18
http://www.cubs.buffalo.edu Face Detection University at Buffalo CSE666 Lecture Slides Resources:

University at Buffalo CSE666 Lecture Slides Resources · Face Detection University at Buffalo CSE666 Lecture Slides Resources:

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Face Detection

University at Buffalo CSE666 Lecture Slides

Resources:

Page 2: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Applications

• Biometrics• Emotional Expression• Person tracking

• Surveillance• Scene analysis• Military applications

• General object location and recognition

Still photo vs. Video, still camera vs. Video, moving camera

easiestmost difficult?

Page 3: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Skin Color Approaches

Kovac et al., "Human Skin Colour Clustering for Face Detection"

Page 4: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Classification Approaches• Select a region from image• Extract some features• Classify feature vector – face or no face

Heisele et al., “Component-based Face Detection”

Page 5: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Neural Network

H. Rowley et al., “Neural Network-based face detection”, PAMI96

Page 6: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

AdaBoost

P. Viola, M. Jones “Robust Real-Time Face Detection”, 2004

• Iteratively chooses which ‘weak classifiers’ (features) are best for currently incorrectly classified images.• Found best feature is added to final classifier in weighted sum.• Additional cascading structure is used in paper

Page 7: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

AdaBoost

P. Viola, M. Jones “Robust Real-Time Face Detection”, 2004

Page 8: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Principal Component Analysis (PCA)

Previous discussion: PCA as a feature extraction method

( )iaA =

∑=

=n

iiiy

1

ax

∑=

=m

iiiy

1

ˆ ax

- representation of original vector in

KL basis vectors

- the projection on subspace spanned by eigenvectors with largest eigenvalues

∑+=

=−n

miiE

1

2 ])ˆ[( λxx

- Karhunen-Loeve transform

ia - eigenvectors of matrix ][ Tx E xxR =

x

m

Axy = - projection of original vector to KL basisx

xAy mm = - first PCA coefficients –‘best’ features

m

Features Representation

- sum of the smallest eigenvaluesmn −

Page 9: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

PCA as appearance model

xAy mm = - first PCA coefficients –

‘best’ features - should provide a good approximation to all the samples (for which PCA was trained)

• Distance from feature space (DFFS) should be small:

m

|ˆ| xx −=DFFS

x

• PCA provides a model on typical samples of the class• Can use it to separate class from non-class samples

mTm yAx =ˆ

Page 10: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

PCA for face detection

• Collect a set of face images for training

• Train PCA

• During face detection, for each query region, if projection of query region onto PCA feature space is close to original region:

then accept region as a face region

θ<− |ˆ| xx

But:• many training images are required • high complexity of PCA training

Page 11: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Active Shape Model (ASM)

Idea:use PCA not on the pixels of face image, but only on selected model points

Training • 133 landmark points in each image represent the face shape• Landmark points correspond to edges – points of large intensity change• PCA training is performed using vectors with all 133 landmark point coordinates

),,...,,( 13313311 yxyxv =

T. Cootes, C. Taylor “Statistical Models of Appearance for Computer Vision”

Page 12: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Active Shape Model (ASM)

Fitting ASM:• Begin with best guess placement of shape on the query image• Iteratively update shape so that

1. PCA model holds 2. The local profile at landmark points coincides with profiles of

training images

θ<− |ˆ| vv

Page 13: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Active Shape Model (ASM)

Page 14: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Active Shape Model (ASM)

Page 15: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Active Appearance Model (AAM)

• A generalization and improvement on ASM• In ASM the correspondence between landmark points and local texture of the image (say, edge property) was calculated heuristically• AAM: model the texture of the image as well as shape of landmark points by PCA

-shape parameters- texture parameterTrain PCA on concatenated vectors-still have one PCA

• Model fitting error

• Search for a good fit:It is possible to train a search algorithm on difference images using training data and perturbing itVery fast searching algorithm

vg

),( gvc =

θ<− |ˆ| cc

)ˆ(Image c−I

Page 16: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Active Appearance Model (AAM)

Page 17: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Active Appearance Model (AAM)

Page 18: University at Buffalo CSE666 Lecture Slides Resources ·  Face Detection University at Buffalo CSE666 Lecture Slides Resources:

http://www.cubs.buffalo.edu

Active Appearance Model (AAM)

• We used implementation by M. Stegmann• Works well for tracking