How do I make my computer detect my emotions?

See more:

C#

Visual-Studio

neural-network

, +

I want my computer to learn what my face looks like when I am in a certain mood(self defined, not just happy and sad) via webcam and then predict the mood I am likely in(in per cent, not binary) by having given it training. Like http://auduno.github.io/clmtrackr/examples/clm_emotiondetection.html but tailored to me and not a generic classifier.

I did some research and it seems that I need to use a haar classifier. I looked into AForge und Accord, finding FaceHaarCascade. But this is not just generic, it also is binary. And it only detects faces, not emotions. At least I need to give it some facial expressions to choose the most likely one from and give the percentage of each.

How do you do it with aforge/accord?

Clarification: I would like to know, how to create a haar classifier with aforge or accord, so I can input 1000 pictures from me in one mood, 1000 pictures in another mood, let machine-learning create decision trees and finally use it. Aforge/Accord already offers this*, but there is no documentation for this task. I also thought that the eyes and nose are more important for mood as you can also laugh in a fake way(with wrinkles around eyes being tense). Basically I would let machine learning decide what makes my face express a certain emotion.

*Accord Vision offers HaarCascade and HaarClassifier whereas HaarCascade offers a method called FromXml which imports HaarCascades from OpenCV XML. For this I need to create an opencv-xml, which leads me to http://note.sonots.com/SciSoftware/haartraining.html but something like haartraining -data haarcascade -vec samples.vec -bg negatives.dat -nstages 20 -nsplits 2 -minhitrate 0.999 -maxfalsealarm 0.5 -npos 7000 -nneg 3019 -w 20 -h 20 -nonsym -mem 512 -mode ALL does not work, this is where I am stuck there. EDIT: It does work with npos < number of included pos. pictures.

Also, using the HaarObjectDetector, it does not offer a threshold output but rather a binary It is the object/It is not the object. What I want is % of each mood, ie similarity to certain faces not a "this is where you are happy/unhappy" type of result.