Presentation is loading. Please wait.

Presentation is loading. Please wait.

Stephan Tschechne Chair for Image Understanding Computer Science Technische Universität München Designing vs. Learning the Objective.

Similar presentations


Presentation on theme: "Stephan Tschechne Chair for Image Understanding Computer Science Technische Universität München Designing vs. Learning the Objective."— Presentation transcript:

1 Stephan Tschechne Chair for Image Understanding Computer Science Technische Universität München stephan.tschechne@gmx.de Designing vs. Learning the Objective Function for Face Model Fitting Abschlußvortrag Diplomarbeit

2 28.6.06 2/15 Technische Universität München Stephan Tschechne Model-based Image Understanding Face Model Fitting Objective Functions Experimental Results Overview:

3 28.6.06 3/15 Technische Universität München Stephan Tschechne Understanding Facial Images  Various Applications  Identification  Mimics  Hands-free Control  Image Database: 850 Natural Images

4 28.6.06 4/15 Technische Universität München Stephan 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)

5 28.6.06 5/15 Technische Universität München Stephan 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)=0.3F(Img,p3)=0.6

6 28.6.06 6/15 Technische Universität München Stephan Tschechne Requirements Formulation of Requirements for Robust Objective Functions: R1: Correct Position of Minimum R2: One Minimum R3: Continuous Behaviour R4: Gradient Vectors Point Away Optimal Objective Functions:

7 28.6.06 7/15 Technische Universität München Stephan Tschechne Traditional Objective Functions Calculation of Objective Function Value ? Intuitive approach: Manual Selection of Salient Features: Distance to Edges..

8 28.6.06 8/15 Technische Universität München Stephan Tschechne Traditional Objective Functions …or Distance to Edges from Skin Colour Images

9 28.6.06 9/15 Technische Universität München Stephan Tschechne Traditional Approach  Problem: Desired Edges are not the Strongest Ones

10 28.6.06 10/15 Technische Universität München Stephan Tschechne Contribution  Robust Objective Function  Better Fulfillment of the Requirements ?!

11 28.6.06 11/15 Technische Universität München Stephan 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

12 28.6.06 12/15 Technische Universität München Stephan Tschechne  Training Data: Feature Values F i  Result R  Deliberately Move Instance to Gather Values  Model Trees Learn: F(Feature Values)  Result F 1 =134 F 2 =66 …  R=0.7 F 1 =54 F 2 =234 …  R=0.0 F 1 =281 F 2 =11 …  R=0.5

13 28.6.06 13/15 Technische Universität München Stephan Tschechne Experimental Results Center: Correct Parameters Axes: Variation of p towards …...translation..deformation

14 28.6.06 14/15 Technische Universität München Stephan 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

15 28.6.06 15/15 Technische Universität München Stephan Tschechne The End. Any Questions ?


Download ppt "Stephan Tschechne Chair for Image Understanding Computer Science Technische Universität München Designing vs. Learning the Objective."

Similar presentations


Ads by Google