Haar cascade versus neural network convolution?
I was wondering if, in the cascading method, the mask size is constant and takes into account the different sizes of the faces in the image, if the original image is zoomed in and out to find overlaps.
I saw that in convolutional neural networks, the size of the weight matrix is constant, say 20 * 20
Other than that, I would like to know if the Caucasian or CNN cascade is faster for face detection. If in this case I could use a hara cascade to first determine the position of the face and then use CNN for face recognition.
Ps: -face is just an object, it can be any other thing like a car, a tree ...
source to share
If you haven't already, read the original Viola / Jones article . As you will see, there is no specific strategy for resizing / changing the characteristics of the hara. But it does make sense, and you should do it for the reasons you suspect. There is no “one way” to do this, so go ahead and try some ideas.
Since the structure does not specify what a weak classifier is, you cannot make any correct statements about learning rate. All I can say is that the hara cascade classifier takes a long time due to the fact that it has 3 layers of different algorithms. But I don't understand why this important learning rate is always slow. If this is a wrt CNN problem, your concept is wrong.
The alta johns' frame is pretty "full". You should first ask yourself why CNN can't solve your problem and then refine your approach instead of jumping to something completely different imo. The most useful combination of both methods I know of is CNN ensembles.
source to share