3

I was wondering if in haar cascade method the mask size is constant and to account for the different sizes of faces in an image , if the original image is shrunk and enlarged to find out overlaps .

I saw that in convolution neural networks the size of the weights matrix is constant at say 20*20

Apart from this , I would like to know if haar cascade or CNN is faster for face detection. If that's the case I could use haar cascade to first detect the positions of the face and then further use CNN to recognise the face.

Ps :-face is simply an object , it could be any other thing like car, tree.....

Chris Thaliyath
  • 86
  • 2
  • 7

1 Answers1

4

If you haven't yet, read the original Viola/Jones paper. As you'll see, there is no specific strategy given about resizing/reshaping the haar features. But it makes sense and you should do it for the reasons you suspect. There isn't "one way" to do it so go ahead and try some ideas.

Since the framework doesn't specify what the weak classifier is, you can't make any proper statements on training speed. All I can say is that haar-cascade-classifier takes a long time due to it having 3 layers of different algorithms. But I don't see why that's important, training speed is always slow. If that's an issue w.r.t. CNN, your concept is flawed.

The viola jones framework is quite "complete". You should first ask yourself why a 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 is CNN ensembles.

runDOSrun
  • 10,359
  • 7
  • 47
  • 57