In my opinion it depends on the environment... Maybe you can do a detection on the 3 methods and then use some kind of classification for enforcing the result. This because:
Ok; the part of the classifier is based on a dataset for training. This you can do after you will have some images of "true hand" and some images of "false hands". And it is more linked to the application specifications. (For example you cans see in Figure 1 on the paper of mittal that there is a "false hand on the chest of the second photo", so to remove those kind of detections). Classifiers are based on many types of features, as SIFT; and the evaluation is based on other things, like SVM. But for more info see this and classification in OpenCV like this one, but there are others too.
I like this suggestion. Make 3 algorithms, let them all make a certainty vote, and then choose the one with then combine the votes into a certainty map of the image, thresholding that map for detections. Like said by @thdrksdfthmn it depends very much on the actual situation of your application. Doing it inside a very controlled environment works perfectly with HSV, but it fails the moment sunlight comes into the game!
Besides that, it could be good to take a look here. Mittal has done some extensive work on hand detection. His LatentSVM part based model works very accurate!
@thdrksdfthmn care to explain more about 'use some kind of classification for enforcing the result'? I'm new here
Asked: 2014-07-24 00:27:40 -0600
Seen: 1,834 times
Last updated: Jul 24 '14
