I am working on the implementation of a location recognition pipeline as detailed in the section 5.4 of the paper Leveraging 3D City Models for Rotation Invariant Place-of-Interest Recognition:
Vocabulary tree + Inverted File System + SIFT detector/descriptors (discarding features not lying on a building using geometric models) + RANSAC for estimating an affine model
I understand most of the elements there but they don't detail much on which types of geometric models are they using, nor what to discard, features from the database of images or from the query image, and neither how are they applying them.
I'm thinking on something like RANSAC for estimation of an homography but is only an assumption.
My question is which geometric models would you suggest me to use in this case (urban environment images of common city landmarks)?