1

I'm working on a project for which I require high accuracy for face detection. For detection I'm using opencv with python.

I'm still not able to achieve my target so I was thinking of manipulating the fields/data in the xml files - haarcascade_frontalface_default.xml and haarcascade_frontalface_alt2.xml.

However, I was unable to understand the tags in it. I want a brief explanation of the major tags only.

Community
  • 1
  • 1
diviquery
  • 569
  • 5
  • 19
  • 1
    no answer to your question, but manual or brute-force manipulation will probably NOT increase the accuracy (in a reasonable amount of time) at all. Have a look at literature about cascade training (probably viola jones). Maybe that'll answer your question. – Micka Jun 07 '17 at 08:24
  • 1
    This post explains the basic tags in a Haar cascade XML file: https://stackoverflow.com/questions/34895186/what-is-the-meaning-of-values-in-stage-xml-and-cascade-xml-for-opencv-cascade-cl/34897460#34897460 – raner Oct 02 '17 at 06:32

1 Answers1

1

Personally I used in more than one occasion the standard opencv's classifiers and they worked fine for me.

But there are a lot of features that could make this classifiers not the best choice for you, like for example the pose of the face (common problem), the excessive luminosity in the captured frame, etc.

You can try to create your own classifier (http://docs.opencv.org/trunk/d7/d8b/tutorial_py_face_detection.html):

  1. with an existing dataset (here you can choose the best one that fits for your goal based on the explanations http://www.face-rec.org/databases/)
  2. or with a custom dataset with pictures (a lot of) taken by you in similar conditions of when your application has to detect the face
luke88
  • 958
  • 19
  • 40
  • is it possible to use an existing dataset and add on some of the pictures taken by me in the positive and negative samples – diviquery Jun 10 '17 at 04:04
  • Yes, it is possible, use image of the same size of the original dataset. But keep in mind that add images that in the feature space are really separate from all the others not necessarily is a good thing. – luke88 Jun 12 '17 at 07:08