Stephan Tschechne Chair for Image Understanding Computer Science Technische Universität München Designing…

Stephan Tschechne

Chair for Image UnderstandingComputer Science

Technische Universität München

[email protected]

Designing vs. Learning the Objective Function for Face Model Fitting

Abschlußvortrag Diplomarbeit

28.6.06 2/15Technische Universität MünchenStephan Tschechne

• Model-based Image Understanding

• Face Model Fitting

• Objective Functions

• Experimental Results

Overview:

28.6.06 3/15Technische Universität MünchenStephan Tschechne

Understanding Facial Images Various Applications

Identification Mimics Hands-free Control

Image Database: 850 Natural Images

28.6.06 4/15Technische Universität MünchenStephan Tschechne

Deformable Face Model 134 Contour Points Perform PCA Point Distribution Model

Description of an Instance: Parameter Vector p = (x,y,scaling,rotation,deform1..deform17)

28.6.06 5/15Technische Universität MünchenStephan Tschechne

Objective Function• Fitting Algorithms Search for Correct p:

Optimisation Problem• Objective Function Calculates Fitting Accuracy • Lowest Value for Correct Solution

F(Img,p1)=0.0 F(Img,p2)=0.3 F(Img,p3)=0.6

28.6.06 6/15Technische Universität MünchenStephan Tschechne

RequirementsFormulation of Requirements for Robust Objective Functions:

R1: Correct Position of MinimumR2: One MinimumR3: Continuous BehaviourR4: Gradient Vectors Point Away

Optimal Objective Functions:

28.6.06 7/15Technische Universität MünchenStephan Tschechne

Traditional Objective FunctionsCalculation of Objective Function Value ?• Intuitive approach:

Manual Selection of Salient Features:• Distance to Edges..

28.6.06 8/15Technische Universität MünchenStephan Tschechne

Traditional Objective Functions

• …or Distance to Edges from Skin Colour Images

28.6.06 9/15Technische Universität MünchenStephan Tschechne

Traditional Approach Problem: Desired Edges are not the Strongest Ones

28.6.06 10/15Technische Universität MünchenStephan Tschechne

Contribution

Robust Objective Function Better Fulfillment of the Requirements

?!

28.6.06 11/15Technische Universität MünchenStephan Tschechne

Learning the Robust Objective Function Training data:

Ground Truth from Image Database Haar-like Features Desired Value from

Optimal Objective Function

Machine Learns Rules with Model Trees

28.6.06 12/15Technische Universität MünchenStephan Tschechne

Training Data: Feature Values Fi Result R Deliberately Move Instance to Gather Values

Model Trees Learn: F(Feature Values) Result

F1=134F2=66 … R=0.7

F1=54F2=234 … R=0.0

F1=281F2=11 … R=0.5

28.6.06 13/15Technische Universität MünchenStephan Tschechne

Experimental ResultsCenter: Correct ParametersAxes: Variation of p towards ….

..translation ..deformation

28.6.06 14/15Technische Universität MünchenStephan Tschechne

Challenges:

Image Database with Natural Images Database High Dimensionality of Parameter Vector Verification of Requirements

Future research: Model Tracking Other Models: 3D, Appearance Models.. Different Features Other Positions for Features

28.6.06 15/15Technische Universität MünchenStephan Tschechne

The End.

Any Questions ?