Updating on my previous post, I found a tiny bug when calculating the inliers. Still don’t know why the normalisation code produces less inliers though. So with the fixed code the inlier results without normalisation for Kyral castle and the house are 437 and 158 respectively, instead of 402 and 151 as reported before. The normalisation and bias random flag can be toggled in CUDA_RANSAC_Homography.h using the #define at the top. The new code can be downloaded from the same link as before:
The code was written on Linux using CodeBlocks. If you don’t have both it should not be difficult to manually create a project in the IDE of your choice. If you try to compile on Windows you will have to get rid of the Linux specific timing code.