KGraph is the next generation indexing technology for K-NN graph construction and nearest neighbor search.
- It only assumes that a distance/similarity can be computed on any pair of data objects, absolutely nothing else about the form of the underlying data.
- It is usually faster than popular libraries like FLANN on the distance metrics they support.
Some of the optimizations are back ported to my previous library nndes. Still, KGraph is about 2x faster than the latest nndes in building a K-NN graph. I have long abandoned LSHKIT because LSH is so hopelessly slow and/or inaccurate.
Speedup over FLANN brutal force search for 10-NN from 1 million SIFT features. At 0.9 recall,
See Benchmark for details.
Free for academic, personal and other non-commercial use. For licensing and other questions, please contact Wei Dong at the following address.