2026-11-13
The lane-detection test: first run
Short answer: the lane model finds the stripes but places them loosely. It got 54 of 327 test photos right; the live model (v39) gets 264. On this first run it is nowhere near the live model.
What a lane model is
Self-driving cars use lane models to find the painted lines on a road. They draw each line as one smooth curve from end to end, instead of placing 8 separate dots the way our current model does. Stripes on a sail look a lot like road lines, so we tried one.
Is it safe to use commercially?
Yes. The lane model code is from UnLanedet (Apache-2.0 licence). The part that "sees" the photo started from Meta's DINOv2, released under the Apache-2.0 licence, which allows commercial use. Nothing came from CULane or any other research-only road-photo set.
How it scored
Same 332 frozen test photos, same answer key and same scoring as the last retrain report. Both columns are the model's raw output before our clean-up steps. This is a model-against-model test and says nothing about what customers see in the app.
| Live model (v39) | Lane model (first run) | |
|---|---|---|
| Photos with all lines right (out of 327) | 264 | 54 |
| Photos passing the strict check | 108 | 0 |
| Typical camber error | 0.3 | 3.5 |
| Typical error at the mast end of a stripe (% of stripe length) | 0.8 | 8.6 |
| Typical error at the leech end | 0.8 | 8.3 |
| Stripes missed | 43 | 452 |
| Extra lines drawn | 39 | 557 |
What it looks like
Green is the answer key. Red is the lane model; blue is the live model. Both are drawn on the upright photo. In every example the red lines follow the right stripes but drift off them by a few pixels and stop at the wrong place; the blue lines sit on the stripe.
What we'd do next
Park it. One run cost about $11 and landed 210 photos behind the live model. That isn't close enough to justify the two extra runs we planned. Its practice score was still climbing when training stopped, so a 3-times-longer run (about $30) is the one more try that could change the picture. We'd do the cheaper tests on the list first.