Why segmentation was harder than the neural network

The expensive vision failures were often upstream of inference.

The network was not the first thing to fail. The pixels reaching it were.

The labels were easier than the card

Lens vignetting changed the belt colour across the frame. With one global background colour the documented threshold rose to ~31; the black card body fell below foreground while its white label remained visible.

Local background model

Build 1.9.33 models the belt per pixel as a smooth 4th-order surface in Lab space, robustly fitted over five rounds so parts are outliers.

16 frames instead of one

Empty-belt capture uses 16 frames over ~2 s: median = reference, largest per-pixel deviation = noise map. Glare spots receive a halo around 1.8× radius; large white parts remain valid.

Physics beats statistics

A learned hole signature solved “carabiner around SD card”: a blob with an enclosed hole cannot physically be a plain SD card, so geometry can split a classification that is statistically plausible but physically impossible.