Presentation is loading. Please wait.

Presentation is loading. Please wait.

Sylvia Pietzsch Chair for Image Understanding Computer Science Technische Universität München Learning Generic and Person Specific Objective.

Similar presentations


Presentation on theme: "Sylvia Pietzsch Chair for Image Understanding Computer Science Technische Universität München Learning Generic and Person Specific Objective."— Presentation transcript:

1 Sylvia Pietzsch Chair for Image Understanding Computer Science Technische Universität München pietzsch@in.tum.de Learning Generic and Person Specific Objective Functions Diplomarbeit

2 07.02.2007 2/16 Technische Universität München Sylvia Pietzsch Overview  Model-based Image Interpretation  Generic Objective Functions  Traditional Approach  Ideal Objective Functions  Learning Objective Functions  Experimental Evaluation  Person-specific Objective Functions  Experimental Evaluation

3 07.02.2007 3/16 Technische Universität München Sylvia Pietzsch Model-based Image Interpretation  Model Contains a parameter vector p that represents the model‘s configurations.  Objective Function Calculates how well a parameterized model fits to an image.  Fitting Algorithm Searches for the model that fits the image best by minimizing the objective function.

4 07.02.2007 4/16 Technische Universität München Sylvia Pietzsch Traditional Approach  Designer selects salient features from the image and composes them.  Based on designer‘s intuition and implicit knowledge of the domain.  shortcomings:  time-consuming  resulting objective function is not ideal

5 07.02.2007 5/16 Technische Universität München Sylvia Pietzsch Ideal Objective Functions P1:Correctness Property: The global minimum of the objective function corresponds to the best model fit. P2:Uni-Modality Property: The objective function has no local extrema or saddle points.

6 07.02.2007 6/16 Technische Universität München Sylvia Pietzsch Example: Comparing Objective Functions a) image b) along perpendicular c) edge values d) designed objective function e) ideal objective function f) training samples g) learned objective function

7 07.02.2007 7/16 Technische Universität München Sylvia Pietzsch Learning the Objective Function (1)

8 07.02.2007 8/16 Technische Universität München Sylvia Pietzsch Learning the Objective Function (2)

9 07.02.2007 9/16 Technische Universität München Sylvia Pietzsch Learning the Objective Function (3) 6 styles · 3 sizes · (5 · 5) locations = 450 features

10 07.02.2007 10/16 Technische Universität München Sylvia Pietzsch Evaluation 1: Used Features  Model trees tend to select the most relevant features.  Edge-based features are hardly used at all.

11 07.02.2007 11/16 Technische Universität München Sylvia Pietzsch Evaluation 2: Robustness Indicators measure the fulfillment of P1 and P2: I1:Correctness Indicator Distance between the ideal position of the contour point and the global minimum of the objective function I2:Uni-Modality Indicator Total number of local minima divided by the size of the considered region

12 07.02.2007 12/16 Technische Universität München Sylvia Pietzsch Evaluation 3: Learning Distance

13 07.02.2007 13/16 Technische Universität München Sylvia Pietzsch Person Specific Objective Functions  Single Images The objective function has to take any appearance of a human face into consideration. ➱ moderate accuracy  Image Sequence The appearance of a person‘s face only changes slightly.  Consider particular characteristics of the visible person, e.g. beard, glasses, bald head,... ➱ increase of accuracy  Challenges:  Learn specific objective functions for groups of persons offline.  Detect the correct group online.

14 07.02.2007 14/16 Technische Universität München Sylvia Pietzsch Evaluation: Fitting Results 45 persons from news broadcasts on TV

15 07.02.2007 15/16 Technische Universität München Sylvia Pietzsch Outlook  Learning objective functions for 3D-Models  Integration of further image features  Compute the image features on the fly  Automatic detection of the visible person: e.g. via AAM parameters

16 07.02.2007 16/16 Technische Universität München Sylvia Pietzsch The End

17 07.02.2007 17/16 Technische Universität München Sylvia Pietzsch Improving the Accuracy of the Training Step  model tree algorithm M5‘ uses standard deviation reduction as splitting criterion  shortcomings:  solution: taking not only the distance to the ideal feature point into consideration but also the direction  bim: Diese Folie passt hier nicht. Ich würde sie ganz weglassen. Oder am Schluss nur kurz ansprechen.

18 07.02.2007 18/16 Technische Universität München Sylvia Pietzsch Learning Person-specific Objective Functions offline step:  subdivide persons into groups  best possible grouping: minimum error between annotated model parameters and model parameters obtained by fitting the model to a set of test images using the objective function of this grouping  learning group-specific objective functions online step:

19 07.02.2007 19/16 Technische Universität München Sylvia Pietzsch The End


Download ppt "Sylvia Pietzsch Chair for Image Understanding Computer Science Technische Universität München Learning Generic and Person Specific Objective."

Similar presentations


Ads by Google