E-mail: [email protected]://web.yonsei.ac.kr/hgjung
MorphableMorphable ModelModel
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Graphics and Vision Research Group, University of Basel, http://gravis.cs.unibas.ch/
Volker Blanz
and Thomas Vetter, “Face Recognition Based on Fitting a 3D Morphable
Model,” IEEE PAMI, Vol. 25, No. 9, Sep. 2003, pp. 1063-1074.
SourcesSources
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
MorphableMorphable Model based Face Recognition SystemModel based Face Recognition System
= Deformable 3D Model= Deformable 3D Model
+ Computer Graphics Simulation of Projection and Illumination+ Computer Graphics Simulation of Projection and Illumination
Intrinsic shape and texture are fully independent of extrinsic parameter
MorphableMorphable ModelModel
- High dimensional vector of shape
- High dimensional vector of texture
- Probability density function of natural faces within face space
MorphableMorphable
ModelModel
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Paradigm 1. (The approach taken in this paper)Recognition can be based on model coefficients, which represent intrinsic shape and texture of faces, and are independent of the imaging conditions
Paradigms for Model Based RecognitionParadigms for Model Based Recognition
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Paradigm 2.
Viewpoint dependentRecognition System
SyntheticViews
Original (PIE images)
Reconstructions renderedInto the originals
Same reconstructions renderedwith standard illumination
Paradigms for Model Based RecognitionParadigms for Model Based Recognition
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
- Vector space representation of faces
- Convex combination of shape and texture vectors Si and Ti
A A MorphableMorphable
Model of 3D FacesModel of 3D Faces
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Database of Three-Dimensional Laser Scans- 100 males, 100 females, aged between 18 and 45- Recorded with a Cyberware 3030PS laser scanner(512 angular steps, 512 vertical steps = 0.615mm)
Preprocessing of Raw Scans- Filling holes and removing spikes with an interactive tool- Automated 3D alignment of the faces with 3D-3D Absolute Orientation- Semiautomatic trimming along the edge of a bathing cap- Vertical, planar cut behind the ears and a horizontal cut at the neck
Constructing Constructing MorphableMorphable
Model from 3D ScansModel from 3D Scans
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
- Correspondence between each face and a reference face is given bya dense vector field
Constructing Constructing MorphableMorphable
Model from 3D ScansModel from 3D Scans
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Correspondence Based on Optic Flow“Objects in image sequences I(x,y,t) retain their brightnesses as they moveacross the image at a velocity (vx ,vy )T”
- Minimizing the following expression
- v is assumed to be constant on each neighborhood R(x0 ,y0 ) 5x5- In each point (x0 ,y0 ), v(x0 ,y0 ) can be found by solving 2x2 linear system
Constructing Constructing MorphableMorphable
Model from 3D ScansModel from 3D Scans
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Definition of Face Vectors
- Our reference face is a triangular mesh with 75,972 vertices
vertices of this mesh be located at
- The definition of shape and texture vectors is based on a reference face I0
- To encode a novel scan I
1) Compute the flow field from I0 to I
2) Convert to Cartesian coordinates
3) Coordinates and color values for the shape and
texture vectors S and T are then sampled at
Constructing Constructing MorphableMorphable
Model from 3D ScansModel from 3D Scans
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Principal Component Analysis
- Compute the eigenvectors s1 ,s2 , … by SVD- The eigenvalues of C are the variances of the data along each eigenvector- Eigenvectors form an orthogonal basis
- PCA provides an estimate of the probability density within face space :
Constructing Constructing MorphableMorphable
Model from 3D ScansModel from 3D Scans
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Segments- To generate a larger variety of different faces linear combinations of shape and texture are formed separately for different regions of the face : eyes, nose, mouth and the surrounding area
- For continuous transitions between the segments, we apply a modification of the image blending technique
Constructing Constructing MorphableMorphable
Model from 3D ScansModel from 3D Scans
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
- Estimate shape coefficient ai and texture coefficient bi
relevant parameters of the 3D scene(pose, focal length of the camera, light intensity, color and direction)
- In an analysis-by-synthesis loop, the algorithm finds model parametersand scene parameters such that the model, rendered by computer graphicsalgorithms, produces an image as similar as possible to the input image
ModelModel--Based Image AnalysisBased Image Analysis
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
- For initialization, the systemcurrently requires image coordinatesof about seven facial feature points,such as the corners of the eyes orthe tip of the nose
- With an interactive tool, the userdefines these points
ModelModel--Based Image AnalysisBased Image Analysis
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Image Positions of Vertices- Object-centered coordinates
Illumination and Color- The illumination model of Phong approximately describes the diffuse and
specular reflection of a surface
Image SynthesisImage Synthesis
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Optimizing1)2)3) 22 rendering parameters
--------
Fitting the Model to an ImageFitting the Model to an Image
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Cost Function- Minimize the sum of square differences
- Minimize feature point error
Maximum A Posteriori estimator
Fitting the Model to an ImageFitting the Model to an Image
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
- Ad hoc choices of and are used to control the relative weights- At the beginning, prior probability end EF are weighted high
The final iterations put more weight on EI and no longer rely on EF
Fitting the Model to an ImageFitting the Model to an Image
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Optimization ProcedureThe core of the fitting procedure is a minimization of the cost functionwith a stochastic version of Newton’s method- Select a set K of 40 random triangles in each iteration
Evaluate EI and its gradient only at their centers- Set the probability of selecting a particular triangle proportional to its area- Area calculation once every 1,000 iterations
The optimum is
- Only consider ai
Fitting the Model to an ImageFitting the Model to an Image
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
- These second derivatives are computed by numerical differentiation fromthe analytically calculated first derivatives, based on 300 random vertices,at the beginning of the optimization and once every 1,000 iterations
- In each iteration, we perform small steps
Fitting the Model to an ImageFitting the Model to an Image
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Reconstructions of 3D shape and texture from FERET images
Results of Model FittingResults of Model Fitting
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Reconstructions of 3D shape and texture from CMU-PIE images
- Illumination in the third image is not fully recovered, so part of thereflections are attributed to texture.
- The error of pose estimates is within a few degrees
Results of Model FittingResults of Model Fitting
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Similarity Measure
1) Mahalanobis distance :2) Cosine of the between angle :3) CW distance :
Results of Model FittingResults of Model Fitting
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
- Overall performance is best for the side-view gallery
Recognition PerformanceRecognition Performance
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
Recognition PerformanceRecognition Performance
E-mail: [email protected]://web.yonsei.ac.kr/hgjung
1% false alarm rate- CMU-PIE : hit rate 77.5%- FERET : hit rate 87.9%
Recognition PerformanceRecognition Performance