The photogrammetric processing should still provide equivelent results given the exact same input, it is simply a soultion to sets/blocks of linearized equations. The desired precision scale should also have no impact on the result. I am guessing the error is in the interest point extraction and matching, but its just a guess.