I am trying to improve face detections and i am using an implementation of viola jones algorithm in OpenCV. I am not getting the results that i desire, so i want to have some preprocessing before using the cascade.detectMultiScale function. The images that i am using contains upper body of a person with some background. Would it help if i use the cascade.detectMultiScale() function on cropped images containing just the face?
Asked
Active
Viewed 403 times
0
-
What ? If you can crop an image such that you have just the face, then you already solved the problem of face detection. – mmgp Feb 07 '13 at 23:10
-
How many images are you using in your training set? Typically the Voila-Jones classifier has pretty good results for face detection. – Max Allan Feb 08 '13 at 21:04
-
@mmgp I was thinking of using some other algorithm to have a rough idea of where a face might be (kind of defining a region of interest) and then use viola jones. But that "other algorithm" may not give reliable results, so i can not make a decision based on that. currently, i was thinking of obtaining training images through a video, so in this case i can use motion detection to identify an approximate position of face (making sure only the face is moving and the background is not) and then using viola jones. Would it yield better detections? – ankitsingh Feb 11 '13 at 07:16
-
@Max i have a set of about 5500 images and out of those only 3500 are getting detected. I am using face detection and then eye detection to make sure it really is a face. As i understand, the accuracy of eye detection in Opencv using the supplied cascades is itself not very good. Any suggestion as to how i can reduce the false positives by only using face detection and not using eye detection? – ankitsingh Feb 11 '13 at 07:20
-
@user1871727 You are not likely to be able to improve the accuracy of the Viola Jones detector. If you look at samples of the images that are not being classified as faces, is there any common theme (such as being rotated away from the camera)? – Max Allan Feb 12 '13 at 16:10
-
@Max i am not able to determine any such theme. I am obtaining my image dataset myself and the camera i am using is pretty crappy. The images taken are not very sharp and sometimes i cannot see a distinct boundary between the face and the background. So i tried using a better camera and the detections are pretty good. So, i have come to a conclusion that the low detection rates are because of the crappy camera. However, i am not able to understand exactly how can i define an image as "bad". Is it the resolution, the amount of blur in it, not enough sharpness? – ankitsingh Feb 13 '13 at 06:26