a basic cuda accelerated implementation of the mbm stereo method (12ms per stereo pair on my machine ~5x faster than if build from plain opencv cuda functionalities) The kernels are developed for compute capability 5+ and cuda 7.5 I expect they will compile for lower versions but this needs to be explicitly enabled in the cmake file by changing the CUDA_ARC_XXX defines.
compile instructions mkdir build cd build cmake .. make -j8
./mbm ...