Train under one regime, deploy under another
Out of distribution, and out of templates.
A correction is calibrated on one systematic regime and deployed on another; templates are dropped or padded with irrelevant ones; the map shrinks; holes and depth-driven noise appear.
Distribution shift
Simulation In this slice the training and test worlds share their sky: only the contamination regime changes. A model that memorised the signal while training therefore subtracts it again at test time — read the clean-transfer view with that in mind (see the limitations).
Method
Metric
rows: trained on · columns: tested on · diagonal = in distribution
Feature ablation and spurious features
Metric
Sample size
Clean signal transfer by map size
flexible correctors overfit small maps hardest
Masks and depth-driven noise
Clean signal transfer under messy observing conditions