komrad36 / kfnn Goto Github PK

Fastest CPU implementation of a brute-force matcher for 128-float descriptors such as SIFT in 2NN mode, i.e., a match is returned if the best match between a query vector and a training vector is more than a certain threshold ratio better than the second-best match. AVX2 and all the SSEs are used to accelerate computation.

Check out the CUDA version, CUDAKfNN, for significantly more speed.

KfNN supports both raw floats and packed (as uint8_t). Just set the 'packed' boolean flag accordingly in the demo. For CPU packed is faster; for CUDA it's slower.

Float descriptors are slow. Check out my K2NN and CUDAK2NN projects for much faster binary description matching. Use a good binary descriptor such as LATCH where possible.

All functionality is contained in the file KfNN.h. 'main.cpp' is simply a sample test harness with example usage and performance testing.