Learning the Appearance of Faces: A Unifying Approach for the Analysis and Synthesis of Images. Thomas Vetter Germany University of Freiburg
Computer Vision & Computer Graphics Computer Graphics can help to solve Computer Vision! | G (p) - I | 2 = min Parameters G ( image ) Parameters V ision ( image ) parameters image G raphics ( parameters )
Analysis by Synthesis 3D World Image Analysis Synthesis Image Model Image Description model parameter
Synthesis of Faces Input Image Modeler Result Database Face Analyzer 3D Head Morphable Face Model
Approach: Example based modeling of faces 2D Image 3D Face Models = w 1 * + w 2 * + w 3 * + w 4 * +...
Cylindrical Coordinates red(h, ) green(h, ) blue(h, ) red(h, ) green(h, ) blue(h, ) h radius(h, ) h
Morphing 3D Faces 3D Blend 3D Morph 1 __ =
Correspondence: A two step process! Correspondence between 1. two examples ( Optical Flow like algorithms). 2. many examples ( Morphable Model ) Reference Example 2 nd Example
= a 1 * + a 2 * + a 3 * + a 4 * +... b 1 * + b 2 * + b 3 * + b 4 * +... Vector space of 3D faces. A Morphable Model can generate new faces.
Manipulation of Faces Modeler
Modelling in Face Space Caricatur Original Average
Modelling the Appearance of Faces A face is represented as a point in face space. Which directions code for specific attributes ?
Learning from Labeled Example Faces Fitting a (linear) regression function
Facial Attributes WeightWeight OriginalOriginal Subjective Attractiveness
Transfer of Facial Expressions = Smile - - Originals: + Smile = Novel Face:
Facial Expressions OriginalOriginal
3D Shape from Images Face Analyzer 3D Head Input Image
Matching a Morphable 3D-Face-Model = R Optimization problem! a 1 * + a 2 * + a 3 * + a 4 * +.. b 1 * + b 2 * + b 3 * + b 4 * +..
Error Function Image difference Plausible parameters Minimize
Optimization Strategies Stochastic Gradient Decent Difference Decomposition
Future Challenges Which Object Classes are linear ? How to built them automatically?