Documentaton - BLOCKS - BLOCK_TRAIN_CRF

Train a superpixel CRF from validation images. If validation images are not available, uses training images.

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

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

Required inputs:

db

The database.

segloc

The unary potentials, in the form output by BLOCK_TEST_SEGLOC()

histq

Superpixel histograms.

qseg

Quick shift superpixels.

Options:

bk.method

The training method. Valid methods are: static, gridsearch. Default 'gridsearch'.

bk.goal

The training goal to optimize. Valid goals are: meanacc, intersection-union. Default 'intersection-union'.

bk.luv

Should the color difference be in the LUV space? Default 1.

bk.max_images

If there are more than max_images in the validation or training data, select max_images and use those. Default 1000.

Fetchable attributes:

params

The learned parameters of the CRF.

paramspace

A structure containing the parameters tried at each iteration.