I’ve updated the CUDA RANSAC Homography code. It now uses OpenCV’s SURF GPU implementation instead of FAST. The CUDA kernel based SVD code has been replaced by a port of GNU GSL’s SVD function. The CUDA SVD code is generic enough that you can easily include it into an existing project, but probably limited to small matrices. You can download it here CUDA_RANSAC_Homography_1.1.tar.bz2.