2026-11-13
Measuring each stripe on a straightened strip (test, not live)
Short answer: not ready to ship. It measures a clean tape very precisely, but on real photos it sometimes jumps onto the wrong edge, such as a sail number or a fold. Your 40-picture vote decides the next step.
What the new step does
- It takes each stripe our current model found and cuts a thin band of the full-size photo along it, straightened out.
- In every sliver of that band it finds the bottom edge of the tape to a fraction of a pixel, then fits one smooth curve through all of them. Glare, seams and numbers get no say, and the curve bridges across them.
- It may only move the line a short distance (1% of the photo's diagonal) and never adds or drops a stripe. If it can't see the tape clearly, it leaves our model's line alone.
Test 1: fake stripes with a known shape
We drew tapes with an exactly known depth and draft onto real sailcloth from our test photos: dark tape, white on white, glare, a seam crossing, sail numbers and blur, 80 of each. Then we compared the readings with the truth.
- The new line sits 0.4 pixels from the true bottom edge. Our current kind of line (8 points) sits 4.6 pixels away.
- No bias. A true 35% draft reads 35.0%. Depth is within 0.05 points of the truth in every condition.
- Steadier numbers on dark tape, glare, seams, numbers and blur. Depth scatters about half as much (0.2-0.3 points instead of 0.5). Draft scatters 3-6 times less (0.2-0.5 points instead of 1.4).
- Weak spot: faint white tape on white cloth. There it sometimes grabs the wrong edge.
Below, an example with sail numbers on top of the stripe. Top: a line like today's. Bottom: the new line on the tape's bottom edge, straight across the numbers. (Made-up stripe on made-up cloth; no customer photos on this page.)
The same with a patch of glare over the stripe:
Test 2: the same sail photographed twice
Customers sometimes scan the same sail twice within a minute or two. The two readings should match. We found 85 such pairs from 43 customers (253 stripes).
| Today's lines | New lines | |
|---|---|---|
| Typical depth gap between the two photos | 1.2 | 1.5 |
| Typical draft gap (steady method) | 2.8 | 2.9 |
The new lines made the two photos agree a little less, except on bottom stripes, where it is a tie. That is the wrong-edge problem showing up on real photos.
Test 3: our 332 test photos
Caveat first: the answer lines on these photos were mostly drawn through the middle of the tape, or loosely. The new step deliberately puts the line on the bottom edge, so it can score worse here even where it is right. This number alone does not decide.
| Today (v39) | New step | |
|---|---|---|
| Photos with every line right (normal check), out of 327 | 264 | 264 |
| Strict check | 109 | 77 |
| Typical depth error, all / top / middle / bottom stripes | 0.3 / 0.6 / 0.3 / 0.2 | 0.6 / 1.2 / 0.6 / 0.2 |
| Typical draft error (steady method) | 2.4 | 3.0 |
| Where the line starts and ends | unchanged | unchanged |
Bottom stripes come out level. The losses are on the short top and middle stripes.
Time per scan
About 0.1 seconds per photo on one processor core (0.03 s per stripe).
Next step
- Your vote on 40 zoomed stripes, where you pick the picture whose line sits right on the bottom edge of the tape (link in the Boats chat).
- Teach the step to trust a column only where it also sees the tape's top edge running parallel above it. A sail number or a fold has no parallel partner. Then rerun these same tests.
- Nothing changes in the app until a version wins.