Fast Matching Pyramids Rotationinvariance

edit

A approach in fast matching seems to be downsizing template and queryimage in pyramids. What would be a good strategy to make this method rotation-invariant in a situation where more than one instances of the (non-scaled) template can arise? I am very sure that the first step has to rotate the

templateImage on smallest pyramid level with big steps of rotation angle

But i do not know how to make a fit, where in each step of the following levels the estimation of the rotation angle gets more preciser. In this SO-Question the last comment refers to the unclear par as "rotate your template roughly on the low resolution levels, and when you trace the template back down to the higher resolution levels"

Can someone give me an idea -perhaps in the form of pseudocode- how such an algorithm could work?