I think you are mixing things up.
HAAR-like wavelets are always rectangular features, which take areas of pixels and subtract those values. The end result is then tresholded to make a decision for a specific feature. More information on which type of haar-like wavelets can be found here. OpenCV supports the use of the basic orientations, but it also supports use of the rotated haar-like wavelets.
The Viola & Jones framework supports the use of this features. Depending on how the rectangles are oriented, one could say they search for specific behaviour, namely gradient information. A strong gradient in vertical direction, will make that there is a large difference between two vertical boxes moving over it.
This is why these features are stated as features looking for edges/lines, which is basically the same. The four rectangleqs type of features go looking for more specific information and gradient structures, like corners and such.
Basically read the Viola&Jones paper for more information on them.
And to answer to why you need different features, you do not want to limit the algorithm to only vertical or horizontal features, but you want to combine horizontal, vertical and rotated features all together.
It applies the boosting algorithm. Again, your should read the paper, in order to get better insight. Basically what happens is during training phase, for each window all possible features are calculated. Next the boosting algorithm tries to define the most descriminate feature (best division of positive and negative data). Only those features are remembered in the final trained model, and need to be calculated at runtime. This is why training of cascade classifiers takes that long.
Asked: 2013-05-30 22:56:44 -0600
Seen: 511 times
Last updated: May 31 '13
