LRClassifierNotes_170613.txt

Logistic Regression Notes

Description of the process and result of generating the logisitc regression classifiers for pixel class labelling added to the repo on 17/06/2013.


Features

Trained the classifier on the following features:

	*Feature*								*Values per pixel*
	rgbColourValues								3
	hsvColour3DHistogram						
	histogram of oriented gradients				
	local binary pattern						
	texture filter response						


This gives a total of 86 features per pixel.


Feature Parameters

The following parameters were used in feature generation:

	*Parameter name*					 *Value*
	increment (filter)						1
	numHistBins								8
	numGradientBins (HOG) 					9
	cellForm (HOG)							(9,9)
	cellsPerBlock (HOG)						(3,3)
	orientationBins	(LBP)					4
	neighbourhoodRadius (LBP)				2
	filter (window size)					15
	
	
100% shonky - I plucked them out of thin air.



Input data

Generated training, crossValidation and test data set using 5% of the MSRC images.

Ensured the training dataset includes at least 1 example of each class; added a check in the data sampling process to provide this option.




Performance

Evaluated three classifiers (note C param required to be > 0)

	*C parameter*		*training accuracy*		*cv accuracy*	*filename*
	0.000001				5.37%					0.0%			
	0.5						6.339%					10.559%
	1.0