ExomeDepth works on coverage, and there's functionality within ExomeDepth to determine an optimal reference / set of references to use. See section 5 here in the vignette. ExomeDepth typically goes with highly correlated samples, hence it's essential that your samples are from the same batch of sequencing. One thing to consider is if you have trios or related individuals in a pedigree, or a mix of probands and healthy samples, then which samples should / should not be included in your reference for a given test.
If you want to use the inter/ intra sample variation, then you'd apply this to the imported normalised counts. You could do something like an rLog or VST transformation before hand, but this will take some experimentation as that paper is quite ambiguous in its methods.