Documentaton - BLOCKS - BLOCK_KER

This block constructs an SVM kernel on two sets of segment ids or segment ids and superpixel ids.

BK = BLOCK_KER() Initializes the block with the default options.

BK = BLOCK_KER(BK) Executes the block with options and inputs BK.

Required Inputs:

hist

Histograms or superpixel histograms.

Options:

bk.kernel

The type of kernel to use. The following are supported:

L2 - k(x,y) = sum x .* y L1 - k(p,q) = sum min(p,q) CHI2 - k(p,q) = sum 2 (p.*q) / (p+q) HELL - k(p,q) = sum sqrt(p.*q) / 4

DL2 - k(x,y) = sum (x-y).^2 DL1 - k(p,q) = sum |p-q| DCHI2 - k(p,q) = sum (p-q).^2 / (p+q) DHELL - k(p,q) = sum (p.^.5 - q.^.5).^2 / 4

 Here p,q denote non-negative vectors, usually l1 normalized
 (histograms). Notice that DL2, DL1, DCHI2 and DHELL are not
 kernels, but the corresponding metrics (this is useful to
 construct RBF kernels). See also KERNEL_FUNCTION() and
 VL_ALLDIST2(). This parameter is required and there is no
 default.
bk.normalize

Normalize the data by the specified norm (L1, L2, ...) before computing the kernel. See also KERNEL_FUNCTION(). This parameter is required and there is no default.

bk.row_seg_ids

The seg_ids of the rows of the kernel matrix. If use_segs and seg_neighbors are set, row_seg_ids is a Nx2 matrx, where the first column denotes the seg_id and the second column denotes the superpixel.

bk.col_seg_ids

The seg_ids of the columns of the kernel matrix. If use_segs and seg_neighbors are set, col_seg_ids is a Nx2 matrx, where the first column denotes the seg_id and the second column denotes the superpixel.

bk.use_segs

Use superpixel neighborhoods.

bk.seg_neighbors

The size of the neighborhood to use. If this parameter is ommited, even if bk.use_segs is true, segs will not be used.

bk.split

Split the job into a number of subtasks. 0 disables splitting.