Detecting Faces in Images: A SurveyBy: Ming-Hsuan Yang, David J. Kriegman, and Narendra AhujaPresented By: Neal AudenaertAgenda{ Introduction{ Approachesz Knowledge-basedz Feature invariantz Template matchingz Appearnce-based{ Databases and Evaluation{ DiscussionAgenda{ Introduction{ Approachesz Knowledge-basedz Feature invariantz Template matchingz Appearnce-based{ Databases and Evaluation{ DiscussionIntroduction{ Domainz Face detection (not recognition)z Still images{ Objectivesz Comprehensive survey of techniquesz Discussion of performance measures{ Limitationsz Methods are not directly comparableChallenges{ Pose{ Structural components{ Facial expression{ Occlusion{ Image orientation{ Imaging conditionsGeneral Tasks{ Face localization{ Facial feature detection{ Face recognition{ Face authentication{ Face tracking{ Facial expression recognitionAgenda{ Introduction{ Approachesz Knowledge-basedz Feature invariantz Template matchingz Appearnce-based{ Databases and Evaluation{ DiscussionSurvey of Techniques{ Knowledge Basedz Top-downz Bottom-up{ Template Basedz Defined templatesz Learned templatesKnowledge-basedFeature invariantTemplate matchingAppearance-basedSurvey of TechniquesXAppearance-basedXXTemplate MatchingXFeature InvariantXKnowledge-basedDet.Loc.ApproachKnowledge-based Feature Invariant Template Matching Appearance-basedKnowledge-Based Top-Down MethodsMain Idea: Use knowledge about what constitutes a face faces to define rulesStrengths: Frontal faces in uncluttered scenesWeaknesses:Translating knowledge into rulesEnumeration of casesKnowledge-based Feature Invariant Template Matching Appearance-basedKnowledge-Based Top-Down MethodsBottom-Up Feature-Based MethodsMain Idea: Describe relationships between invariant features using statistical models Strengths: Improved invariance for different poses and lighting conditionsWeaknesses: Corruption of individual due to illumination, noise, or occlusionKnowledge-based Feature Invariant Template Matching Appearance-basedBottom-Up Feature-Based Methods{ Facial Features{ Texture{ Skin Color{ Multiple FeaturesKnowledge-based Feature Invariant Template Matching Appearance-basedTemplate MatchingMain Idea: Find correlation values with a standard face pattern for face contour, eyes, nose, and mouthStrengths: Simple to implementWeaknesses: Cannot deal with variation in scale, pose, and shapeAlternatives: Multiresolution, multiscale, subtemplates, and deformable templatesKnowledge-based Feature Invariant Template Matching Appearance-basedAppearance-Based MethodsMain Idea: Use statistical analysis and machine learning techniques to learn “template” characteristicsStrengths: Most successful approachWeaknesses: Relatively complex to implement, high-dimensionality requires many training examplesKnowledge-based Feature Invariant Template Matching Appearance-basedOverview of Techniques{ Eigenfaces{ Distribution-Based Methods{ Neural Networks (ANN){ Support Vector Machines (SVM){ Sparse Network of Winnows (SNoW){ Naïve Bayes Classifier{ Hidden Markov Models (HMM){ Information Theoretic Approaches{ Inductive LearningKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveEigenfaces{ Pedro?{ Main Idea: Calculate distance between an instance and exemplary data in a reduced dimensional spacez Build a face mapKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveDistribution-Based Methods{ Fit a distribution model to examples{ Project example into reduced dimensional space{ Build classifier to decide face/non-faceKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveSung and PoggioKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveSung and PoggioKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveSung and Poggio{ Mahalanobisz PCA for each cluster{ Representative sample of non-face images?z Bootstrap approachKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveYang, Ahuja, Kriegman{ Method 1:z Factor Analysis{ Instead of PCA{ Does not define a mixture modelz Estimate mixture model using EM{ Method 2:z Fisher’s Linear Discriminantz Class decomposistion using Kohonen’s Self Organizing Mapsz ML decision rule to detect facesKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveNeural Networks{ Two class pattern recognition { Advantage: capture complex class conditional desnsity{ Drawback: Requires extensively tuningKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveSupport Vector Machines{ Estimating hyperplane is expensive{ Evaluation is fastKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveSparse Network of Winnows{ Detect images with:z Different features and expressionsz Different posesz Different lighting conditions{ Primitive and multiscale features{ Tailored for domains wherez Number of features is largez Features unknown a priori Knowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveNaïve Bayes Classifier{ Estimate joint probability of local appearance z c.f. bottom-up methods{ Emphasize local appearancez Some patterns are more unique{ Detects some ratated and profile facesKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveHidden Markov ModelKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveHidden Markov Model{ Alternativesz Karhunen Loeve Traformcoefficients as input to HMMz Use HMM to learn face to non-face transitionKnowledge-based Feature Invariant Template Matching Appearance-basedEigenfacesDistributionANNSVMSNoWBaysianHMMInfo. Theory InductiveInformation Theoretic Approaches{ Markov Random Fields (MRF)z Model context-dependent entities{ Kullback relative inforamtionz Maximize information-based
View Full Document