Hi all,
I'm building my own smile detector, because the existing of haarcascade_smile.xml based on Viola-Jones not good, I want to make a classifier with 3-classes: normal – opening mouth (not smile) – smile. First, i’m using Logistic Regression (in ml lib of Opencv 3) to make 2 binary classifier: normal – opening mouth, normal – smile. It’s better than Viola-Jones, but it still get many false-positives. So in your experience, is that HOG+Linear SVM or LBP+Linear SVM better? Or could you suggest any approaches?
Otherwise, I have to apply this smile detector in mobile with Objective-C or Java. If I use the algorithms in cv::ml of Opencv, I can save this model as xml file, and load it in Objective-C. It's great because I just training one time and re-use it without training for each prediction, so my mobile application will be "realtime" when using with camera (meaning it's can run ~ 30fps). So, if you suggest any way to make my own detector, please consider this condition.
Many thanks!