# LBP features and cascade classifier training, model details

So I figured out the complete meaning of every parameter inside the XML model generated when doing haartraining. Basically it is

• a set of stages with nodes, with a node for each weak classifier of a stage, lets assume stumps, containing its associated feature index and the treshold applied to the retrieved feature value over all the training samples of that stage.
• a set of features where each feature is described by 2 rectangles, of which a weight is added as a fifth parameter. Then in the integral image the sum of the rectangle intensities is calculated, multiplied with the weight and then all summed. A sort of weighted sum.

Now I am trying to do the same for the LBP model and find an explanantion there. However even in the case of a stump we have other data present. Take this part for example:

<internalNodes>
0 -1 46 -67130709 -21569 -1426120013 -1275125205 -21585
-16385 587145899 -24005
</internalNodes>

<_>
<rect>0 0 3 5</rect>
</_>


Lets assume that the first and second parameter of the node are the same as with HAAR indicating the level of the stump, meaning that we have a 2 leaf decision tree here. 46 is then the index of the feature associated with it. Now that feature has a single rectangle as seen below.

My questions:

• Are the 8 values after the index some thresholds? I am assuming this due to the fact their are 8 and a LBP calculation is a comparison of 8 sums compared to the center value. If these are blocks instead of pixels and they are calculated on the integral image I would somewhat get it, but how to get positive and negative values then ... anyone an idea here?
• How should I interpret the rectangle as feature? It doesnt seem to be x y w h as with rectangles in HAAR.

All help apreciated!

edit retag close merge delete

@PhilippWagner you have quite done the research on LBP i saw on your webpage. Any ideas?

( 2015-01-31 08:42:38 -0500 )edit

Sort by » oldest newest most voted

Nevermind, found the explanation, and now it becomes all clear. The 8 values combined make the 256 bit LUT which is constructed by comparing the LBP regional subrectangles to the center subrectangle. The rect stored in the features section is the top left region used to calculated the LPB feature.

More detailed info on this: http://w3facility.org/question/unders...

Basically accept that a 256 bit LUT can be used to bitshift certain values and to retrieve the 0's and 1's without going into too much details.

more

1

( 2015-01-31 10:16:31 -0500 )edit

Official site

GitHub

Wiki

Documentation