero instrument flag, zero exposure weight,
+          invalid background RMS, or off-tile) gets weight 0 in prepare_ngmix_weights. Its
           image value still matters: ngmix's metacal deconvolves, shears and
           reconvolves an InterpolatedImage of the whole image and copies the
           weights through, so a zero-weight pixel's content spreads into the
           weighted pixels within about a PSF width. DES's ngmixer fills for
           that reason: "it may be important for codes that take moments or use
-          FFTs". In the committed default (`BLEND_HANDLING = noisefill`),
-          masked pixels are replaced with independent noise at the per-pixel
-          background RMS when supplied, or the stamp's robust noise scale
-          otherwise. With `BLEND_HANDLING = uberseg`, the image is left
-          untouched and weights are zeroed on defects and neighbour-side
-          pixels. [LINT] the prepare_ngmix_weights docstring says noisefill
-          keeps the weight of filled pixels (the code zeroes it), and the
-          ngmix_runner comment says noisefill fills neighbour pixels (it fills
-          flagged pixels and leaves neighbours untouched). The committed fill
+          FFTs". Under every BLEND_HANDLING, the committed uberseg included,
+          and the default DEFECT_FILL =
+          noise, defect pixels are replaced with independent noise at the
+          per-pixel background RMS when supplied, or the stamp's robust noise
+          scale otherwise. The tile VIGNET holds -1e30 on other detections'
+          footprints and beyond the tile's edge (written by SExtractor). The runs of entirely -1e30 stamp
+          rows and columns that start at a stamp border (the off-image part of
+          a rectangle clip) are off-tile: the epoch holds the object's own
+          light there, so they are defects, flagged 2**10. The other markers,
+          including a neighbour footprint that completes an interior row
+          beside an off-tile band, are neighbour pixels, not defects;
+          blend_handling decides their treatment. The committed fill
           uses the unsymmetrized defect set: DES symmetrized its masks, but
           four-fold symmetrization quadruples m and still leaves an additive c1
           (symmetrized_4fold_noise). The cost of not symmetrizing, a hole in
@@ -1458,7 +1461,11 @@ analyses:
           interpolate:
             label: Interpolate short bounded runs; noise-fill the rest
             description: >-
-              Not implemented. PR #916 measures c1 = -1.3e-3 for a column
+              DEFECT_FILL = interpolate: row or column runs of at most
+              MAX_INTERPOLATED_RUN = 3 defect pixels that stop short of the
+              stamp border are Clough-Tocher interpolated from the clean pixels
+              within SUPPORT_RADIUS = 4 px; other defects are noise-filled.
+              PR #916 measures c1 = -1.3e-3 for a column
               8 px from a 0.5 arcsec galaxy and +2e-6 when the weight is also
               zeroed on its quarter-turn orbit. For a 3-px bleed 6 px from that
               galaxy, PR #916 measures m11 = +0.89% when the fill is
@@ -1476,11 +1483,11 @@ analyses:
               with symmetrization against 0.7e-4 without it.
             insights: [mask_bad_column_symmetrize, mask_des_defect_practice, mask_fixed_orientation]
           raw:
-            label: "No fill: raw defect values (BLEND_HANDLING = uberseg)"
+            label: "No fill: raw defect values"
             description: >-
-              With BLEND_HANDLING = uberseg, defect pixels keep weight 0 but
-              their image values stay untouched in the image that metacal
-              deconvolves, shears and reconvolves.
+              Defect pixels keep weight 0 but their image values stay
+              untouched in the image that metacal deconvolves, shears and
+              reconvolves.
             excluded: true
             excluded_reason: >-
               Metacal acts on every pixel regardless of weight, so raw defects
@@ -1490,12 +1497,15 @@ analyses:
       blend_handling:
         label: Neighbour treatment before metacal
         rationale: >-
-          Covers only pixels shared with a neighbour; defect_fill is coupled to
-          it through BLEND_HANDLING. noisefill (the default; the committed
-          config sets no key) leaves neighbours fully weighted and untouched.
-          uberseg zeroes the weight of pixels nearer a neighbour's coadd
-          segmentation footprint than the target's (DILATE_NEIGHBOUR, default 1)
-          and leaves the image untouched, as official uberseg does:
+          Covers only pixels shared with a neighbour; defect_fill does not
+          depend on it: defects are filled the same way under every
+          BLEND_HANDLING, and the epoch cuts never count neighbour pixels.
+          The committed choice is uberseg (the workflow's default
+          blend_handling), which ignores the -1e30 neighbour markers in the
+          tile VIGNET and zeroes the weight of pixels nearer a neighbour's
+          coadd segmentation footprint than the target's (DILATE_NEIGHBOUR,
+          default 1), leaving the neighbour side's image untouched, as
+          official uberseg does:
           esheldon/meds get_uberseg returns a weight map (a
           nearest-segment-pixel Voronoi split). DES Y1's fiducial metacal and
           the last Y3 config ran on uberseg-weighted stamps with raw neighbour
@@ -1512,24 +1522,31 @@ analyses:
           with simulations (mask_des_y1_uberseg_only,
           mask_blend_bias_detection). Noise-filling the neighbour side would
           instead cut the target's own light along an unsheared boundary, a
-          sharp edge that rings in the FFTs. The recommended comparison arm is
-          uberseg (weight-only), with defect_fill held equal across arms.
-        default: none
+          sharp edge that rings in the FFTs. The alternative, noisefill
+          (ngmix's own default, and the workflow's blend_handling:
+          noisefill), gives weight 0 to the pixels marked -1e30 in the tile
+          VIGNET on other detections' footprints (about 93% of
+          neighbour-footprint pixels on a simulated tile), excluding the
+          off-tile rows and columns, which are defects, and replaces them with
+          noise; unmarked neighbour pixels keep their weight and light. It is
+          the comparison arm, with defect_fill held equal across arms.
+        default: uberseg
         options:
-          none:
-            label: No neighbour treatment (BLEND_HANDLING = noisefill)
+          noisefill:
+            label: Noise-fill the marked neighbour pixels (BLEND_HANDLING = noisefill)
             description: >-
-              Neighbour pixels keep their full weight and image values, so the
-              fit sees all neighbour light, which biases shapes toward
-              neighbours (Jarvis et al. 2016); this masks less than even their
-              plain segmentation map. A candidate cause of the FLAGS=2 B-modes
+              Pixels marked -1e30 in the tile VIGNET get weight 0 and noise.
+              The fill stops at the marked footprint, so unmarked neighbour
+              pixels and the neighbour's wings keep their weight and light,
+              which biases shapes toward neighbours (Jarvis et al. 2016). A
+              candidate cause of the FLAGS=2 B-modes
               investigated in #814.
             insights: [mask_uberseg_neighbour_bias]
           uberseg:
             label: UberSeg, weight-only
             description: >-
-              As in DES Y1/Y3. Needs the coadd segmentation stamp
-              (SEG_VIGNET_PATH). DILATE_NEIGHBOUR absorbs the coadd-vs-epoch
+              As in DES Y1/Y3. Needs the coadd segmentation stamp (the tile
+              catalogue's SEG_VIGNET). DILATE_NEIGHBOUR absorbs the coadd-vs-epoch
               overlay offset: ShapePipe reuses one coadd seg stamp for every
               epoch where MEDS reprojects it.
             insights: [mask_uberseg_weight_only, mask_uberseg_neighbour_bias, mask_des_y1_uberseg_only, mask_blend_bias_detection]
@@ -1553,11 +1570,13 @@ analyses:
       central_defect_veto:
         label: Per-epoch veto on a defect near the stamp centre
         rationale: >-
-          Not on develop; implemented on feat/defect-fill-veto (7777181b).
-          There an epoch is dropped when a defect pixel lies strictly closer
-          to the stamp centre than its fill's radius, beside the
-          masked-fraction cut in the epoch loop. The veto reads only the
-          defect mask, so it selects on nothing shear-responsive; for the same
+          An epoch is dropped when a defect pixel lies strictly closer to the
+          stamp centre than its fill's radius, beside the masked-fraction cut
+          in the epoch loop; EPOCH_CENTRAL_DEFECT_RADIUS = 0 disables it.
+          The tile VIGNET's -1e30 neighbour markers are not defects: all epochs
+          share the tile VIGNET, so a neighbour inside the radius would drop
+          every epoch. Off-tile pixels are defects, so an object near the
+          tile edge is vetoed. The veto reads only the defect mask, so it selects on nothing shear-responsive; for the same
           reason the radii are fixed rather than scaled by galaxy size: 10 px
           for noise-filled pixels (EPOCH_CENTRAL_DEFECT_RADIUS) and 7 px for
           interpolated ones (EPOCH_INTERPOLATED_DEFECT_RADIUS,
@@ -1568,16 +1587,18 @@ analyses:
           wide defects through the elliptical PSF sit at the 1% bound (m11 =
           -0.98%; -0.24% at 11 px), and noise fill needs 14 px for 0.7 and 0.9
           arcsec galaxies. Measured on feat/defect-fill-veto and
-          feat/defect-interpolation.
-        default: disabled
+          feat/defect-interpolation. The 7 px radius applies only under
+          DEFECT_FILL = interpolate; with the committed noise fill every defect
+          takes the 10 px radius.
+          Values:
+          EPOCH_CENTRAL_DEFECT_RADIUS = 10;
+          EPOCH_INTERPOLATED_DEFECT_RADIUS = 7.
+        default: fixed_radii
         options:
           disabled:
-            label: No central veto (committed code)
+            label: No central veto (EPOCH_CENTRAL_DEFECT_RADIUS = 0)
           fixed_radii:
             label: Fixed radii, 10 px for noise-filled and 7 px for interpolated defects
-            description: >-
-              Implemented on feat/defect-fill-veto and feat/defect-interpolation,
-              not on develop.
           size_scaled_radius:
             label: Veto radius scaled by galaxy size
             excluded: true
@@ -1587,14 +1608,16 @@ analyses:
       epoch_masked_fraction_cut:
         label: Per-epoch masked-fraction cut
         rationale: >-
-          [HARDCODED] on develop, an epoch whose stamp has more than 1/3 of its
-          pixels flagged (any nonzero flag bit, including the tile-coverage bit
-          2**10 set where the tile vignet is off-image) is dropped from the
-          multi-epoch fit; an object with no surviving epoch has no shape.
-          Zero-weight and invalid-RMS pixels are not counted. On
-          feat/defect-fill-veto the cut counts the raw, unsymmetrized defect
-          set (flagged, zero-weight and invalid-RMS pixels) against
-          EPOCH_MASKED_FRACTION_CUT, default 1/3. Before the cut, an epoch is
+          An epoch whose stamp has more than EPOCH_MASKED_FRACTION_CUT
+          (default 1/3) of its pixels in the raw, unsymmetrized defect set is
+          dropped from the multi-epoch fit: flagged pixels (any nonzero
+          exposure flag bit), zero-weight pixels, invalid-RMS pixels and
+          off-tile pixels (whole -1e30 rows and columns at the tile VIGNET's
+          border, flag 2**10), the set defect_fill fills. On a 51-px stamp an object
+          within about 8.5 px of the tile edge fails the 1/3 cut. The other
+          -1e30 markers, neighbour footprints, are not counted: every epoch
+          shares the tile VIGNET, so a large neighbour would drop them all. An object with no surviving epoch has no
+          shape. Before the cut, an epoch is
           dropped silently if its galaxy stamp is all zeros or its
           background-subtracted sigma_mad is not positive. DES was stricter:
           Y1 rejected any epoch with a masked or zero-weight pixel, or with
@@ -1610,12 +1633,12 @@ analyses:
             label: 1/3 of the stamp in the defect set
             description: >-
               Drop an epoch only if more than 1/3 of the stamp pixels are
-              defects (on develop, flagged pixels).
+              defects.
           ten_percent:
             label: 10% (DES Y3 / Y6)
             description: >-
-              Not a default; EPOCH_MASKED_FRACTION_CUT = 0.1 on
-              feat/defect-fill-veto. Matches DES Y3 max_zero_weight_frac and
+              Not a default; EPOCH_MASKED_FRACTION_CUT = 0.1. Matches DES Y3
+              max_zero_weight_frac and
               Y6 max_masked_fraction.
             insights: [mask_multi_epoch_drop]
           any_masked:

src/shapepipe/modules/make_cat_package/make_cat.py

@@ -113,7 +113,8 @@ def save_sextractor_data(final_cat_file, sexcat_path, remove_vignet=True):
     sexcat_path : str
         Path to SExtractor catalogue to save
     remove_vignet : bool
-        If ``True`` will not save the ``VIGNET`` field into the final catalogue
+        If ``True`` will not save the ``VIGNET`` and ``SEG_VIGNET`` stamp
+        fields into the final catalogue
 
     Returns
     -------
@@ -127,6 +128,7 @@ def save_sextractor_data(final_cat_file, sexcat_path, remove_vignet=True):
     data = np.copy(sexcat_file.get_data())
     if remove_vignet:
         data = remove_field_name(data, "VIGNET")
+        data = remove_field_name(data, "SEG_VIGNET")
     cat_size = len(data)
 
     tile_name = os.path.basename(sexcat_path)