As stated before the value that comes from the saliency map (an other analysis techniques) is not universal. It is a value that depends on a particular training and neural network. We may however say that the value should be in relation to other analysis value, such that we could get from one to the other by a monotonic transformation. Monotonic equivalence can be measured with rank correlations (B.2).
The rank correlation may however be too pessimistic, and they may emphasize the ordering too much. Among the correlations that do not emphasize the ordering so much is the ordinary correlation coefficient:
By some called the Pearson correlation coefficient. This is however rather metric.
An other correlation is the Lin concordance correlation [36] used for reproducibility:
As the saliency value vary depending on the specific training set and neural network this measure can not be used.
One may argue that the comparison between maps rather should show if there is difference between the blobs (The small areas with the highest saliency): A qualitative description accounting new blobs or blobs with an other form --- a higher contrast.
