SSIP Project 2 GRIM GRINS Michal Hradis Ágoston Róth Sándor Szabó Ilona Jedyk Team 2
SSIP OUR TEAM
SSIP Our team Michal Hradis Brno University of Technology, Czech Republic Brno University of Technology, Czech Republic Main Function BOSS
SSIP Ágoston Róth Babes-Bolyai University Kolozsvár, Romania Babes-Bolyai University Kolozsvár, Romania Main Function Listening to the Boss Our team
SSIP Our team Sándor Szabó University of Szeged, Hungary University of Szeged, Hungary Main Function Listening to the Boss
SSIP Our team Ilona Jedyk Technical University of Lodz, Poland Technical University of Lodz, Poland Main Function Listening to the Boss
SSIP Our task Localize face Recognizing of face expressions –neutral – surprised – angry – smiling Assumptions – pictures of single frontal face
SSIP Recognizing facial expression – TECHNIUQUES Method for classification –Support Vector Machine – best results –AdaBoost - good –Linear Discriminant Analysis – junk –Neural networks – ???? Method for feature selection (e.g. using PCA)
SSIP Face detection AdaBoost classifier with Haar-like features Training - CBL Face Database Multiple detections
SSIP AdaBoost “Strong” classifier constructed as linear combination of “week” classifiers Greedy selection of week classifiers from large set of features Feature (h(x) = {-1, 1}) –simple guess about sample class –high error ( )
SSIP AdaBoost conclusion Adventages –Low computation cost –High number of features (1000 – ) –High number of samples Disadvatages –Gready selection – suboptimal result
SSIP Recognizing facial expression AdaBoost classifier with Haar-like features Database of face expression –MMI face database –photos of SSIP participants –Automatic face extraction with our face localization –100 – 200 samples per class
SSIP Decision Neutral Angry Surprised Happy
SSIP Program Program in C++ Using Open CV Library AdaBoost Training –Form VUT Brno Inputs: –Expression classifiers (text file) –Face detector (text file) –Detector configuration (text file) –Image with single frontal face Outputs: –Face image –Expression classification
SSIP Results
SSIP Conclusion It really works –75% corect recognition –State of the art around 90 % Not so good performance –Low number of training samples –Haar-like features are not well suited for this task Feature work –Use Gabor wavelets as features
SSIP References Intel, “Open Computer Vision Library, Reference Manual” Recognizing facial expression: machine learning and application to spontaneous behavior umber= umber= A Short Introduction to Boosting tut-ppr.pdf tut-ppr.pdf
SSIP Thanks for your attention