I think this is one of the most thorough and independent comparisons out there. All the authors of imputation software (IMPUTE2, BEAGLE and MACH) have performed some sort of comparison in their papers pinpointing marginal benefits for their respective tools. The bottom line of golden helix' comparison is:
Neither tool distinguishes itself as superior.
In case of IMPUTE2 you have the benefit of multiple reference panels + you don't have to care about population composition in the reference panel.