Repository navigation
Check that sim-side tile SExtractor reproduces the DR6 detections data uses #929
Description
Activity
Measured on CFIS.186.307:
workflow/config/cfis/default_tile.sex(SExtractor 2.25.0, DR6 weight map,gauss_3.0_7x7) run on the DR6 image, matched by position to the DR6 catalogue.all mags 20 < MAG_AUTO < 24.5 DR6 / SExtractor objects 41,246 / 41,201 21,249 / 21,245 Completeness (DR6 with a SExtractor match) 99.89% 99.98% (4 missing) Purity (SExtractor with a DR6 match) 100% 100% Matched centroid offset (p99) 7e-5 px same |ΔMAG_AUTO| > 0.01 0.2% 6 objects (max 0.048) |FLUX_RADIUS ratio − 1| > 1% 0.7% 7 objects The numbers are the same for any match radius from 0.1 to 3 px, and completeness is flat in FLUX_RADIUS, including sizes near the PSF (FLUX_RADIUS 1.5–2.2 px). The two segmentation maps flag exactly the same pixels. No object is matched more than once in either direction. Every one of the 45 DR6-only objects is a faint deblended child (median mag 25.4; 4 fall inside 20–24.5) inside the footprint of one of 12 bright parents (mag 11.5–17.8). In those footprints, MegaPipe's deblender splits off children that SExtractor 2.25 leaves merged. The MegaPipe SExtractor version isn't recorded in the catalogue, so a version difference is my inference.
Verdict: the detection configs are effectively identical.
default_tile.sexreproduces DR6 detection, deblending, photometry and sizes object for object in the shear range. The only difference is deblending inside bright-star or bright-galaxy footprints, which is at the 2e-4 level and mostly in regions that get masked. One tile is enough for this, because it tests the configuration and code, not the sky: the agreement is exact to floating-point precision, and more tiles would only add more of the same rare edge cases.What this doesn't cover: whether detection on the simulated images matches detection on the data. The config is the same, but the noise, background and weight maps in the sims are not the data's. Next step: compare the detection count, the MAG_AUTO/FLUX_RADIUS distributions and the deblend rate between a sim tile and its data tile (for example 202.301 on Nibi, whose 35,924/36,065 match at 2 px is lower than the 99.89% here, which is worth understanding), and record the result above in
detection.detection_threshold_policy.Scripts, logs and plot on candide:
/n17data/cdaley/scratch/detection-dr6-vs-sex/(dr6_vs_sex_186307.png).Claude Opus 5.5 on behalf of Cail
Since #897, data takes its tile galaxy sample from the UNIONS DR6 catalogue, while image simulations detect with ShapePipe's own tile SExtractor run, configured with MegaPipe-like parameters (
default_tile.sex: threshold 1.0, minarea 3, 7x7 FWHM-3 kernel, DEBLEND_MINCONT 0.002, BACK_SIZE 512). Shear calibration from the sims assumes the two selections match: detection-dependent multiplicative bias, blending, and the neighbour footprints used for masking (catalogue_neighbour_marking). Nothing checks this on output. SExtractor version, weights and background inputs can all differ from MegaPipe's run.Proposed check: run
tile_detection=sextractoron a handful of data tiles and match against DR6, comparing completeness and purity vs magnitude and size, and the deblend rate. The Nibi run already has one such pair (tile 202.301: 35,924 of 36,065 objects match within 2 px). If they agree, cite the result in thedetection.detection_threshold_policyrecord; if not, tunedefault_tile.sexor record the mismatch as a calibration systematic.Claude Opus 5.5 on behalf of Cail