Can one whole-sail outline catch both misses and junk?

Mixed. It's strong on the miss side — 91.3% of missed stripes still sit inside the sail outline, meaning a search confined to that shape could find them. It's weak as a junk filter — only 36.4% of junk lines fall outside the sail, so it won't clean up most of our false lines.

What it means for SailScan

One global sail shape is a useful guardrail for narrowing WHERE to look for missed stripes (nearly all of them are on the sail, unsurprisingly), but it's not a general-purpose junk filter — most of our junk lines are drawn on the sail too, just in the wrong place, so a whole-sail boundary check won't catch them. Worth keeping as one signal, not a standalone fix.

91.3%

misses inside the sail shape

36.4%

junk caught outside

misses checked

The detail

Check	Result
Missed stripes that fall INSIDE the one-sail-shape fit	63 of 69 (91.3%)
Obvious junk lines that fall OUTSIDE the sail shape (correctly caught)	28 of 77 (36.4%)
Control: correctly-matched real stripes inside the shape (ceiling)	98.2%

We fit one rough outline of the whole sail (from SAM3) per photo, then checked two things: do the stripes we MISS still fall inside that outline (so a band-search inside it could find them)? And do our JUNK lines fall outside it (so the outline could veto them)?

Source: 332-photo holdout, frozen 2026-08-25. Sail outline = SAM3 full-frame mask, reused from the boom-fence spike.

raw data (JSON) →