This is a purely educational project attempting to create a simple library for sparse feature matching and homography estimation. The short term goal is to learn more about feature matching. The long term goal is to create a suite of tools that allow me to do experiments with projective and camera based multiple view geometry.
I use a simple Forster-Harris feature detector (corner detector as some call it) to measure the 'goodness' of a point as the center of a feature. I then reduce the 'corner score map' by doing a non-adaptive non-maximal suppression sub-sampling. Since the long term goal is to use point correspondences to create a homography describing the relationship between the given images, I do not need a terribly large (or consistent) number of point correspondences; really, the things that matter are that the point correspondences are high quality (accurate and precise), that they cover a wide range of the image, and that they are not collinear.
I am currently in the process of implementing the MOPS feature description/matching algorithm as described in Brown, Szeliski, and Winder's "Multi-Image Matching using Multi-Scale Oriented Patches."
The plan here is to use the normalized Direct Linear Transformation (DLT) using the point correspondences computed above. I am pretty new to this stuff and I have no idea how well this will work aside from some reading, so we will see how this goes and maybe try an iteratively based refinement of some sort.