machine learning - Defining an (initial) set of Haar Like Features -

when comes cascade classifiers (using haar features) read methods adaboosting used select 'best' features detection. works if there initial set of features begin boosting.

given 24x24 pixel image there 162,336 possible haar features. might wrong here, don't think libraries opencv test against of these features.

so question how initial features selected or how generated? there guideline initial number of features?

and if 162,336 features used initially. how generated?

i presume, you're familiar viola/jones' original work on topic.

you start manually choosing feature type (e.g. rectangle a). gives mask can train weak classifiers. in order avoid moving mask pixel pixel , retraining (which take huge amounts of time , not better accuracy), can specify how feature moves in x , y direction per trained weak classifier. size of jumps depend on data size. goal have mask able move in , out of detected object. size of feature can variable.

after you've trained multiple classifiers respective feature (i.e. mask position), proceed adaboost , cascade training usual.

the number of features/weak classifiers highly dependent on data , experimental setup (i.e. type of classifier use). you'll need test parameters extensibly know type of features work best (rectangle/circle/tetris-like objects etc). worked on 2 years ago , took quite long time evaluate features , feature-generation-heuristics yielded best results.

if wanna start somewhere, take 1 of 4 original viola/jones features , train classifier applying anchored (0,0). train next classifier (x,0). next (2x,0)....(0,y), (0,2y), (0,4y),.. (x,y), (x, 2y) etc... , see happens. you'll see it's ok have less weak classifiers, i.e. can proceed increase x/y step values determine how mask slides. can have mask grow or other stuff save time. reason "lazy" feature generation works adaboost: long these features make classifiers better random, adaboost combine these classifiers meaningful classifier.