For a dense sampling of a D dimensional feature space F ⊂ RD, a training
set of size M ≫ 2
D is required. For the cases where M . 2
D, a modification of
the basic algorithm improves the classification performances for sparse datasets,
at the expense of the query time