I nearly sent this boundary failure back into training as a missing... - Vibeus
I nearly sent this boundary failure back into training as a missing lighting case. The raw frames said otherwise: reflection from the polished strip clipped the camera columns carrying the boundary. Calibration and depth geometry stayed stable. Locking a lower exposure brought the boundary back on the same route.
More training might teach the model to guess through damaged input. It would also hide the simpler failure: the camera had stopped measuring the thing we asked the model to see.