Package boofcv.alg.scene.nister2006

Class LearnNodeWeights<Point>

java.lang.Object

boofcv.alg.scene.nister2006.LearnNodeWeights<Point>

public class LearnNodeWeights<Point>
extends Object

Learns node weights in the HierarchicalVocabularyTree for use in RecognitionVocabularyTreeNister2006 by counting the number of unique images a specific node/word appears in then computes the weight using an entropy like cost function. Functions must be called in a specific order:

1. reset(boofcv.alg.scene.vocabtree.HierarchicalVocabularyTree<Point>)
2. addImage(java.util.List<Point>)
3. fixate()

Each image used to train the tree should have it's descriptors passed in to addImage(java.util.List<Point>).

Field Summary

Fields

Modifier and Type	Field	Description
boolean	checkEveryNodeSeenOnce	Sanity check.
ConfigLength	maximumNumberImagesInNode	If a node has more than this number of images passing through it, it's weight is set to 0.
protected HierarchicalVocabularyTree<Point>	tree	Tree which has been learned already but with unspecified weights

Constructor Summary

Constructors

Constructor	Description
LearnNodeWeights()

Method Summary

Modifier and Type	Method	Description
void	addImage(List<Point> descriptors)	Adds a new image to the weight computation.
void	fixate()	Call when done.
void	reset(HierarchicalVocabularyTree<Point> tree)	Initializes and resets with a new tree.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Details

tree

protected HierarchicalVocabularyTree<Point> tree

Tree which has been learned already but with unspecified weights

maximumNumberImagesInNode

public ConfigLength maximumNumberImagesInNode

If a node has more than this number of images passing through it, it's weight is set to 0. The idea being that there is little information gained by considering this node, but it adds significantly to the cost of searching the tree as many images now need to be considered.

checkEveryNodeSeenOnce

public boolean checkEveryNodeSeenOnce

Sanity check. If true it will make sure every node is observed by an image. This is true when the training set and the set of images passed into this are the same. This is set to false by default because in rare situations floating point noise can cause a false positive.

Constructor Details

LearnNodeWeights

public LearnNodeWeights()

Method Details

reset

public void reset(HierarchicalVocabularyTree<Point> tree)

Initializes and resets with a new tree. Reference to the passed in tree is saved.

addImage

public void addImage(List<Point> descriptors)

Adds a new image to the weight computation. It's highly recommended that the same images used to train the tree be added here. This will ensure that all nodes are filled in with a valid weight.

Parameters:

descriptors - Set of all image feature descriptors for a single image

fixate

public void fixate()

Call when done. This will compute the weight using an entropy like function: weight[i] = log(N/N[i]) where N[i] is the number of times a specific node/work appears at least once in the images.