org.openimaj.demos.sandbox.tldcpp.detector

Class NNClassifier

java.lang.Object

org.openimaj.demos.sandbox.tldcpp.detector.NNClassifier

public class NNClassifier
extends Object

The third and most powerful, but equally most slow parts of the DetectorCascade. Holding a list of falsePositives and truePositives, a classification score can be ascribed to a new patch which can be used as a confidence that a given patch is positive. This is calculated using the correlation between the new patch and the false positive and falst negatives such that: confidence = dP / (dN + dP) and dP = max(corr(patch,truePositives)) dP = max(corr(patch,falsePositives)) if no true positives have been seen, classify will always return 0 if not false positives have been seen, classify will always return 1 classify is used by filter such that if the confidence of a patch is larger than thetaTP the patch is though to be a good patch for the object.

Author:

Sina Samangooei (ss@ecs.soton.ac.uk)

Field Summary

Fields

Modifier and Type	Field and Description
boolean	enabled whether this stage is enabled
float	thetaFP Used as the lower bound of a historysis threshold (i.e.
float	thetaTP Used as the upper bound threshold

Constructor Summary

Constructors

Constructor and Description
NNClassifier() Sets thetaFP as 0.5f and thetaTP as .65f

Method Summary

Modifier and Type	Method and Description
float	classifyBB(FImage img, Rectangle bb)
float	classifyPatch(NormalizedPatch patch)
boolean	filter(FImage img, int windowIdx)
List<NormalizedPatch>	getPositivePatches()
void	learn(List<NormalizedPatch> patches) Given a list of patches, classify each patch.
void	release() clear falst positives and true positives

Methods inherited from class java.lang.Object

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

Field Detail

enabled

public boolean enabled

whether this stage is enabled

thetaFP

public float thetaFP

Used as the lower bound of a historysis threshold (i.e. if a detection is made with a confidence over thetaTP, the next detection can be a little worse by matching this)

thetaTP

public float thetaTP

Used as the upper bound threshold

Constructor Detail

NNClassifier

public NNClassifier()

Sets thetaFP as 0.5f and thetaTP as .65f

Method Detail

release

public void release()

clear falst positives and true positives

classifyPatch

public float classifyPatch(NormalizedPatch patch)

Parameters:

patch -

Returns:

The confidence that a given patch is a postive

classifyBB

public float classifyBB(FImage img,
                        Rectangle bb)

Parameters:

img -

bb -

Returns:

confidence of the bb in image

filter

public boolean filter(FImage img,
                      int windowIdx)

Parameters:

img -

windowIdx -

Returns:

Filter a window by getting its confidence and returning true if confidence > thetaTP

learn

public void learn(List<NormalizedPatch> patches)

Given a list of patches, classify each patch. If the patch is said to be positive and has a confidence lower than thetaTP add the patch to the true positives If the patch is said to be negative and has a confidence higher than thetaFP add the patch to the false positives

Parameters:

patches -

getPositivePatches

public List<NormalizedPatch> getPositivePatches()

Returns:

the positively classified patches