Figure 4
From: Physical imaging parameter variation drives domain shift
Smoothed, mean training schedule over 5 runs. Resnet50 trained on 1000 A images per class (under vs over 58 years). 150 images for each testing class from each dataset. This illustrates the clear generalization gap between A test I.I.D test sets and all other O.O.D test sets. A 5–10% drop in accuracy on test images that visually similar illustrates that PIP variation is a large contributing factor of domain shift. See supplementary Fig. S2. for examples with other Resnet networks each with similar results.
