A model for image analysis was trained using data captured with imaging system A. I then deployed the model on imaging system B. System B has better image contrast than system A. Features in the last fully connected layer indicate a domain-shift (TSNE/feature distributions shows distinct differences between features extracted from training versus deployment data). This shift is likely due to the instrumentation/contrast change. I performed an image perturbation analysis to empirically validate this claim. Nevertheless, the performance on system B (deployment) is better than the performance on system A, even though model was trained with data from system A.
The improvement makes sense since the better contrast could highlight more relevant features. However, in literature the consensus is that under domain-shift performance should have decreased. Are there any works that highlight improved performance under domain-shift in image analysis? Any thoughts on potential reasons why a model could perform better under domain-shift?
