Hello maurello,
I mean that for the project which is based on multiple poorly connected components it may be impossible to estimate the tie point variance. However, you can try to use gradual selection filters, if you need to clean up some outliers.
And you are correct, that resetting the alignment will result in some missing parts of the final model, so probably you shouldn't sacrifice the alignment in order to get variance calculation working for one or two major alignment components.
But also please note that the cameras with just a few valid tie points will not be used for the depth maps generation procedure, and if there are clusters of such cameras, you may still observe poorly reconstructed geometry in the related area or even the holes in the model.